9	Understanding Sentence Structure and Meaning

As we saw in Chapter 6, the mind is exquisitely adept at relating syntactic structure to meaning, able to compute the meanings of long, complex sentences, even those containing numerous clauses nested within each other and encoding intricate relationships among their elements, of which this particular sentence is an excellent example.

Despite our great parsing prowess, we still stumble over some sentences. We’ve all had the experience of finding ourselves re-reading certain sentences over and over, not quite sure how to unravel their meanings even though none of the individual words is especially complicated or unfamiliar. Some sentences just feel knotty or clunky. They’re the sentences that an English teacher or an editor might single out with the remark “awkward sentence structure,” or “this feels clumsy.”

Let’s see if you have an intuitive sense of the kinds of sentences that strain language comprehension. Of the perfectly grammatical sentences listed on the next page, some slide through the mind with ease, while others seem to create hard-to-read “word piles.” Make a note of the ones you find taxing or confusing. For the time being, rely solely on your editorial instinct, but by the end of this chapter, you should have some scientific understanding of why they cause problems for the reader. (And if you ever do find yourself working as an editor, you’ll be able explain to your authors exactly what’s gone awry with many of their bad sentences.)

The boy watched the ball with the yellow stripe float down the river.

Susanne put the toy soldier in the box into the cupboard.

The soup and the vegetables simmered lightly on the stove while the hostess served some wine.

The patient told the doctor that he was having some trouble with about his sore knee.

The baker is going to sue the plumber who botched the installation of his new pipes last week.

The suspect who was seen last night at the crime scene has confessed to the murder.

The mouse the cat chased keeled over.

The boy who stalked the prom queen has been caught.

The teachers taught new math techniques passed the exam with flying colors.

Woody confided to his therapist that he occasionally thought about murdering his mother.

While the pop star sang the national anthem played in the background.

The cruel man beat his puppy with a thick stick.

Samantha explained to her son’s dentist that the boy’s grandparents kept feeding him candy.

Friedrich cooked the ribs and the corn had already been roasted.

The farmer slaughtered the goose that had intimidated his hens.

The gang leader hit the lawyer with a wart.

As Marilyn smiled the photographers snapped dozens of pictures.

The bartender told the detective that the suspect will try to escape the country yesterday.

The coffee in the red mug was already cold by the time the secretary had time to drink it.

The administrator who the intern who the nurse supervised had accused fudged the medical reports.

At least nine of these sentences would be recognized by most psycholinguists as potential troublemakers, containing elements known to cause problems for readers or hearers. Obviously, being able to recognize what makes sentences tricky to read is exceptionally useful if you’re a writer or editor. But if you’re a psycholinguist, it also gives you some insights into how human beings go about structuring the words of incoming speech or text into sentences with complex meanings. We can figure out a great deal about how human language comprehension works in real time by observing it at its limits and seeing where it breaks down.

In this chapter, you’ll get a chance to explore the nature of difficult sentences, whether written or spoken. But this isn’t just a story about awkward sentences. As you’ll see, in addition to learning how glitches in processing come about, you will ultimately gain an appreciation of why it is that, most of the time, comprehension glides along perfectly smoothly.

9.1 Incremental Processing and the Problem of Ambiguity

As you saw in Chapter 8, the word recognition system is not timid about making rapid, reasonable guesses about which words are in the midst of being pronounced; it begins to link sounds with possible words and their meanings from the very first moments of an utterance, activating a large number of possibilities in the memory store and gradually winnowing these down to the very best candidates.

But what about sentences? Unlike individual words, their meanings don’t just depend on retrieving meanings that are pre-stored in memory. The meaning of each sentence has to be constructed anew based on the syntactic relationships among all the words in the sentence. Otherwise, we wouldn’t be able to understand sentences we’d never heard before, having never had the opportunity to memorize their meanings. This process of structure-building during comprehension is referred to as parsing. Psycholinguists often use “the parser” as a convenient term for the collection of structure-building mechanisms and procedures; the term does not refer to an individual person. So, how long does it take for the parser to build these meaningful structures? Does each word in the sentence have to be uttered and fished out of the hearer’s memory first before syntactic grouping and structuring can begin to take place?

parsing The process of assigning syntactic structure to the incoming words of a sentence during language comprehension. The structure-building mechanisms and procedures collectively are often referred to as “the parser”; the term does not refer to an individual.

It seems not. The parser is just as eager as the word recognition system in generating guesses about possible meanings, even on the basis of very partial information from the speech stream. Like word recognition, understanding sentences is an exercise in incrementality—that is, meaning is built on the fly as the speech comes in, rather than being delayed until some amount of linguistic material has accumulated first. One way to demonstrate that understanding follows hot on the heels of speech is by means of a shadowing task, in which subjects are asked to repeat the words of a speaker’s sentence almost as quickly as the speaker produces them. People who are very good at this can follow at a lag of about a quarter of a second—roughly a lag of one syllable behind the speaker. This suggests that within hundreds of milliseconds of a word, not only has its meaning been recognized, but it has been integrated with the syntactic structure and meaning of the sentence. We know that people are actually analyzing the sentence’s meaning rather than just parroting words or even just the sounds of words, because shadowing gets considerably slower when meaningful sentences are replaced with nonsensical sentences (but with recognizable words), and slower yet if the sentence that has to be repeated is in a foreign language or uses made-up words.

incrementality The property of synthesizing and building meaning “on the fly” based on partial information as speech unfolds, rather than delaying processing until some amount of linguistic material has accumulated.

shadowing task An experimental task in which subjects are asked to repeat the words of a speaker’s sentence almost as quickly as the speaker produces them.

But there’s a downside to this “hungry” aspect of language processing. Like an overeager student who blurts out the answer before the teacher has finished asking the question, the parser’s guesses, based on only partial information from the sentence, may not all turn out to be the right ones. You might remember that this was also a factor with the eager word recognition system. Along with the correct target word (for example, candy), soundalike words, especially those that begin with the same sounds, also become activated (candle, Canada, cantaloupe). Evidence of such spurious activation can be seen in priming tasks or in patterns of eye movements to visual displays. In other words, extreme eagerness or incrementality in word recognition leads to an explosion of potential ambiguity of meanings, at least for a short period of time. The processing system ultimately has to suppress a number of possible interpretations that started out as being perfectly consistent with the uttered speech. This happens at the level of the sentence as well.

WEB ACTIVITY 9.1

Processing language on the fly Try your hand at a shadowing task. You’ll first be able to listen to someone shadowing speech, and then have an opportunity to shadow a variety of sensible, senseless, and foreign-language sentences.

https://oup-arc.com/access/content/sedivy-2e-student-resources/sedivy2e-chapter-9-web-activity-1

“Garden path” sentences

Successful interpretation is often a matter of not getting tripped up by the incorrect meanings that also become activated during comprehension. But sometimes, one of the alternative meanings causes so much disruption during parsing that it becomes truly difficult to recover the correct meaning of the sentence. Consider the following, probably the single most famous sentence in psycholinguistics, brought to life by Tom Bever (1970):

The horse raced past the barn fell.

On a first reading, many people think this sentence makes no sense, or is ungrammatical with some words left out. But it’s perfectly grammatical. Don’t believe me? It means exactly the same thing as this:

The horse that was raced past the barn (by someone) fell.

Still don’t see it? It has exactly the same structure and a very similar meaning to this:

The horse driven past the barn fell.

Assuming you’ve eventually been able to parse the sentence correctly, the puzzle becomes this: What is it that makes that first sentence so dastardly difficult to understand, much more so than the other two, even though all of them are permitted by the grammar of English? The answer is that only the first has the potential for a temporary ambiguity; until the very last word of the sentence, it’s most natural to understand the sentence as being structured so that the word horse is the subject of the verb raced—that is, the horse is doing the racing, rather than being raced. This interpretation would have worked out just fine if the sentence had continued this way:

The horse raced past the barn and fell.

You can see why people tend to have the feeling that the sentence is missing a word or two. In order to get the right reading for our original sentence, you have to ignore this highly tempting meaning for the first six words of the sentence, and instead interpret them as being structured such that the phrase raced past the barn is a separate clause that adds more information about the horse. This is called a reduced relative clause because it’s just like the (non-reduced) relative clause that was raced past the barn with some of the function words taken out. In this more complicated structure, notice that the horse acts semantically as the direct object of raced—that is, it is being raced, rather than doing the racing—but also acts as the subject of the main clause verb fell. Figure 9.1 shows one way in which to graphically capture the two different interpretations. (And, if grammatical terms like “direct object” and “relative clause” aren’t part of your daily vocabulary, Box 9.1 offers a quick refresher.)

reduced relative clause A grammatical structure in English involving a relative clause in which certain function words have been omitted (e.g., the reduced relative clause raced past the barn derives from the full relative clause that was raced past the barn). This structure often leads to ambiguity.

Here are a few more sentences with reduced relative clauses, and with the same potential for ambiguity as our “horse-raced” example—though a number of them seem to cause somewhat milder reading hiccups, for reasons we’ll see later:

The swimmers drowned in the lake were not found until the following spring.

The general presented copies of the report was aware of the problems.

The boat floated downstream will never get past the rapids safely.

Difficult sentences like these are poetically called garden path sentences, since their effect is to lead readers down a garden path, misleading them into interpreting the sentence one way, but then veering off in another direction entirely. English offers a generous variety of opportunities for garden path sentences, aside from the reduced relative/main clause ambiguity. For example:

garden path sentences Sentences that are difficult to understand because they contain a temporary ambiguity. The tendency is for hearers or readers to initially interpret the ambiguous structure incorrectly, and then experience confusion when that initial interpretation turns out to be grammatically incompatible with later material in the sentence.

Figure 9.1 Tree structures illustrating two possible syntactic interpretations of the string The horse raced past the barn … . (A) A main clause interpretation. (B) An interpretation involving a reduced relative clause, with the anticipation of additional content to be uttered as part of the verb phrase (VP).

After the police stopped the car …

Here, the car could be read as either the direct object of the verb stopped (as in After the police stopped the car they noticed its license plate was missing), or as the subject of the following clause (After the police stopped the car sped off down the road). The first of these possible continuations probably feels more graceful to you, while the second is a bit bumpy, presumably because you’ve given in to the desire to interpret a noun phrase immediately following a verb as a direct object of that verb. This creates problems when you then encounter the second verb sped off, which nonsensically appears to have no subject, so you need to go back and reanalyze the sentence so that the car is interpreted as its subject.

Now try supplying a plausible continuation for this sentence:

The married man promised his mistress that he would soon abandon …

If you suggested that the next words should be his wife, then you’ve taken the heavily traveled path of interpreting that he would soon abandon as introducing a sentence complement of the verb promised—that is, what the married man promised was that he would soon abandon someone, perhaps his wife. But another structure is possible, as evident in this sample continuation:

The married man promised his mistress that he would soon abandon a diamond ring.

Since it’s decidedly bizarre to abandon a diamond ring, the reader is nudged into a different interpretation and, if successful, will eventually settle on a reading in which that he would soon abandon is a relative clause, providing additional information about his mistress (similar to The married man gave the woman that he would soon abandon a diamond ring.)

Once you start looking, you’ll be able to diagnose many awkward-sounding sentences as garden path sentences. For instance, each of the following sentences contains a temporary ambiguity. Try to identify the ambiguous fragment of the sentence and then think about another possible way to continue the sentence that would be consistent with that alternative reading of the ambiguous string of words:

The investigation revealed the error resulted from gross negligence.

Sam ate the hot dog and the vegetables went untouched.

The gangster shot the cop with the gun.

This company only hires smart women and men who may be dumb but who have clout.

BOX 9.1

Key grammatical terms and concepts in English

Subject

The noun phrase (NP) that appears to the left of a verb phrase (VP) and combines with it to form a sentence. The subject is often described as “what the sentence is about.” When the sentence is in the active voice, the subject is typically (but not always) the cause or the instigator of the event described by the verb:

Copernicus made an important discovery.

The third man on the left is acting suspiciously.

She died on August 12, 1542.

Direct object

A noun phrase that appears inside of the verb phrase, to the right of the verb. Not all verbs take a direct object; whether a verb allows/requires one or not must be specified as part of the lexical knowledge of that verb. When the sentence is in the active voice, the direct object is usually the entity that is acted upon, or comes into being, as a result of the actions of the subject:

Copernicus made an important discovery.

The president fired his chief of staff.

The police stopped the car.

Cyrano wrote an eloquent love letter.

Indirect object

Occasionally, two noun phrases occur inside a verb phrase; in this case, only one is the direct object; the other NP is the indirect object. Rather than expressing the acted-upon entity, the indirect object usually expresses the recipient of the acted-upon thing. An indirect object either appears immediately after the verb or is introduced by a preposition:

Cyrano wrote Roxanne an eloquent love letter.

Cyrano wrote an eloquent love letter for Roxanne.

Copernicus explained his discovery to many skeptical theologians.

The hooded man passed the bank teller a note.

Sentence complement

A clause, or a sentence unit, that appears inside the verb phrase, to the right of the verb. It is often (but not always) introduced by the complementizer word that, and may occur on its own, or together with a noun phrase:

His friends warned Copernicus that the authorities were planning his arrest.

The president claimed he had created many new jobs.

Main versus subordinate clauses

A sentence may have multiple sentence units contained within it, but usually only one is the main clause: this is the sentence that is at the top of the tree (as in Figure 9.1), while the other subordinate clauses are embedded inside other phrases. Subordinate clauses might appear within the verb phrase as sentence complements, attached to nouns as relative clauses, or introduced by adverbial words such as although, despite, after, etc. Here, the subordinate clauses appear in bold:

The horse raced while the cow stared.

Though her audience loved her, Marilyn was riddled with doubt.

The man who was acting suspiciously turned out to be an escaped convict.

She died because she had no money for the operation.

Conjoined clauses

Two or more constituents of the same type can be conjoined by and or but, including clauses. When a main clause is conjoined with another main clause, they carry equal weight, and both are considered to be main clauses:

The police stopped the car and the driver jumped out.

Copernicus made a discovery but he hesitated to reveal it.

Relative clause

A sentence unit that is embedded within a noun phrase, usually (but not always) introduced by a relative pronoun such as who or that:

The boy who stalked the prom queen has been caught.

The discovery that Copernicus made was controversial.

Notice that, confusingly, that can also be a complementizer introducing a sentential complement inside a verb phrase (see the entry “Sentence complement”).

Active versus passive voice

This is one of the most commonly misunderstood grammatical concepts. The active voice is simply the default structure for expressing an event or a situation:

Many theologians denounced Copernicus.

The principal drove the prom queen home.

The passive voice is an alternative way of expressing the same event:

Copernicus was denounced by many theologians.

The prom queen was driven home by the principal.

A passive sentence can only be related to an active voice sentence that contains a direct object. The passive sentence rearranges the order of the noun phrases so that the direct object from the active version becomes the subject of the passive version, and the subject from the active sentence (usually the instigator of the action) becomes embedded in a prepositional phrase introduced by the word by. In the process, the auxiliary verb to be (or occasionally to get) is inserted between the subject and the verb, and the verb appears in the past participle form (for example, drove becomes was driven; past tense verbs that end in -ed appear in exactly the same form as past participles—more on that later in the chapter). Thus:

The orangutan is being fed too often by the staff.

Copernicus got arrested for his controversial ideas.

The prom queen is chosen by the students every year.

A noteworthy feature of passive sentences is that the by phrase may be dropped, thereby leaving the instigator of the action completely implicit, unspoken, or unknown:

The orangutan is being fed too often.

Mistakes were made.

Copernicus was denounced.

Many people mistake other structures for passive sentences because they contain superficial similarities such as the presence of to be. But if a sentence does not contain all of the elements described above, it is not passive. Try identifying the passive sentences among the following examples:

(a) The maid is stealing money.

(b) Those ideas were considered revolutionary.

(c) The prisoners were being tortured on a daily basis.

(d) Now you are just being picky.

(e) Copernicus got famous for his controversial ideas.

(f) Every car is searched at the border.

For many relative clauses, it’s possible to drop the relative pronoun (such as that or who), leading to a reduced relative clause. When this happens to a relative clause that’s in the passive voice, the auxiliary to be (or to get) also gets dropped:

The astronomer (who got) arrested for his ideas was controversial.

The prom queen (who was) driven home by the principal was drunk.

Answer to the passives quiz: The passive sentences are (b), (c), and (f).

Visiting relatives can be annoying if they overstay their welcome.

The government plans to raise taxes were defeated.

The army houses soldiers and their families.

Once you’ve identified the potential ambiguities, you can check your interpretations against Figure 9.2, where some of these examples are graphically mapped out.

WEB ACTIVITY 9.2

Ambiguity in music In this activity, you’ll see how musical structure, like sentence structure, often requires the listener to disambiguate between competing interpretations.

https://oup-arc.com/access/content/sedivy-2e-student-resources/sedivy2e-chapter-9-web-activity-2

Measuring processing difficulty

Many garden path sentences are easy enough to identify because the reader becomes aware of some processing tangle or even of an initial misreading. But psycholinguists often want to study subtler ambiguity effects that bypass conscious awareness. Once again, careful timekeeping is a useful tool for detecting mild (but real) processing disruptions.

A garden path effect is detected by measuring how long people take to read the disambiguating region of the sentence—that is, the first point at which only the correct interpretation is consistent with the unfolding sentence, with other alternatives having been ruled out. In the following examples, disambiguating regions are in bold:

Figure 9.2 Alternative interpretations for three ambiguous sentence fragments. (A) Two possible structures for The investigation revealed the error … . (B) Two ways of interpreting Sam ate the hot dog and the vegetables … . (C) Two structures for The government plans … . Notice that in (C), the ambiguity hinges on whether the word plans is analyzed as a noun or a verb.

The general presented copies of the report was aware of the problems.

The investigation revealed the error resulted from gross negligence.

For purposes of comparison, it’s often possible to create virtually identical sentences that don’t contain hazardous ambiguities, simply by introducing disambiguating function words such as those shown here in italics:

The general who was presented copies of the report was aware of the problems.

The investigation revealed that the error resulted from gross negligence.

LANGUAGE AT LARGE 9.1

Crash blossoms run amok in newspaper headlines

“VOLINIST LINKED TO JAL CRASH BLOSSOMS.” That’s the headline of a 2009 article that ran in the newspaper Japan Today, prompting an American editor on an online discussion forum about copyediting to query, “What’s a crash blossom?” In fact, the article was about the blooming musical career of a violinist whose father had died in a 1985 Japan Airlines crash. The headline is a garden path sentence rivaling the famous “horse raced” example in its inscrutability, and it involves a similar ambiguity of structure. It contains a reduced relative clause, and would be somewhat more understandable had it been written like this:

The violinist who has been linked to the JAL crash blossoms.

There’s also a second ambiguity contained in the word sequence JAL crash blossoms. Under the intended meaning, crash is supposed to be a noun, and blossoms is supposed to be a verb. But it’s tempting to read the phrase such that crash is acting as a modifier (much like an adjective), and blossoms is a noun. Here, a lexical ambiguity involving the syntactic category of the word blossoms contributes to a completely nonsensical way of structuring the sentence overall. (Another example of this is with a sentence like He saw her duck, where duck could be either a noun or verb, resulting in very different readings.)

Function words like who, that, and was can serve the practical function of disambiguating a sentence, but in many cases such words are, perversely, optional. When they get dropped, this can lead to potential syntactic ambiguity. And in newspaper headlines, where there is a premium on terse expression, the disambiguating words are often swept away.

Hilarious results frequently ensue. Spotting garden path sentences in headlines (or crash blossoms, as they’ve come to be called at the suggestion of journalist John McIntyre) is a favorite sport of language geeks, and you can find many examples discussed on the popular linguistics blog Language Log (just do a search for “crash blossoms” on the blog’s website). Here are a few examples of especially humorous crash blossoms that have been honored in the blogosphere:

POLICE INVESTIGATE DEATH BY BALZAC (Balzac turns out to be a town, not the famous nineteenth-century French writer)

CHINESE COOKING FAT HEADS FOR HOLLAND

MANSELL GUILTY OF MISSING BUSINESSMAN’S MURDER

SMOKING MORE DANGEROUS FOR WOMEN THAN MEN

SISTERS REUNITED AFTER EIGHTEEN YEARS IN CHECKOUT COUNTER

GREECE FEARS BATTER MARKETS AGAIN

BRITISH LEFT WAFFLES ON FALKLAND ISLANDS

MINERS REFUSE TO WORK AFTER DEATH

JUVENILE COURT TO TRY SHOOTING DEFENDANT

You might try keeping a log of the crash blossoms you find in headlines over the period of, say, a week. Some might be funny, like the ones above in which a second meaning is readily accessible. Others might be nothing more than uninterpretable word mash-ups. Then try rewriting the headlines so that you preserve as much of their meaning as possible while removing the hazardous ambiguities.

If the bold regions in the first pair of sentences (that is, the sentences containing the potential ambiguity) take significantly longer to read than the same bold regions in the second pair (the unambiguous versions), then a garden path effect is considered to have occurred and is taken as evidence that the reader ran into processing trouble from having initially been misled, at least in part, by the wrong interpretation of the ambiguous portion of the sentence. You can get a more detailed feel for the reading time task in Method 9.1.

METHOD 9.1

Using reading times to detect misanalysis

How do you know for sure that a sentence is hard to read because readers are deceived into taking the wrong fork in the interpretive road? As a psycholinguist, you have several options. Most of the work on ambiguity resolution has relied on measures of reading time. A cheap and low-tech way of gathering these is by means of a self-paced reading task. Subjects read through sentences on a computer one word or phrase at a time, pressing a button to advance through the sentence, while a program records the amount of time they spend reading each segment. One downside to this method is that it’s a bit unnatural, and breaks up the flow of reading by requiring subjects to keep pushing a button. Researches have tried to make the task a touch more natural by embedding it in a moving window paradigm, simulating somewhat the experience of bringing each word into focus by moving the eyes over the sentence. In this experimental situation, subjects first see the entire sentence with each character covered over by dashes:

self-paced reading task A behavioral task intended to measure processing difficulty at various points in a sentence. Subjects read through sentences on a computer screen, one word or phrase at a time, pressing a button to advance through the sentence. A program records the amount of time each subject spends reading each segment.

moving window paradigm A version of the self-paced reading task in which dashes initially replace each alphabetic character in a sentence, and participants press a button to successively “uncover” each portion of the sentence. This method of presentation simulates a fairly natural reading rhythm.

  	--- ------- --------- ------ -- --- ------ --- ----- -- --- -------.

Each button press reveals a portion of the sentence and covers up the previous one, so a sequence of button presses would reveal chunks like this:

  1.	The general --------- ------ -- --- -- --- ----- -- --- -------.

  2.	--- ------- presented copies -- --- ------ --- ----- -- --- -------.

  3.	--- ------- --------- ------ of the report --- ----- -- --- -------.

  4.	--- ------- --------- ------ -- --- ------ was aware -- --- -------.

  5.	--- ------- --------- ------ -- --- ------ --- ----- of the problem.

Even so, there may be a tendency for subjects to apply a steady rhythm to their button presses and artificially fit their reading into that time frame. When they hit an especially difficult phrase, they may have already pressed the button to advance to the next phrase, even if they haven’t fully processed the earlier phrase. This can lead to spillover effects, where the reading time that should be recorded for a particular phrase ends up being recorded for the following one instead, as subjects struggle to catch up in processing. This is a bad outcome, since so much of the theorizing about garden path sentences hinges on firm knowledge about exactly which regions of the sentence cause hiccups in processing. In general, there’s a concern that by breaking up the flow of reading, and in a way that doesn’t allow subjects to backtrack, you’re not getting the most accurate picture of how processing works in typical reading situations.

You can encourage a smoother flow of reading and also get a more detailed snapshot of reading with a slightly higher-tech version of the reading time measure. Subjects read sentences presented on the screen in their entirety while their eye movements are being recorded. This allows you, the researcher, to measure the amount of time readers spend on each region of interest. It also allows you to get a picture of how long subjects take reading each portion on their first pass, how often they backtrack and where, and how long they spend rereading portions of the sentence.

By and large, there’s a lot of overlap in the data gathered by these two methods—certainly enough that the self-paced reading task remains a viable, easy way to detect most robust garden path effects. But the finer detail of eye tracking can provide clues that might help distinguish between competing theories. For instance, as you’ll see in Section 9.2, different theories make different predictions about the nature of the parser’s initial analysis, as opposed to later revisions to that analysis. Some researchers have argued that first-pass reading times align neatly with the parser’s first guesses, while backtracking, or regressive eye movements, are informative about later revision processes. It’s doubtful that eye movement behavior lines up this cleanly with the inner workings of the sentence processing system, but the distinction between earlier versus later reading time can often add an extra dimension to the reading time profile.

Regardless of how they’re collected, interpreting reading times is tricky, because the measure is somewhat of a blunt instrument. All kinds of factors can cause increases in reading times, not just misanalyzing the syntactic structure of the sentence. Reading times can be elevated because uncommon words are used, because the sentence depicts an unexpected event, because a low-frequency structure is used (regardless of ambiguity), or because the sentence is somehow unnatural in a particular context. Ideally, in order to be able to attribute long reading times to misanalysis, you want to—wherever possible—compare sentences that contain an ambiguity to unambiguous versions of the same sentences, with the same meaning, containing the same content words. This is a standard design feature of experiments that rely on reading times. Usually, each subject sees only one version of each sentence, so the second version cannot be predicted, and reading times across subjects are compared.

9.1 Questions to Contemplate

1. How do we know that people attempt to interpret sentences incrementally, as they unfold in time?

2. What are some of the consequences of incremental sentence understanding?

3. What would count as evidence of a garden path effect—that is, evidence that a reader has initially interpreted an ambiguous fragment incorrectly?

9.2 Models of Ambiguity Resolution

Some of the most passionate discussions in psycholinguistics (often referred to as the “parsing wars”) have revolved around questions of ambiguity resolution and how to explain garden path effects. Part of me would like to say that this passion was fueled by a burning desire to equip writers and editors with a definitive list of do’s and don’ts of sentence construction—no doubt the skills of an entire generation of professional writers might improve as a result. But the reality is that researchers were mobilized into spending long hours (and indeed, many years) in studying ambiguity resolution because this area of psycholinguistics became the battleground for several “Big Ideas” in the field.

To work up to the Big Ideas, we start with a simple question: Why is it that one reading of an ambiguous string of words often seems to be so much more attractive than another? For instance, in the classic “horse-raced” example, why are we so tempted to read the sentence as a simple main clause, and why does it seem to be so hard to access the reduced relative clause reading? Notice that this kind of confusion is strikingly different from what happens in word recognition. Remember that with words, multiple candidates are usually activated in parallel, with competitors dropping off in activation over time until a winner remains standing. We don’t ever seem to have the experience of hearing the syllable can- and committing so strongly to the word candle that we can’t seem to recover and recognize the word as cantaloupe once it’s been fully uttered. But with full-blown garden path sentences, there’s a strong early commitment to one particular structure, a strategy that sometimes leads to epic failure when this commitment turns out to be wrong. When faced with a syntactic ambiguity, why don’t we just remain noncommittal until there’s enough information to help us decide on the right structure?

The garden path theory

Explanations for the existence of severe garden path effects have followed several very different approaches. One of the earliest accounts of these troublesome sentences is known as the garden path theory, proposed by Lyn Frazier and her colleagues (see Frazier & Fodor, 1978, for an earlier version; and Frazier & Clifton, 1996, for a reformulation). These researchers noticed that for many garden path sentences, people seemed to be drawn toward a structure that was simpler than what eventually turned out to be the correct structure. For example, if you look back at Figure 9.1, depicting the sentence fragment The horse raced past the barn, it’s evident that the more alluring structure—with its single clause and basic word order—is not as complex as the reduced relative clause, which involves a passive structure and two clauses. Frazier and her colleagues argued that a number of other garden path effects followed the same pattern (for example, look back at the sentence fragments diagrammed in Figure 9.2A). They proposed that when faced with an ambiguity between a simpler and a more complex structure, people have a strong preference for the simpler structure, and this causes an interpretation crash when the sentence reveals itself to be consistent only with the more complex structure.

garden path theory A theory of parsing that claims that an initial “first-pass” structure is built during comprehension using a restricted amount of grammatical information and guided by certain parsing principles or tendencies, such as the tendency to build the simplest structure possible. Evaluations of plausible meanings or consideration of the context only come into play at a later stage of parsing.

Elaborating on this observation, Frazier and colleagues suggested that, rather than activating multiple meanings at once (as is the case in word recognition) the parser computes only one structure and its associated meaning. In order to be able to interpret sentences incrementally, the parser needs to build sentences very quickly. To achieve this blazing speed, the parsing system builds whatever structure is easiest to build and runs with it. If the whole thing runs aground at some later point, then a reanalysis of the sentence is initiated.

It may have crossed your mind, though, that factors other than simplicity could influence how an ambiguous phrase is interpreted. For example, some interpretations might be more plausible than others, or fit in better with the preceding context. Wouldn’t this information be taken into consideration while assembling the structure of a sentence? According to the garden path theory, no, not during the parser’s first attempt at building structure. This may seem odd, since information about plausibility or contextual fit could be a great help in averting a parsing meltdown. Why would the parser ignore such useful information? This is where the Big Ideas come in.

On the face of it, a parsing system that disregards helpful information looks like a badly designed system because it would result in a number of otherwise avoidable errors—it just seems needlessly dumb. But it’s worth pointing out some intellectual historical context. At the time that psycholinguists began trying to explain how ambiguity resolution works, cognitive psychologists were finding that sometimes, human information processing does appear to be a bit dumb. Evidently, we humans are saddled with certain inherent processing limitations. Intelligent, thoughtful analysis consumes a lot in the way of processing resources, and it takes a lot of time. In the 1970s and 1980s, it was becoming increasingly apparent to psychologists that because of our limitations, we can easily deplete the processing resources that are needed for solving some difficult cognitive problems. As a work-around, we often rely on cognitive processes that are fast and easy but that fail to take into consideration all of the relevant information.

This notion had come to be a highly influential one. The work of well-known psychologists such as Amos Tversky and Daniel Kahneman (1974) had shown that in many situations, people rely on very quick but error-prone heuristics—that is, shallow but very fast information-processing shortcuts that often lead people to leap to “illogical” or incorrect conclusions based on very superficial cues. (As an example, consider this problem: The surface area of water lilies growing on a pond doubles every day. If it takes 24 days for the pond to be completely covered, in how many days is the pond half covered? If you answer 12, then you’ve fallen prey to a common heuristic. Some more examples are given in Box 9.2.) Cognitive psychologists began talking about an important distinction between these fast, automatic, but often buggy processes, as opposed to slow, deliberate, but ultimately more accurate thought machinery. They concluded that an enormous amount of human cognition depends on the faster but dumber cognitive processes, and that we have a limited mental budget to spend on the more “intelligent” ones, which require considerably more effort and processing time. Such distinctions continue to be important in psychology to this day. You can find an accessible overview of this body of research in Daniel Kahneman’s 2011 book Thinking, Fast and Slow.

heuristics Shallow but very fast information-processing shortcuts that often lead to incorrect conclusions based on superficial cues.

BOX 9.2

Two common psychological heuristics

Rather than do an exhaustive search through their store of knowledge and use all the cognitive tools that are available, people often resort to “quick and dirty” judgments that are efficient, but error-prone. Two common heuristics found by cognitive psychologists are the availability heuristic and the representativeness heuristic.

In the availability heuristic, people’s estimates of frequency are often based on how easily examples come to mind:

■ They may believe that more Americans have died in terrorist attacks since September 11, 2001, than have died from having furniture fall on them, despite the fact that the reverse is true. (News reports about death by bookcase do not appear in the news very often, and therefore it is hard to recall any instances of such deaths.)

■ They may believe that words that begin with “r” are more frequent than words that have “r” as the third letter, since it is easier to retrieve examples of words that begin with “r.”

■ They may overestimate how frequently they take out the garbage relative to their domestic partner, since it is easier for them recall instances in which they themselves completed the chore.

In the representativeness heuristic, people base their estimates of the likelihood of a category based on judgments of surface plausibility, disregarding the statistical probability of the category:

■ They may believe that a person who is described as a “shy poetry lover” is more likely to be a student of medieval literature than a student of business administration, even though business admin students far outnumber medieval lit students.

■ They may make logically nonsensical judgments based on representative events, such as estimating that (b) is more likely than (a), despite the fact that (b) is a subset of the cases described in (a):

(a) A massive flood will occur somewhere in North America next year in which more than 1,000 people drown.

(b) An earthquake will occur in California sometime next year, causing a massive flood in which more than 1,000 people drown.

The garden path theory is very consistent with this line of thinking and represents one of several attempts to apply the ideas about fast versus slow cognition to problems in language processing. The idea is that the parser is able to conserve processing resources by relying on a set of quick and simple structure-building heuristics in its first guess at sentence structure. In this first pass, a great deal of information that could be pertinent to resolving the ambiguity is ignored because this would take too much time and mental energy to integrate—in fact, if the parser did have to consider all potentially relevant information in the first-draft structure, doing so might strain memory capacity beyond the breaking point. All things considered, a fast-and-cheap solution may be the best strategy, even if it results in frequent errors; if an error is detected, a deeper analysis of the sentence is triggered.

If you’ve worked your way through the Digging Deeper discussion in Chapter 8, this theoretical approach should seem very familiar to you. There, I discussed the tension between modular and interactive models of word recognition. In a modular system, “higher-level” information doesn’t have a direct effect on “lower-level” processes. Instead, it comes into play at a later stage, once the lower-level processes have been completed. For example, in a modular model of word recognition, information about the context or lexical knowledge can’t affect the activation levels of individual sounds; it can only help decide which sounds are most likely or appropriate, once they’ve been activated from the bottom up. In an interactive model, on the other hand, information from the higher levels can flow from the top down to affect the activation of sounds.

When it comes to parsing, the garden path model is an instantiation of a modular system. Structures are initially built by a lean, fast parser on the basis of a very limited amount of purely syntactic information, and entirely without the benefit of semantic knowledge or knowledge about the context of the sentence. Once a partial structure is built, it gets sent off to the higher-level semantic interpretation component, which then evaluates its sensibleness. But no information about the meaning of the sentence fragment to that point can influence the parser’s first stab at structure-building. For some researchers, the division between syntactic and semantic knowledge could be seen as a division between fast, automatic mental processes and slow, thoughtful ones, with semantic information lagging behind syntactic processes (and sometimes cleaning up some messes in interpretation).

In Chapter 8, I suggested a metaphor for thinking about the distinction between modular and interactive systems: A highly modular system is structured like a company or factory in which lower-level employees do their jobs with very limited knowledge and responsibilities, and then simply pass on their work to higher-level workers who later make all the important decisions. On the other hand, an interactive system is more like a company in which all levels of workers have a shared stake in making decisions, with information flowing freely from upper to lower levels. As in business, the competing models offer different advantages: the first option may be fast and cheap, but lacks a certain flexibility (which you may have experienced, if you’ve ever asked a customer service representative to deal with a problem outside of his area of expertise). The second option is “smarter” in some ways, but maybe not so cost-effective on a large scale.

The constraint-based approach

In the garden path model, because the parser initially spits out a single preferred structure in the face of ambiguity, it’s expected that people will often run into an interpretive dead end and be forced to reconsider the sentence’s meaning; this is an unavoidable by-product of the parser’s efficiency. But quite a few researchers have argued that the parser actually shows a lot more subtlety and intelligence in its initial analysis than the garden path model is willing to give it credit for. The core of the argument is that people experience much less confusion in processing sentences than you’d expect if they were blindly computing the “preferred” structure without considering other potentially helpful information.

According to the early version of the garden path account, the parser only looks at the syntactic categories of words and the language’s rules for assembling them, building the fastest and easiest structures it can. But in that case, all sentences that make use of dispreferred structures (where easier, or preferred, ones are available for the same string of words) should be equally likely to lead to a garden path effect, because the wrong structure is initially built. But this doesn’t seem right. For example, compare these two sentences:

The dog walked to the park wagged its tail happily.

The treasure buried in the sand was never found.

Both sentences involve reduced relative clauses. Most people find the first far more difficult to interpret than the second (if your own parser has become benumbed to the distinction as a result of reading too many garden path sentences in the last half hour, try them out on a fresh reader). This became one of the key arguments against the garden path theory, as dissenting researchers began to look for evidence that similar structures resulted in extremely variable garden path effects. It would be especially problematic for the garden path theory if some sentences saddled with dispreferred structures, such as reduced relative clauses, didn’t show any measurable garden path effects at all. For example, if we compared the following two sentences, we might not find any difference in reading times for the region in bold:

The treasure buried in the sand was lost forever.

The treasure that was buried in the sand was lost forever.

This would suggest that in the first sentence of the pair, which begins with a potentially ambiguous string of words, readers don’t seem to have been misled by the other (supposedly preferred) interpretation, and interpretation goes as smoothly in reading the final portion of the text as if there had been no ambiguity at all.

In one influential paper (MacDonald et al., 1994), the authors used a witty expository device to make the point that reduced relative clauses don’t always cause interpretive snarls. They pulled out many examples of reduced relative clauses from the writings of various psycholinguists, some of whose theories in fact did predict that such structures would invariably cause garden path effects. Some of these examples are reproduced in Box 9.3, along with a few other examples I dug up myself from various sources.

The main competitor to the garden path theory was the constraint-based approach. It parted ways with the garden path account on a number of points. First, constraint-based theorists argued that a broad range of information sources (or constraints) could simultaneously affect the parser’s early decisions, including semantic or contextual information, which was often thought of as being too subtle or slow to drive initial parsing (specific constraints will be discussed in detail in Section 9.3). Parsing speed, they argued, is not necessarily bought at the expense of subtlety or intelligence in processing. As a result, only a very small percentage of temporary ambiguities might cause any discernible processing difficulties.

constraint-based approach The main competitor to the garden path theory, this approach claims that multiple interpretations of an ambiguous structure are simultaneously evaluated against a broad range of information sources (or constraints) that can affect the parser’s early decisions.

Second, they argued that syntactic ambiguity resolution actually looks very much like ambiguity resolution in word recognition. In both cases, multiple interpretations are accessed in parallel, and our minds very quickly try to judge which one is more likely to be correct based on very partial linguistic evidence, along with various sources of information—such as context, or statistical frequencies—that make one interpretation seem more likely to be the right one. These sources of information (or constraints) can serve to either ramp up or suppress the activation of one interpretation relative to the other.

BOX 9.3

Not all reduced relatives lead to processing implosions

We’ve seen how a relative clause like the horse [that was raced past the barn] can become ambiguous and potentially impede sentence processing when the clause is “reduced” by removing function words (becoming the horse raced past the barn). But are such reduced relatives always a problem for the person processing them?

Maryellen MacDonald and her colleagues (1994) discovered examples of reduced relative clauses in the writings of various psycholinguists, many of whom were proponents of theories predicting that reduced relatives invariably lead to garden path effects. MacDonald’s point was not that these psycholinguists are bad writers, oblivious to the fact that their sentences might cause readers to stumble, but (more damningly) that their theories are flawed, since these sentences seem to pose no real burden for readers. Here’s a sampling of the sentences unearthed by MacDonald et al., with the reduced relative clauses in color type (color was added for this textbook, not by the authors of the original papers):

Thus, a noun phrase followed by a morphologically ambiguous verb (e.g., “The defendant examined”) will be temporarily ambiguous.

Trueswell, Tanenhaus & Garnsey, 1994, p. 287

Referential information provided by the discourse is of no help.

Britt, Perfetti, Garrod & Rayner, 1992, p. 305

Recent research reported by Ferreira and Clifton (1986) has demonstrated that syntactic processing is quite independent and that the initial syntactic analysis assigned to a sentence is little influenced by the semantic information already analyzed.

Frazier & Rayner, 1987, pp. 520–521

In all cases, the examples cited here were not the only reduced relatives in these articles.

MacDonald et al., 1994, p. 678

Of course, the use of reduced relatives is not limited to academic writers. Here are some examples I discovered among the wilds of the popular print media:

The reward money offered to find Jyrine, a helpless toddler, was a paltry $1,000.

“What happened to my child?” Essence, September 2007, p. 224

The food prepared for the now non-existent media event is donated to homeless shelters.

“911 Coverage of Sept. 11 attacks overshadowed a world of events,” Denver Post, October 11, 2001

Of the two reduced relatives in the following example, only the first one (italicized) is potentially ambiguous. Neither seems to pose any difficulty for processing:

… many Western Europeans believe that a job offered to an older worker is a slot taken away from a younger one.

Foreign Affairs, May/June 2007, p. 55

This final, lovely example has one reduced relative clause nested inside another (the embedded clause is italicized):

… many Russian politicians influenced by more than 40 years of communist propaganda aimed against the west finally became good “Homo Sovieticus.”

Letter to the editor, San Francisco Chronicle, June 6, 1995

So why does it often seem that we consider only one interpretation of a sentence, with the other seeming to be completely inaccessible? The response of constraint-based theorists was that flagrant garden path effects are in fact the results of an unfortunate coincidence: the various sources of information that are available in the ambiguous portion of the sentence just happen to overwhelmingly point to the wrong interpretation, causing the parser to suppress the competing correct one, much as a lexical competitor is suppressed over time. By the time readers or hearers get to the disambiguating region, it’s too late—they’ve already pushed the eventually correct alternative beyond accessibility.

In other words, the idea is that a sentence like The horse raced past the barn fell starts out activating both interpretations, but since all the available constraints in the ambiguous region are heavily biased toward the simple main clause, the reduced relative reading is too weak to ever become a serious candidate and becomes deactivated. On the other hand, things are a bit different in an easier sentence like The treasure buried in the sand was never found. In this case, the constraints are more equivocal, creating a more even balance between the two alternatives. When the disambiguating region is reached, the new information in that region provides the decisive evidence between the alternative meanings (both of which are still active), and readers end up settling on the correct meaning without too much difficulty.

This is really a dramatically different view of the nature of ambiguity effects than the one offered by the garden path model. For the garden path account, comprehension mishaps arise because the inherent limitations on human cognition force the parser to make decisions about structure without considering all the evidence, so these decisions will often be wrong. For the constraint-based model, the parser has considered all the evidence and has made a rational prediction based on prior experience with language—it’s just that in some cases, the outcome turns out to fly in the face of all reasonable predictions.

9.2 Questions to Contemplate

1. What ideas and insights from psychology motivated the garden path model, and how well do you think these insights fit with the problem of parsing?

2. How does the constraint-based model explain devastating garden path effects, such as the difficulty with The horse raced past the barn fell?

3. What kind of evidence would you look for to support the garden path model?

WEB ACTIVITY 9.3

When do ambiguities pose processing problems? In this activity, you’ll see a number of examples of temporary ambiguities. You’ll informally gauge the difficulty of each example—does the degree of processing difficulty seem variable, even for structurally similar sentences? You’ll be prompted for potential explanations for your intuitions.

https://oup-arc.com/access/content/sedivy-2e-student-resources/sedivy2e-chapter-9-web-activity-3

9.3 Variables That Predict the Difficulty of Ambiguous Sentences

Let’s turn to testing the predictions made by each model. The garden path account predicts that the parser will crash whenever it’s forced to build a structure that falls outside of the preferred structure-building routines—it will always build the wrong one first. But the constraint-based account says that garden path effects will be highly variable, even for the so-called dispreferred structures. Processing difficulty may range from devastating to nonexistent, depending on the specific words that appear in those structures, or even the specific context of the sentence. This is because the extent to which one reading over another will be favored should depend on the biasing strength of all the constraining sources of information together, of which general structural information is but one source. For the garden path theory, only structure should determine the presence of a garden path effect, since this is the only source of information that’s considered in the early phases of parsing. Garden path effects, then, should be impervious to other sources of information. To evaluate the predictions of the garden path and constraint-based theories more concretely, let’s look more closely at the varied bouquet of information sources that bear on the resolution of structural ambiguity.

Thematic relations associated with verbs

Our knowledge about verbs includes knowing the kinds of events they describe, or their thematic relations—that is, information about how many and what kinds of participants must be involved, and what roles the various participants play. For example, when we hear the verb bite, we know that a biting event involves at least two participants—the biter and the bitee. We also know that the biter must be an animate entity in possession of teeth and the ability to close those teeth around something. The bitee, on the other hand, could be either another animate entity or an inanimate object: you can bite your brother, but you can also bite your brother’s finger or a piece of toast. A parser that had access to this semantic information (as predicted by the constraint-based account) could make much smarter guesses about the likely structure of ambiguous word strings in at least some cases.

thematic relations Knowledge about verbs that captures information about the events they describe, including how many and what kinds of participants are involved in the events, and the roles the various participants play.

Consider a sentence like The treasure buried in the sand was lost forever. The verb bury involves at least two participants, one doing the burying, and one being buried. Which of these participants lands in subject position depends on the sentence’s structure. The main clause structure requires the burier to appear as subject; but treasure is a terrible fit for this role. So, as soon as the reader gets to The treasure buried …, there’s semantic pressure to shift to the alternative reduced relative clause reading, in which the buried entity appears as subject. This means the reader is less likely to be led down the garden path to an interpretation that later blows up.

But thematic relations aren’t quite as helpful in a sentence like The dog walked to the park wagged its tail happily. The dog fits perfectly well here as the participant who is doing the walking—which is what the main clause structure requires in subject position—so there’s nothing to force the reader toward the reduced relative reading. Unfortunately, the main clause interpretation turns out to be wrong.

The constraint-based model predicts that when information from thematic relations is strong enough to steer people away from the normally dominant reading of a structure, it should diminish or eliminate a garden path effect, as measured using reading times. In fact, this is what a number of studies have found (e.g., Trueswell et al., 1994).

The syntactic frames of verbs

Your knowledge of verbs also includes the idiosyncratic preferences of verbs for specific syntactic frames. You can’t just put any verb into a generic verb slot in a sentence structure and expect things to turn out well. Hence, the ungrammaticality of:

*The soldier buried.

*Alice fell the ball.

*Samantha sneezed that Billy was in prison.

*The mom put the cookies.

*The mom put the cookies about the jar.

*Frank said the report.

Some verbs, like fall, have to be intransitive verbs, with just a subject and no direct object. Some (for example, buried) are transitive verbs, with a subject and a direct object. Some (for example, put) are ditransitive verbs, calling for both a direct object and an indirect object (which may be introduced by a preposition). Finally, some (like said) are sentential complement verbs, which introduce a clause rather than a direct object noun phrase. (See Table 9.1 for some additional examples.) Through your experience with language, you’ve learned which verbs can be slotted into which frames—and whether a verb can appear in multiple frames and, if so, under what circumstances. (The verb say, for example, can be transitive, but selectively so, as in Tom said his prayers, Marta said her piece, etc.).

intransitive verbs Verbs that occur with a subject but no direct object.

transitive verbs Verbs that take both a subject and a direct object.

ditransitive verbs Verbs that occur with a direct object and an indirect object (which may be introduced by a preposition).

sentential complement verbs Verbs that introduce a clause rather than a direct object noun phrase (NP).

The garden path model (at least in its original incarnation) claims that in its first pass, the parser only cares about identifying coarse grammatical categories, leaving aside these fancy details about verbs’ syntactic frames. But constraint-based advocates have countered that the parser has access very early on to information about the syntactic frames that are linked with specific verbs. Access to such information would lead to much “smarter” parsing decisions. Notice that walk is allowed to appear with or without a direct object, whereas bury absolutely can’t get by without one:

TABLE 9.1 A variety of syntactic frames for verbs

Intransitive verbs (appear with a subject only)

The neighbors’ dog reeks.

Gerhardt finally relaxed.

Mariah sings beautifully.

The magician vanishes.

The tornado touched down last night.

Transitive verbs (select for a direct object noun phrase)

This theory vexes me.

I rarely wear socks.

The author signed his book.

The queen poisoned the courtesan.

The engineer inspected the plans.

Ditransitive verbs (select for two noun phrases, one of which may appear inside a prepositional phrase)

Please lend me the money.

Parents should send their teens to boarding school.

Devon presented his fiancée with a ring.

Frances hid the affair from her husband.

The teacher tossed Alisha the ball.

Sentential complement verbs (select for a clausal unit)

The workers complained that their bosses harassed them.

I agree that the king should be deposed.

The president assumed his staff would cover his mistake.

The teacher regretted that she had promised the kids a dollar for every book they read.

Verbs that fall into more than one category^a

NP-bias verbs: e.g., accept, repeat, advocate, maintain, reveal, print

S-bias verbs: e.g., conclude, decide, promise, worry, prove

^aMany verbs fall into more than one of the four categories, but may occur much more often in one syntactic frame than another. The verbs given as examples here can appear with either a sentential complement (S) or a direct object noun phrase (NP), but have a bias for one or the other. Try constructing sentences with both kinds of frames for each verb.

Raj walked his dog to the park.

The dog walked.

The dog walked to the park.

The soldier buried the land mine.

*The pirate buried.

*The pirate buried in the sand.

Based on this knowledge, upon getting The treasure buried in the …, the parser would be able to avoid the normally seductive main clause reading, since this reading lacks the obligatory direct object that bury demands. This would leave only the reduced relative clause interpretation as viable because, in this structure, the syntactic process of passive formation has taken the treasure (i.e., what would be the direct object in the active voice) and stuck it into the subject position. A fair bit of experimental evidence has shown that knowledge of syntactic frames plays a central role in ambiguity resolution (e.g., Tanenhaus et al., 1989). And, as you’ll see shortly, this knowledge about syntactic frames is fairly nuanced, not just limited to what’s grammatical or not, but also tuned in to which syntactic frame is most statistically likely for a given verb.

Frequency-based information

As you saw in Chapter 8, frequent words are recognized more quickly than less common words. It also stands to reason that more common structures would be easier to build than less common ones, just as a building crew would be more efficient in building a house from a plan that they’ve used many times before than from one that they’ve used less frequently. Could some portion of garden path effects be due to different frequencies of the alternative structures? For example, are reduced relative sentences so hard in part because their structure is much less common in English than the competing structure of simple past-tense main clause?

To ask this question, we can compare similar structures across languages that happen to have different patterns of frequency for these structures. Take the following sentence:

Someone shot the maid of the actress who was standing on the balcony with her husband.

Who was standing on the balcony with her husband? Most English speakers say it was the actress, preferring to attach the relative clause to the nearby noun. But a funny thing happens when the sentence is translated almost word for word into Spanish:

Alguien disparó contra la criada de la actriz que estaba en el balcón con su marido.

The majority of Spanish speakers say that it was the maid who was on the balcony with her husband. This is interesting because, as it happens, relative clauses in Spanish attach more frequently to the first of two nouns, while the reverse is true in English. English and Spanish readers, then, seem to be resolving the ambiguity in a way that reflects their personal experience with these structures (Mitchell & Cuetos, 1991).

Parsing also seems to be sensitive to frequency-based information related to specific verbs and the structural frames in which they commonly appear. Many verbs, for example, can appear as either transitive or intransitive, or with or without prepositional phrases—but they often appear much more commonly in one syntactic frame than another. Frequency-based expectations about syntactic frames can feed directly into the structures that are built on the fly by the parser. An especially ingenious case study showing effects of syntactic frame frequencies is presented in Researchers at Work 9.1.

RESEARCHERS AT WORK 9.1

Subliminal priming of a verb’s syntactic frame

Source: J. C. Trueswell & A. E. Kim (1998). How to prune a garden path by nipping it in the bud: Fast priming of verb argument structure. Journal of Memory and Language 39, 102–123.

Question: Does reading a verb immediately activate its associated syntactic frame in a way that can bias the interpretation of a syntactic ambiguity?

Hypothesis: The rapid (subliminal) presentation of a verb that is biased toward one syntactic frame (e.g., a direct object) will immediately activate that frame and influence the subsequent processing of an ambiguous string, such that continuations that are consistent with the preferred frame will be read faster than continuations that are inconsistent with that frame.

Test: Thirty-two participants read sentences (16 target sentences and 54 distractor sentences) presented in a self-paced reading format, as described in Method 9.1. Half of the target sentences contained a temporary ambiguity that hinged on interpreting the noun after that verb as either a direct object or as the subject of an embedded clause (i.e., sentential complement). For example: The talented photographer accepted the fire … could in theory continue with a direct object interpretation:

The talented photographer accepted the fire from his fellow camper.

However, all critical sentences continued with a sentential complement:

The talented photographer accepted the fire could not have been prevented.

The remaining half of the sentences used a grammatical word to mark the upcoming embedded clause, eliminating any potential ambiguity:

The talented photographer accepted that the fire could not have been prevented.

As the subjects read through the target sentences, the main verb (e.g., accepted) was preceded by the very brief (39 ms) presentation of a “priming” verb. The presentation of this verb was so rapid that subjects were not able to consciously recognize it, experiencing it as a brief flash that appeared before the main verb. Half of the priming verbs were statistically biased toward a syntactic frame in which they took a direct object (e.g., obtained; as in obtained the documents yesterday), while the other half were biased toward a frame in which they were followed by a sentential complement (e.g., realized; as in realized the documents were fake). The ambiguity manipulation was crossed with the priming verb manipulation, yielding the following four conditions, with four sentences in each condition:

Ambiguous sentence: Priming verb biased toward direct object

Unambiguous sentence: Priming verb biased toward direct object

Ambiguous sentence: Priming verb biased toward sentential complement

Unambiguous sentence: Priming verb biased toward sentential complement

Results: The critical findings involved the magnitude of the garden path effect, that is, the difference in reading times between the unambiguous and ambiguous sentences versions of the sentence at the region marked in bold:

The talented photographer accepted the fire could not have been prevented.

It’s at this point that readers are forced to give up any expectations they may have had that the ambiguous sentence would continue with a direct object.

Overall, readings times at this word were longer for the ambiguous sentences relative to the unambiguous sentences, suggesting that participants were surprised to some extent to encounter the sentential complement continuation. However, this surprise, or garden path effect, was stronger when the priming verb had been biased toward the direct object frame (e.g., obtained) than when the priming verb was biased toward a sentential complement frame (e.g., realized) as shown in Figure 9.3.

Conclusions: Recognizing a verb involves activating its associated syntactic frame; once a syntactic frame is active, it can immediately be used to help resolve a temporary ambiguity.

Questions for further research

1. Does the effect of the priming verb’s syntactic frame depend on participants’ linguistic experience? For example, would the same effect show up if the participants were second-language learners of English or had relatively poor vocabulary knowledge?

2. Does using the information about a verb’s frame require cognitive effort? That is, would the effect of the verb be reduced if participants’ cognitive resources were taxed—for example, by requiring them to remember a 3-digit number while reading the target sentences?

Figure 9.3 Garden path effects in milliseconds for each word position, calculated as reading times for ambiguous sentence continuations minus reading times for unambiguous sentence continuations. (From Trueswell & Kim, 1998, J. Mem. Lang. 39, 102.)

In addition to the frequency of syntactic frames, people are also sensitive to how frequently verbs appear in the passive or active voice. Some verbs (such as entertained) rarely show up in passive structures (e.g., Hamlet was entertained by the actors), whereas others (such as accused) more commonly do (The writer was accused of plagiarism). Thus, we might expect entertained to lead to a much stronger bias for the main clause interpretation than accused, resulting in a stronger garden path effect for the first than for the second of these two sentences:

The audience entertained at the gala left in high spirits.

The suspect accused at the crime scene was soon released.

Several researchers (e.g., Trueswell, 1996; MacDonald et al., 1994) have indeed found that the severity of garden path effects for reduced relative clauses is affected by the frequency of past-tense/past-participle readings of ambiguous verb forms.

The importance of context

Let’s return to that famous racing horse, and the ugly sentence in which it appears. What happens when the sentence shows up in a story like this:

Farmer Bill and Farmer John were racing their horses through the field. Farmer Bill rode his horse along the fence, while Farmer John raced his horse past the barn. Suddenly, the horse raced past the barn fell.

Does the last sentence of the story suddenly seem less difficult? If so, why would that be?

Several researchers (e.g., Altmann & Steedman, 1988; Crain & Steedman, 1985) have pointed out that attaching a modifier phrase (such as a relative clause or a prepositional phrase) usually only makes sense when that modifier is needed to pick out one of two or more similar entities—here, to distinguish between two horses. In this story, when the final sentence is encountered, there’s pressure to interpret the phrase raced past the barn as providing some more information about the horse; without this modifier, the simple noun phrase the horse wouldn’t provide enough information to distinguish between the two horses, leading to confusion about which one was being referred to. But in a scenario that doesn’t involve two horses—for instance, in a story about a single horse, or an out-of-the-blue sentence where the presence of two horses hasn’t been established—the use of the modifier is unnatural. Speakers and writers don’t generally go around providing extra, unnecessary information in the form of a modifier phrase. Without the proper contextual support for relative clause reading, readers will be more likely to default to the alternative main clause reading.

Reading time studies have shown that the right context can reduce or eliminate garden path effects, even for complex structures such as a reduced relative clause (e.g., Spivey-Knowlton et al., 1993). But the effects of context have been seen most vividly in studies of spoken language in which a visually present context provides a strong constraint for interpretation. In these experiments, researchers have used evidence from eye movements to infer the interpretations that are unfolding in the minds of listeners as they listen to ambiguous language.

As you’ve seen in Chapter 8, when people look at a visual scene while listening to spoken language, they try to relate the scene in front of them to what they’re hearing. Since people look at images that correspond to the words they think they’re hearing, this makes it possible to track eye movements to the visual scene as a way of inferring which word(s) they’re activating based on spoken input. It takes a bit more ingenuity, but it’s also possible to set up visual displays to test between competing structural interpretations, as demonstrated by Michael Tanenhaus and his colleagues (1995). For example, imagine sitting at a table with the objects shown in Figure 9.4A and imagine hearing an instruction like, “Put the apple on the towel into the box.”

Where do you think you’d be looking at various points in the sentence? Naturally, upon hearing apple, people tend to look at the only apple in the display. But when they hear towel, there are now two possibilities. Which towel should they look at? It depends on which sentence structure they’re entertaining. If they’ve attached on the towel directly to the verb phrase, they’ll interpret the instruction as meaning they should pick up the apple and place it on the empty towel since it makes no sense to put the apple on the towel that it’s currently sitting on. Of course, they’ll hit trouble when they get to into the box, since they’ll already have interpreted the empty towel as the intended destination. On the other hand, if they’ve attached on the towel to the noun apple (specifying which apple should be moved), then they’ll continue looking in the upper left square, which contains the object that corresponds to this noun phrase. They’ll then have no trouble integrating into the box, since they’re still waiting at this point to hear the destination phrase that will attach to the verb phrase. Critically, there should be no need to look at the empty towel, since this object is irrelevant for carrying out the instruction as understood. In other words, it’s possible to use eye movements to the irrelevant towel as evidence of a garden path interpretation.

Figure 9.4 Visual displays and eye movement patterns used in the 1995 study by Tanenhaus et al. (A) Visual display with only one referent corresponding to the word apple. Letters indicate the typical sequence of eye movements and their timing relative to the spoken instruction. (A' and B' correspond to the unambiguous version of the instruction.) (B) Visual display with two referents corresponding to apple. Note that for this display, the sequence and timing of typical eye movements (relative to critical words in the speech stream) are the same, regardless of whether the instruction was ambiguous or unambiguous. (Adapted from Tanenhaus et al., 1995, Science 268, 1632.)

And evidence for a garden path is exactly what you see when the ambiguous sentence Put the apple on the towel into the box is compared with its unambiguous version Put the apple that’s on the towel into the box. In the ambiguous version, people look at the irrelevant towel more often than in the unambiguous version (see Figure 9.4A).

This method turns out to be especially sensitive to the effects of context, particularly if that context can be set up visually so that hearers don’t have to hold it in memory, as they might with a story context. Remember that attaching extra information to the noun is most natural in a situation where it’s needed to distinguish between more than one similar entity. It’s perfectly easy to visually set up just such a context, as shown in Figure 9.4B, where there are now two apples in the scene, only one of which is on a towel. In this visual context, when subjects hear Put the apple on the towel, there’s a strong incentive to interpret the phrase on the towel as specifying which of the two apples is to be moved.

As it happens, the effects of the visual context are powerful; when faced with a scene like the one in Figure 9.4B, upon hearing Put the apple on the towel into the box, subjects very rarely look at the empty towel—no more so, in fact, than when they hear the unambiguous version of the instruction. Hence, the eye movement record shows no evidence of a garden path interpretation when the context is appropriate to the normally less preferred structure.

Even more subtle effects of context have been demonstrated by Craig Chambers and his colleagues (2004). In their experiment, they set up displays such as those illustrated in Figure 9.5, simulating a cooking environment in which objects needed to be manipulated according to spoken instructions such as, “Pour the egg in the bowl over the flour” (or its unambiguous counterpart, “Pour the egg that’s in the bowl over the flour”). You’ll notice that in both versions of the visual display, there are two eggs, so the phrase in the bowl is potentially helpful in identifying which of the two eggs is the target of the instruction. However, in Figure 9.5B (unlike in Figure 9.5A) the display contains only one egg in liquid form that could possibly be poured—the other egg is still inside its shell. Hence, upon hearing Pour the egg, the subject might infer that since the action could only target one of the eggs, there’s no need for additional information to specify which of the eggs is under discussion. If hearers are indeed making these inferences based on the visual context, then they might be tempted to misinterpret the ambiguously attached phrase in the bowl as designating the location where the egg is to be poured, rather than as further information about the targeted egg. Such an interpretation would be revealed in increased glances to the empty bowl, the “false goal” in the display. In fact, Chambers and his colleagues showed that just such a garden path effect did emerge in the visual displays with a single pourable egg (see Figure 9.5C). However, when two pourable eggs appeared in the display, there was no evidence of misinterpretation, with eye movements being similar for both the ambiguous and unambiguous versions of the instruction. This shows an impressive ability to closely integrate the incoming stream of speech with detailed aspects of the visual context. Contrary to the predictions of the garden path model, people seem to be able to recruit context quickly enough to avert garden path effects for a number of syntactic ambiguities.

Figure 9.5 Sample visual displays and eye movement data from Chambers et al. (A) Display in which both referents corresponding to egg are compatible with the action described in the instruction “Pour the egg in the bowl over the flour.” (B) The second referent corresponding to egg (i.e., a hard-boiled egg) is incompatible with the instruction. (C) The study’s results show that when the second referent is incompatible with the action, participants are less likely to interpret the phrase in the bowl as helpfully distinguishing between the two potential referents; instead, they interpret the ambiguous phrase as referring to the goal location, as evident from the increased amount of time spent looking at the false goal (the empty bowl). (Adapted from Chambers et al., 2004, J. Exp. Psychol. Learn. Mem. Cogn. 30, 687.)

You might be inclined to wonder just how often the need for such context-sensitive inferences arises in actual spoken language. After all, isn’t it true that we can communicate our intended meaning through intonation, even when a given string of words has more than one meaning? This is a common intuition, but while intonation can sometimes help to disambiguate speech, it certainly doesn’t sweep away all or even most potential ambiguities (see Box 9.4). Rather, intonation seems to be just one more useful cue among many that we can use intelligently to help us untangle strings of words and their many potential meanings.

BOX 9.4

Doesn’t intonation disambiguate spoken language?

Many people assume that in spoken language, unlike in written language, true ambiguity simply doesn’t arise; the belief is that while a string of words may be ambiguous, the way that you say those words will be different depending on your intended meaning, allowing your hearer to avert any possible misinterpretation. And in many cases, this is true. For example, I’m sure that you can come up with six or seven different ways to utter the word really, each of them clearly communicating something different from the others.

But let’s look at the issue a little bit more carefully in the context of the garden path sentences that we’ve come to know and love. Here, the situation is a bit different from your ability to clarify your intended use of the word really—in that case, intonation was serving to signal a general communicative purpose such as asking a question, expressing incredulity or emphasis, or indicating doubt. In the case of garden path sentences, though, intonation is being asked to signal one structure rather than another. Under what conditions would it be possible for intonation to disambiguate structure? (Actually, it would be more accurate to talk about the disambiguating potential of prosody, a notion that includes intonation, but also rhythmic aspects of speech such as pausing and lengthening of sounds.)

prosody The rhythm, stress, and intonation of a spoken phrase or sentence.

One possibility might be that particular sentence structures match up with specific prosodic patterns in a one-to-one fashion. But this turns out not to be true. The relationship between syntax and prosody is much more slippery than this. There’s a general tendency, for example, to insert pauses or stretch out the last syllable of a phrase just before the boundary between major constituents (you saw evidence of this in Language at Large 6.1, in our discussion of constituents in literary language). For example, ambiguities that revolve around whether a clause is to end or to continue lend themselves especially well to prosodic disambiguation. Consider the following ambiguous fragment:

While the pop star sang the national anthem…

which, as you’ve seen, could continue like this:

… played in the background

or it could continue like this:

… the people stood respectfully

In the first continuation, the desire to insert a pause is very strong, in keeping with the convention of inserting a comma:

While the pop star sang, the national anthem played in the background.

But boundaries between constituents that are smaller than clauses are marked with much subtler prosody—and in fact, whether or not they’re marked at all can depend on factors such as the length of the phrases that are involved, so you’d be more likely to mark the boundary here:

The patient told the doctor [S that he was having some ongoing trouble with his injured knee].

than here:

The patient told the doctor [S that he smoked].

But even if prosodic patterns don’t match up neatly with syntactic structures, it’s still possible that speakers would choose to use cues like pauses and syllable lengthening as a courtesy to the hearer, in order to mark the correct way in which the words should be grouped into constituents. This, too, turns out to be not quite true. A number of studies have shown that speakers fail to reliably disambiguate speech for their hearers. For example, David Allbritton and his colleagues (1996) recorded the speech of trained actors as well as psychology undergraduates while they read sentences aloud, and found that neither group spontaneously disambiguated sentence structure through prosody. When instructed, the actors at least were able to use prosody to signal their intended meaning, but only when both meanings of the ambiguous sentence were presented to them side by side. This finding is consistent with a number of other similar experiments.

Things get a bit better when speakers and hearers are involved in an interactive task such as a game in which the speaker instructs the hearer to move objects around on a board. For example, Amy Schafer and her colleagues (2000) found that under these circumstances, when speakers became aware of their partners’ misinterpretations, they did use prosody somewhat consistently to disambiguate their instructions. But overall, the evidence suggests that in many linguistic settings, it’s quite difficult for speakers to anticipate ambiguities that will prove difficult for their hearers and to be helpful in the prosodic patterns that they produce.

Still, some systematic alignment between syntax and prosody does occur, probably as a by-product of the processes that are involved in planning and uttering speech, as you’ll see in Chapter 10. Prosody is more helpful for some structures than for others, but there’s enough information there to be of some use to hearers for the purposes of disambiguation, as demonstrated by Tanya Kraljic and Susan Brennan (2005). So, while prosody certainly doesn’t render spoken language unambiguous, hearers can use it as one more helpful cue, along with the many other sources of information discussed in Section 9.3.

WEB ACTIVITY 9.4

Does prosody disambiguate? In this activity, you’ll listen to sound files of speakers uttering ambiguous sentences. You’ll be asked to determine whether you can reliably understand the intended meanings of the sentences based on the speaker’s prosody.

https://oup-arc.com/access/content/sedivy-2e-student-resources/sedivy2e-chapter-9-web-activity-4

Humans parse language despite rampant ambiguity

By now, we’ve compiled quite a list of variables that can either aggravate or relieve garden path effects. The overall experimental evidence suggests that the human parser is quite a bit more flexible and intelligent—and less error-prone—than was originally envisioned. In order to predict whether a specific sentence will seriously derail interpretation for readers or hearers, it’s not enough just to know that it contains a potential ambiguity. People are able to combine a large variety of informative cues to help resolve an ambiguity. If a number of different cues all conspire to point to the wrong interpretation, readers or listeners will run into a dead end, and have to backtrack in their processing attempts. But situations in which these cues overwhelmingly point to the wrong structure are extremely rare.

This turns out to be a good thing for language comprehension because, as it happens, language is even more inherently ambiguous than psycholinguists might have imagined. The sheer scope of linguistic ambiguity became apparent to computer scientists once they started building computer programs whose goal was to understand sentences produced by humans. They found that for many partial sentences of English, the programs would offer not one or two, but sometimes thousands of ways of parsing the sentence in conformity with the rules of English grammar. Most of these massively ambiguous sentence fragments are not at all problematic for humans—we filter out almost all of the grammatically legal structures as nonsensical or enormously unlikely to ever be uttered. But currently, computers are much less skilled at coping with ambiguity than we are. Attempts to build machines that can understand us have highlighted just how impressive we are at resolving ambiguities in language.

Is it possible to have a language without ambiguity? As far as I know, naturally occurring languages all involve ambiguity, but it is certainly possible to invent an ambiguity-free artificial language, as was done by James Cooke Brown (1960) when he created the language known as Loglan. In Loglan, words never have more than one meaning, and they’re combined in ways that resemble logical formulas—these formulas convey conceptual content in a transparent way so that you can never have one string of words that means more than one thing. Brown originally invented Loglan with the idea that it would be interesting to see whether a completely unambiguous language could help its speakers organize their thinking more clearly. But more recently, enthusiasts of Loglan (and a related language known as Lojban) have argued that unambiguous languages could prove to be very useful in the interaction between humans and computers, since ambiguity has such a catastrophic effect on computers’ ability to understand language.

WEB ACTIVITY 9.5

Create and recognize “easy” versus “hard” ambiguous sentences In this activity, you’ll be guided through a series of steps in which you’ll generate sentences that contain a potential temporary ambiguity. Some of these will be likely to cause very severe garden path effects, while others will be likely to cause only minimal disruption. You’ll hone your editorial instincts and learn to recognize the potentially troublesome sentences.

https://oup-arc.com/access/content/sedivy-2e-student-resources/sedivy2e-chapter-9-web-activity-5

Would humans benefit from using an unambiguous language among themselves? Despite the hype from some Loglan and Lojban enthusiasts, it’s not clear that they would. As you’ve seen, people are extremely good at cutting through much of the ambiguity inherent in language. And as discussed in Chapter 8, ambiguity might serve to lessen the demands on language production. In Box 8.3, I introduced the idea that ambiguous words allow languages to reuse easy-to-produce bits of language with minimal disruption to comprehension. In the case of syntactic ambiguities, it may be that a language that contains ambiguity allows for a simpler syntactic system, or allows speakers some important flexibility in how they order words and phrases. The notion that ambiguity might offer some important benefits is echoed by author Arika Okrent (2009), who describes her attempts at learning Lojban. She suggests that “composing a sentence in Lojban is like writing a line of computer code. Choose the wrong function, drop a variable, forget to close a parenthesis, and it doesn’t work.” As a result, she notes, watching live conversations in Lojban is much like watching people slowly doing long division in their heads.

9.3 Questions to Contemplate

1. Would the constraint-based model predict very similar garden path effects for a native speaker of a language and someone who is fairly new to it? What about the garden path model?

2. What kinds of cues might the parser use to anticipate the structure for the fragment The students studied in the library… ? What structure do you think would be predicted?

3. What does sentence processing in real-life contexts reveal about the nature of the parsing system?

9.4 Making Predictions

In the previous section, you learned that it’s rare for ambiguity to be truly disruptive for language processing—the human parser seems to be very adept at coping with uncertainty about structure. We regularly make use of our experience with language patterns, our knowledge of the world, or our awareness of the specific context to choose the most likely meaning when several are possible. This allows us to process language incrementally and on the fly, even when the fragments we’ve heard are highly ambiguous. However, researchers have shown that we may go even further than this. Not only do we choose among multiple possible structures for incoming fragments of speech, but we can sometimes anticipate the meanings of sentences, even before we’ve heard critical portions of these sentences.

Figure 9.6 A sample visual display for testing anticipatory eye movements. In one experiment of the 2003 study by Kamide et al., the subjects’ eye movement data revealed that upon hearing The girl will ride the … , participants were already predicting the appropriate direct object (i.e., the carousel). (Adapted from Kamide et al., 2003, J. Mem. Lang. 49, 133.)

Eyeing the content yet to come

Anticipatory, or predictive, language processing has been demonstrated in a number of eye-tracking experiments. In one of these studies, Yuki Kamide and her colleagues (2003) showed people visual displays illustrating several characters and various objects, and observed their eye movements as they listened to sentences involving these characters and objects. For example, while viewing a picture such as the one in Figure 9.6, subjects would hear one of the following sentences:

The man will ride the motorbike.

The man will taste the beer.

The girl will ride the carousel.

The girl will taste the sweets.

The researchers found evidence that people were anticipating the identity of the direct objects in the sentence. That is, upon hearing The girl will ride the … they were already more likely to be looking at the carousel than at any of the other objects in the display. (In contrast, hearing The man will taste the … prompted glances at the picture of the beer.) Listeners’ guesses about the upcoming direct object reflected the combination of the previously heard subjects and verbs—that is, people didn’t just look at the rideable objects (the carousel and the motorbike), or the objects that they think girls like (the carousel and the candy). They looked at whichever object was most likely to be involved in an event involving the already-mentioned subject and verb. This suggests that they had integrated linguistic information about the subject and verb, used this information to build a partial meaning of the unfolding sentence, and were generating predictions about the most likely way to fill in the sentence’s meaning.

In another experiment, Kamide and her colleagues leveraged the grammar of Japanese to show that predictions about meaning are based on a subtle interplay between linguistic and contextual information. Unlike in English, in which the verb usually appears before the direct object, in Japanese the verb typically appears at the end of the sentence after both the subject and its direct and indirect objects. For example, here’s the Japanese translation for the sentence The waitress will merrily bring the hamburger to the customer:

weitoresu-ga	kyaku-ni	tanosigeni	hanbaagaa-o	hakobu.
waitress	customer	merrily	hamburger	bring.

Japanese subjects heard sentences such as these while viewing visual displays (see Figure 9.7).

Figure 9.7 A sample of the type of display used in the Kamide et al. study to probe the integration of grammatical and contextual information in generating predictions about the upcoming content in the sentence (see text). (Adapted from Kamide et al., 2003, J. Mem. Lang. 49, 133.)

What kind of predictions were listeners making after they’d heard just the first two or three words of the sentence? Since waitresses usually bring people food, it would be sensible for the listeners to assume that this is the direction the sentence was taking, and to throw some anticipatory glances at the hamburger. But waitresses can also do many other things involving customers; they can greet them or tease them, take their orders, recite the menu, and so on. What information could the Japanese hearers use to constrain the set of possibilities?

One potentially useful source of information is the presence of case markers in Japanese—these are grammatical tags that appear on nouns to indicate whether the noun is a subject (in which case it appears with the -ga tag), an indirect object (-ni), or a direct object (-^o). In the above sentence, the noun meaning “customer” is tagged with an indirect object marker. This means that the customer could be involved in, say, an event of bringing, which does involve an indirect object; but he couldn’t be the target of a teasing event, which involves a direct object, but no indirect object. In fact, if the waitress were teasing the customer, we’d have to tag the noun with the direct object tag -^o rather than -ni:

weitoresu-ga	kyaku-o	tanosigeni	karakau.
waitress	customer	merrily	tease

Is predictive language processing sensitive to such linguistic subtleties? Kamide and her colleagues showed that it can be. Upon analyzing their subjects’ eye movements, they found that after hearing only the first three words (that is, the equivalent of waitress customer merrily), the subjects were more likely to look at the hamburger than at the distractor object (the garbage can)—but only if the word for “customer” was tagged with an indirect object marker, signaling that the word was grammatically consistent with an event in which the waitress brought the customer some food. When the case marker was inconsistent with such an event, the subjects were not lured by the picture of the hamburger.

Brain waves reveal predictive processing

One of the great advantages of the eye-tracking method, as compared with button-press response time measures, is that it allows researchers to get a moment-by-moment view of what’s happening in people’s minds as they process language. This is because the target behaviors (that is, eye movements) are tracked continuously as speech unfolds, rather than being probed at one particular point in time after the presentation of some critical stimulus, as is the case in response time studies. Naturally, this makes eye-tracking a powerful method for studying predictive processing, where we want to see what’s happening in the mind before people encounter some critical stimulus. But the technique comes with certain limitations as well:

1. You have to be able to visually depict the content of the sentence, which restricts the abstraction and complexity of the stimuli.

2. Eye movements can only reveal how people respond to those aspects of the sentence that are visually depicted, but they’re fairly uninformative about whether people are generating predictions about content that isn’t present in the display. In a sense, the visual displays may be acting as a sort of multiple choice test situation—we can see whether people are favoring certain hypotheses over others among the options that are provided, but we can’t easily tell whether they’re generating hypotheses that aren’t made visually available as options, or whether they would do so in a more open-ended language-processing situation. This is important because, presumably, much of daily language processing takes place in less visually constraining contexts.

One way around these limitations is to study brain wave activity while people hear or read language. This approach offers researchers many of the advantages of eye-tracking methods, without the constraints imposed by visual presentation. Like the eye-tracking studies, experiments that measure event-related potentials (ERPs) can tap into moment-by-moment responses to language as it unfolds.

A number of researchers have used ERP techniques to ask: Do people anticipate the actual words that will appear downstream, rather than just the general content that’s yet to come? In one such study, Katherine DeLong and her colleagues (2005) had subjects read sentences like these (with words presented as text one at a time):

The day was breezy so the boy went outside to fly a kite.

The day was breezy so the boy went outside to fly an airplane.

If people are generating predictions about upcoming words, they’re probably more likely to come up with kite than with airplane. If you remember the discussion of N400 brain waves back in Chapter 3, unexpected words result in heightened N400 activity compared with more predictable words. (This is the case regardless of whether “predictability” reflects the semantic plausibility/anomaly of a word in the context, or simply a contrast between a common versus rare word.) Hence, subjects should display more energetic N400 activity upon reading the word airplane at the end of the sentence.

This is exactly what DeLong and her colleagues found (Figure 9.8). But more importantly, they found heightened N400 activity for the “airplane” sentences at the previous word, the article an. Notice that in the sentences above, the more predictable word kite should be preceded by the article ^a, following the English rule of phonology that says the article an only appears before vowel-initial words. By cleverly setting up a contrast in the articles that could appear with predictable and unpredictable words, the researchers were able to probe for clear evidence of word anticipation. The earliness of the N400 effect suggests that even before getting to the word airplane, subjects must have generated some expectations about the specific shape of the upcoming word; otherwise they wouldn’t have been surprised by the occurrence of the article an.

Figure 9.8 ERP data from the study by DeLong et al. (2005). This graph plots the ERP waveforms at the midline central (vertex) recording site for words that were determined to be of high versus low predictability based on a separate sentence completion task. Note that negative values are plotted upward. Both articles and nouns that were low in predictability showed significantly greater negativity between 200 and 500 ms after the onset of the word (N400) compared with articles and nouns of high predictability. (Adapted from DeLong et al., 2005, Nat. Neurosci. 8, 1117.)

Predictions in language and hockey: A matter of expertise?

Together, these eye-tracking and ERP studies portray language processing as extremely active and adaptive: Instead of waiting for the speech signal to trigger the process of word recognition and structure-building, we quickly pull together all the information that’s available to us at any point in time, and generate expectations about the sentence’s upcoming content and form. Naturally, these preliminary expectations might turn out to be wrong, in which case we have to be able to quickly revise them.

Based on findings like those above, some researchers have argued that the notion of predictability in language can offer a unified approach to understanding the processing of both ambiguous and unambiguous sentences. They see language processing—whether or not it involves ambiguity—as the business of making predictions about a sentence’s future trajectory, much as you might make predictions about the trajectory of objects moving in physical space. In fact, some researchers (e.g., Clark, 2013) claim that a central goal of mental activity in general is to make predictions about future events.

In language, this might involve predictions about how an ambiguous structure will pan out. But it can also involve generating expectations about the type of event that is being described, about whether a just-uttered word like the will be followed by an adjective or directly by a noun, or even stylistic expectations about whether the speaker is likely to refer to a youngster by using the word child or the more informal kid in a particular social situation. A number of researchers have worked at defining predictability in a way that will extend to all of these cases. For example, John Hale (2001) has used the term surprisal to capture this notion, arguing that processing difficulty is a general reflection of whether a sentence’s continuation is highly predictable, and whether the unfolding sentence ends up conforming to the most likely outcome. Surprisal is lowest when the probability that a sentence will continue in a particular way is very high, and the sentence does in fact continue in the most likely manner. The two examples that follow (only one of which involves a syntactic ambiguity) illustrate a low level of surprisal upon encountering the words in bold; hence, the bolded words should be very easy to process:

surprisal A measure that’s inversely related to the statistical predictability of an event such as a particular continuation of a sentence. Processing difficulty is thought to reflect the degree of surprisal at specific points in the sentence, so that less predictable continuations result in greater processing difficulty.

The horse raced past the barn and fell.

Lin likes to eat bacon with her eggs.

Surprisal is highest (leading to greatest processing difficulty) when the sentence continues in a way that was extremely improbable, as shown by the words in bold in these two examples:

The horse raced past the barn fell.

Lin likes to eat bacon with her cheesecake.

Intermediate levels of surprisal occur when the unfolding sentence is reasonably consistent with a number of possible outcomes, one of which turns out to be realized, as illustrated by these words in bold:

The landmine buried in the sand exploded.

Lin bought some potatoes.

This broad notion of surprisal does a good job of mapping onto processing times across a variety of sentences (Hale, 2001; Levy, 2008), and has been shown to be correlated with the size of the N400 effect upon encountering words in a sentence (Frank et al., 2015). High surprisal in sentences has also been linked to higher degrees of brain activation in areas that are involved in processing syntactic structure, the left inferior frontal gyrus (IFG) and the left anterior temporal lobe (ATL), bolstering the claim that extra processing effort is needed to understand sentences that violate expectations (Henderson et al., 2016).

But linguistic expectations don’t come out of thin air; a great deal of knowledge and skill goes into them. In the physical world, your ability to make predictions about where a moving object will land depends on your experience with moving objects. In much the same way, making linguistic predictions is dependent on your experience with language—with its grammar, with its common words and structures, and with how people tend to wield it in a variety of situations. Now, predicting an object’s motion trajectory is usually far simpler than predicting the path of a sentence. But to push the analogy a bit further, think of language processing as similar to the skills that are required for playing hockey at a very high level, and predicting the movement of the puck. According to author Adam Gopnik (2011), hockey requires, above all, a heightened situational awareness—the ability to “see the ice,” to quickly size up all the relationships among the fast-moving players, grasp what is happening, and anticipate what will happen next. Some players develop this skill to an elite level:

Well, hockey, obviously, which is played at incredibly high speed, reveals and rewards situational and spatial intelligence at a degree of difficulty that no other sport possesses. So much so, that the greatest of all hockey players, Wayne Gretzky, had, besides his other significant skills as a fine-edge skater, almost nothing else that he’s specifically good at. That is his gift—the gift of spatial and situational intelligence: knowing what’s going to happen in three seconds, anticipating the pattern approaching by seeing the pattern now, sussing out the goalie’s next decision and Jari Kurri’s or Luc Robitaille’s eventual trajectory in what would be a single glance if a glance were even taken. Gretzky is the extreme expression of the common skill the game demands.

To many psycholinguists, this is as apt a metaphor as there is for the “common skill” that is demanded by everyday language processing. Perhaps if all of us spent as much time playing hockey as we do using language, we too would have the expertise of a Wayne Gretzky.

Like hockey, making language-based predictions does appear to be an acquired skill. Arielle Borovsky and her colleagues (2012) set out to learn whether children, like adults, make predictive eye movements, in a study very similar to the eye-tracking experiment by Kamide et al. (2003). They tested adults and children age 3 to 10 years, presenting their subjects with spoken sentences such as The pirate hides the treasure, along with visual displays of various objects related to the sentences. They measured the point in time at which a subject began to look more often at the picture of the treasure than at the other objects in the display, in anticipation of the direct object.

As in the earlier study by Kamide and colleagues, adults in the Borovsky et al. study regularly made predictive eye movements very soon after hearing the beginning of the verb. Children did too, although their eye movements were somewhat more sluggish than the adults’, suggesting that the predictions about the upcoming sentence content were not being generated as quickly. Of special interest, though, was the fact that children with relatively large vocabularies launched speedier anticipatory looks than their peers with smaller vocabularies. Knowledge of language, more than age itself, was correlated with the speed of their predictions. And it might surprise you to learn that, even among the adults, vocabulary size was correlated with the speed of predictive processing.

In the language arena, as in hockey, it appears that some people are more expert than others at anticipating what is going to happen next. In fact, as psycholinguists have taken a closer look, they’ve found quite a bit of variability in the extent to which language prediction effects are seen in eye movements and brain wave activity. Predictive skills are less robust among speakers of a second language that has been learned in adulthood. In an extension of the eye tracking study by Kamide and colleagues (described on p. 372–373), Sanako Mitsugi and Brian MacWhinney (2016) found that native speakers of Japanese, but not third- or fourth-year learners of Japanese, were able to use Japanese case markers to make predictions about upcoming referents. Likewise, Martin et al. (2013) ran a version of the ERP study led by DeLong et al. (described on p. 374) and found no evidence that second-language learners anticipated specific words, even when they were highly predictable in the sentence context. Moreover, there’s evidence that people are less active in making linguistic predictions if they have low literacy skills (Huettig et al., 2011) or are of advanced age (Federmeier, 2007).

Findings like these have caused some researchers (e.g., Huettig & Mani, 2016) to wonder just how common it is to strongly anticipate language structure and content. Do the results of the eye tracking and ERP studies we’ve seen (usually conducted with young, highly educated subjects) truly reflect language processing skills that are universal and essential? Or do they tap into fairly elite skills that are possessed only by those who are steeped in their language and are at the top of their language processing game? In other words: is highly predictive language processing more like what we all do when we predict the path of a falling rock, or what Wayne Gretzky does when he’s working his magic on the ice? And if it’s a skill possessed only by some, do they deploy it all the time? Or is it a heightened state of awareness that is brought into some situations—such as lab studies—but not into more routines ones, just as Gretzky might turn down the dial of his situational awareness while strolling through his neighborhood?

WEB ACTIVITY 9.6

Linguistic predictions In this activity, you’ll explore a range of language phenomena in which expectations about upcoming structure or content could potentially guide language processing.

https://oup-arc.com/access/content/sedivy-2e-student-resources/sedivy2e-chapter-9-web-activity-6

Much more work needs to be done to understand the source of the variability of this skill, the situations in which it is wielded, and the consequences of making linguistic predictions (or not making them). Do active predictors gain an important edge in language understanding? Does it lead to deeper comprehension or better retention? Is prediction mentally exhausting? These are all questions that await answers.

9.4 Questions to Contemplate

1. What is the most convincing evidence that people can actively predict yet-to-be-uttered linguistic material, and what makes this evidence so convincing?

2. How can you explain sentence processing in a way that captures how people understand both syntactically ambiguous and unambiguous strings of words?

3. What are some of the reasons that the strong predictiveness seen in some eye tracking and ERP studies many not generalize to everyday language use?

9.5 When Memory Fails

Let’s take another look at the difficult sentences introduced at the beginning of this chapter. As you’ve seen, ambiguity is a common culprit in creating havoc for language comprehension. But sentences can be fiendishly hard to process for other reasons. For example:

The mouse the cat chased keeled over.

The administrator who the intern who the nurse supervised had accused fudged the medical reports.

It’s easy to see that these sentences involve more than one clause or sentence unit, and that they encode multiple events. (For instance, in the first sentence, there’s an event of the cat chasing the mouse, and another event in which the mouse keels over.) But encoding multiple events alone can’t explain the difficulty, because it’s certainly possible to express these same events in a sentence without unduly straining the processing system:

When the cat chased the mouse, the mouse keeled over.

The mouse keeled over when the cat chased it.

And oddly, if we want to say that it was the cat that keeled over rather than the mouse, the next sentence seems to be considerably easier to understand:

The cat that chased the mouse keeled over.

So, it’s not just that the sentence is generally complex or the ideas inherently difficult to understand, it’s that there’s something taxing about the particular structure with which those ideas are expressed. What is that something?

Memory failure or faulty predictions?

Here’s an example of two sentences that cause unequal brain strain, with the first typically read faster than the second, even though they’re made up of exactly the same words:

The senator who spotted the reporter shouted.

The senator who the reporter spotted shouted.

What’s the key difference? Both contain a relative clause attached to the subject of the main clause, the senator. But in the first sentence of the pair, the senator is also linked to the subject role of the embedded verb spotted. (Notice what happens in the question/answer pair Who spotted the reporter?/The senator spotted the reporter.) This type of relative clause is called a subject relative clause. In contrast, in the second sentence, the senator is linked to the direct object role of the embedded verb spotted. (Notice: Who did the reporter spot? The reporter spotted the senator.) This yields an object relative clause. As I did in Chapter 6, I’ll notate the relationships by using the same color for the phrase the senator and the underscore marking the position in the sentence that it links to:

The senator who ____ spotted the reporter shouted.

The senator who the reporter spotted ___ shouted.

So, why are object relative clauses harder to process than subject relative clauses? There are two competing classes of explanations for the difficulty: one that highlights the inherent limitations of the parser, and another that emphasizes its tendency to predict structure.

According to the first account, a version of which is articulated in detail by Ted Gibson (1998), the trouble crops up because complex sentences involve holding some parts of a sentence in memory until they can be related to other parts of the sentence—as the distance between these related parts gets bigger, so do the demands on memory and the prospects of processing failure. And, in some cases, multiple dependencies have to be held in memory at the same time, adding even more strain on memory.

In a subject relative clause, an NP (the senator) has to be integrated with the embedded verb (spotted), as well as the main clause verb (shouted), leading to two dependencies:

The senator who ___ spotted the reporter shouted.

This turns out to be fairly easy: at spotted, there’s no intervening NP to get in the way of the integration of the two components, and by the time the reporter is reached, the first dependency has already been resolved. Things are a bit different with the object relative version:

The senator who the reporter spotted ___ shouted.

Readers have to hold the senator in memory while encountering the reporter, and it’s the intrusion of this second NP that makes the integration with spotted that much harder. The sentence gets harder still if you add yet another intervening NP:

The senator who the reporter next to the president spotted ___ shouted.

In addition, while the intervening phrase the reporter is being encountered and stored, neither of the two dependencies involving the senator has yet been resolved, adding to the memory burden. You can see the effect of piling up the number of unresolved dependencies by taking the original object relative sentence and adding yet another object relative clause:

The senator who the reporter who the president detested ___ spotted ___ shouted.

For most people, this sentence disintegrates into utterly unrecognizable word mush. But again, there’s nothing difficult about the ideas expressed in the sentence, as we can see if we unravel the sentence and present the same content in a way that avoids the memory crash:

The president detested the reporter who ___ spotted the senator who ___ shouted.

There’s another memory-related issue that may play a role in these horrendous sentences: the semantic similarity of the nouns reporter and senator. Some researchers (e.g. Lewis, 1996) have pointed out that it’s a general property of memory that similar items tend to blur together and become harder to retrieve separately. It’s possible that when complex sentences involve very similar nouns that have to be held in memory at the same time, this might make it hard to retrieve the right noun in the right place in the sentence. The experience might be analogous to reading a long Russian novel populated by a large cast of characters, all of whose names sound alike—it becomes onerous to keep track of who did what to whom. Indeed, readers often seem to fare better with object relative sentences that contain very dissimilar nouns. For example, compare:

The senator who the president summoned arrived.

The helicopter that the president summoned arrived.

These explanations all derive from the very reasonable assumption that the process of building syntactic structure taps working memory resources, which, sadly, are known to be finite. But not all researchers focus on memory limitations in explaining the difficulty of object relative clauses. Some have suggested that the difficulty is a by-product of the parser’s tendency to predict upcoming structure based on previous experiences with language; in this case, experience with the statistical likelihood of different structures leads the parser astray. In a sense, you might think of the subject relative/object relative asymmetry as simply another kind of very transient garden path effect. To see this, have several people you know (but not from this class) complete the following fragment:

The beauty queen who …

How many of them produced an object relative clause as opposed to a subject relative clause? Based on the statistical patterns of English, chances are that the subject relatives came up more frequently, with most people providing a verb rather than a second NP as the next element to continue the fragment. In that case, according to an experience-based account, they’d have trouble reading sentences with object relative clauses (for example, The beauty queen who the judge disqualified broke into tears) not because they’d drained some finite processing resources, but because their perfectly intelligent and reality-based expectations about upcoming structure were not met once they ran into the more unusual sentence structure.

Both of these accounts are well grounded in reasonable assumptions about language processing, assumptions for which there’s a fair bit of independent evidence. Luckily, there are ways to pull apart the two explanations. For instance, they make somewhat different predictions about where in the object relative clause readers should experience the greatest difficulty. For experience-based accounts, this should happen right upon encountering the second NP, where expectations of a subject relative structure are confounded (as in The beauty queen who the judge disqualified broke into tears.) For memory-based accounts, the trouble spot should be at the embedded verb, where the noun has to be integrated with the verb, over an intervening NP (The beauty queen who the judge disqualified broke into tears.) These are subtle points, but the evidence so far suggests that both explanations contribute to the patterns of difficulty seen with such sentences.

Case markers as disentanglers: A Russian lesson

Other languages sometimes offer opportunities to disentangle competing explanations in ways that English can’t. For example, Russian, like Japanese, allows a fair bit of freedom in its word order, made possible by the fact that case markers indicate the role that each participant plays in the event. So, a sentence with a subject relative clause can be built using different word orders (see Levy et al., 2013). In the following example, NOM refers to nominative case, indicating subject status, and ACC refers to accusative case, indicating object status:

(a)	Slesar’,	kotoryj	udaril	elektrika	so	vsego	rasmaxa,	ushel	domoj.
	Repairman	who-NOM	hit	electrician-ACC	with	all	strength	went	home

The Russian sentence marked ^a uses a word order that is similar to its English translation (“The repairman who hit the electrician with all his strength went home”), although it adds case markers and has no articles corresponding to English “the.” But sentence b below has quite a different word order, even though it means exactly the same thing: In this version, the word for “electrician” appears before the verb (marked in bold), much like the English word order for object relative clauses. Hence, it should lead to a greater memory load.

(b)	Slesar’,	kotoryj	elektrika	udaril	so	vsego	rasmaxa,	ushel	domoj.
	Repairman	who-NOM	electrician-ACC	hit	with	all	strength	went	home

Likewise, a sentence with an object relative clause—translated here as “The repairman who the electrician hit with all his strength went home”—in Russian has the following two options:

(c)	Slesar’,	kotorogo	elektrik	udaril	so	vsego	rasmaxa,	ushel	domoj.
	Repairman	whom-ACC	electrician-NOM	hit	with	all	strength	went	home

(d)	Slesar’,	kotorogo	udaril	elektrika	so	vsego	rasmaxa,	ushel	domoj.
	Repairman	whom-ACC	hit	electrician-ACC	with	all	strength	went	home

Both object and subject relative clauses use word orders that should be fairly easy from a memory standpoint (sentences a and d) as well as word orders that should be more difficult (sentences b and c). However, not all word orders are equally common; within subject relative clauses, it’s more common for the verb to precede the noun (a is more common than b), whereas the opposite is true for object relative clauses (c is more common than d). The experience-based account predicts that the most common sentences (a and c) should be the easiest to process, diverging somewhat from the predictions of the memory-based account.

A study led by Roger Levy (2013) exploited this property of the Russian language to pit the two theories against each other. Both accounts partially accounted for reading times at the critical verb (shown in bold in the preceding examples). In line with the memory-based account, sentences b and c showed longer reading times than the other two; sentence b was more difficult than c, however—a fact that was predicted by the experience-based account but not by the memory-based story.

9.5 Questions to Contemplate

1. What are the crucial properties of sentences that are most likely to tax limited memory resources during sentence understanding?

2. Why is there so much overlap between the predictions of memory-based and experience-based explanations for difficulties in processing complex sentences?

3. If you were looking to disentangle the predictions of the two accounts using data from other languages, what general patterns would you look for in the language, and what kinds of evidence would you need to gather?

9.6 Variable Minds

So far, I’ve been using expressions like “the parser” and “the language processing system” as if sentence comprehension involved something like a software program, and each one of us had exactly the same copy of it installed in our heads. Many language researchers will admit that when you look closely at individuals’ data, you might see quite a bit of variation—for example, some people might display enormous garden path effects for certain sentences, while others might show subtle or nonexistent slowdowns. But until fairly recently, this kind of variability was treated as unsystematic experimental “noise.” Building and testing theories has typically focused on the commonalities that can be found across large numbers of language users, not their differences. But differences between individuals are becoming a very interesting part of the language-processing story. A recent fMRI study (Mahowald & Fedorenko, 2016) has revealed that people have unique patterns of brain activation while processing language, and that these activation patterns seem to be stable within each individual across experimental sessions (see Figure 9.9). And we’ve already seen that some groups of people (such as second-language learners or older people) show weaker evidence of predictiveness in language processing,

Figure 9.9 Sample activation for five participants across two sessions contrasting brain activity in a sentence reading task and a non-word reading task. In the reading task, participants read meaningful sentences such as A RUSTY LOCK WAS FOUND IN THE DRAWER, whereas in the non-word task participants read lists of non-words such as DAP DRELLO SMOP UL PLID KAV CRE REPLODE. These maps indicate areas of the brain that showed greater activity during the sentence reading task relative to the non-word task. Note how activation patterns differ across individuals but are stable with individuals between sessions. (From Mahowald & Fedorenko, 2016, NeuroImage 139, 74.)

Such differences hint at distinct cognitive profiles across individuals or groups. Far from being “noise,” they could play an important part in how people use and understand language. There are obvious practical implications to understanding how individuals might vary in their language processing abilities. Such knowledge could lead to targeted treatments or training programs. And knowing the limits of what a specific audience might find easy to process could allow for better design of text or spoken language materials.

Studying individual differences could also dramatically expand our understanding of the cognitive work that takes place during language processing. Conceptually, research on individual differences is similar to research in neuropsychological disorders: By studying disorders that are linked to specific abnormalities of the brain, researchers can make inferences about the mental operations that are involved in various tasks. A natural extension of this logic leads to looking at variation that occurs within the general population. By seeing which abilities cluster together in people who are especially good or especially bad at specific tasks, we may get some deep insights about the mental powers that go into language processing.

Individual differences in memory span

The earliest research to focus on a connection between individual cognitive profiles and language processing was in the domain of memory. As you saw in the previous section, a number of researchers have pointed the finger at working memory limitations to explain why some sentences are hard to understand. And that’s interesting, because it seems that some people are blessed with longer memory spans than others, and that these are stable individual traits that can be measured by standard tests. So, if some sentences are difficult because they bump up against limits on memory capacity, then we should be able to predict, based on someone’s working memory span, whether their interpretation of such sentences is especially prone to crashing—and maybe even whether they experience the prose of Henry James as pleasantly stimulating or as a depressing slog.

In an important paper, Marcel Just and Patricia Carpenter (1992) looked at several aspects of language processing in subjects with high memory spans versus subjects with lower spans. They measured memory span using a particular test known as the reading span test (developed by Meredyth Daneman and Patricia Carpenter, 1980). This task, which is intended to mimic the memory pressures on the parsing system, requires subjects to keep certain words active in memory while reading and understanding text. You try it—read the following sentences, and as you do, be sure to remember the last word in each sentence:

reading span test A behavioral test intended to measure an individual’s verbal working memory. The test involves having the individual read a sequence of sentences while holding the last word of each sentence in memory. The number of words successfully remembered corresponds to that individual’s memory span.

Despite their persistent disagreements, they managed to agree how to best educate their child.

At last, she spotted the rowboat coming across the bay, tossed about on the tall waves.

Long periods of unstructured reading and thinking sometimes lead to the most fertile ideas.

Quiz time! Without looking back, can you remember the last word of every sentence? If you could, the reading span test would continue, inflicting upon you one sentence after another until you failed to recall all of their last words. This breaking point would reflect the limits of your memory span on this particular test.

Just and Carpenter argued that there are several consequences to having a lower memory span. The most obvious is that it becomes very hard to understand sentences that make heavy demands on verbal working memory. They reported evidence that low-span subjects were especially slow at reading object relative clauses (which have been argued by some researchers to be memory-hogging sentences) compared to the less taxing subject relative clauses. In contrast, the higher-span subjects experienced a more subtle slowdown for sentences with object relative clauses, presumably because their extra processing capacity could accommodate these structures fairly easily.

You’ll remember, though, that not all researchers agree that object relative clauses are difficult because they stretch memory resources too thin. The experience-based argument was that these sentences are hard simply because they don’t occur very frequently, hence they defy the word-by-word predictions that parsers make over the course of a sentence as it unspools. Notice that in this debate, individual differences become an important source of evidence that can shed some light on very general questions about the nature of parsing. Namely, if it turns out that we can systematically predict how hard it is for individuals to parse object relative clauses by looking at their memory spans, then this looks like pretty strong support for the view that such sentences do in fact draw heavily on memory resources, and that this is what makes them potentially difficult. In other words, the different parsing outcomes of different individuals can tell us something about the language processing system.

But all this depends on the assumption that the reading span test truly is a valid measure of pure memory abilities—an assumption that some researchers have questioned. Maryellen MacDonald and Morten Christiansen (2002) have argued that some people do better on the reading span test not because they have a more spacious memory, but because they are better readers. Better readers, they claim, are better able to read and understand sentences efficiently, consuming less in the way of processing resources, which simply makes them appear to have more processing capacity. And how do people get to be better readers? Generally by spending more time immersed in written language than their peers. The memory test, then, is just a stand-in for the amount of reading “training” a person has had. And why would those with more experience at reading have an easier time with unusual structures like object relative clauses? Possibly because they’ve encountered these unusual structures many more times than people who read less (especially given that written language is far more likely to use uncommon structures than spoken language; see Box 9.5). Under this view, the way to get better at coping with the intricate syntax of Henry James is not to do mental calisthenics that extend memory capacity—rather, it’s simply to read more complex prose.

While MacDonald and Christiansen argued that the reading span test captures linguistic experience rather working memory, other researchers claim that the test is based on the wrong notion of memory in the first place. As you saw in Section 9.5, one way to think about the memory strain that’s imposed by hard sentences such as object relatives is in terms of the amount of information that has to be held in memory at a time. But another view is that these sentences are hard because items in memory can blur together, putting more emphasis on the kind of information in memory. (This latter account explained why The helicopter that the president summoned is so much easier than The senator that the president summoned.) As argued by Julie Van Dyke and her colleagues (2014), the reading span test is irrelevant to this problem. Instead, the sharpness and detail of word representations is more likely to matter—someone who has very rich representations of the words senator and president is less likely to confuse them in memory. Hence, they argue, an individual’s vocabulary knowledge is a more relevant predictor of difficulty with such sentences.

WEB ACTIVITY 9.7

Test yourself on various memory tasks To what extent do you think these tests do or don’t tap into the same kinds of memory resources that are needed to process complicated sentence structures?

https://oup-arc.com/access/content/sedivy-2e-student-resources/sedivy2e-chapter-9-web-activity-7

This ongoing debate over the reading span test shows how important it’s going to be for researchers to be really precise if individual differences are to shed light on the mental operations that lie beneath the surface of language. They will need to figure out if memory tests really do measure what they’re intended to measure. And, just as important, they will need to develop and test very specific hypotheses about the relevant mental operations.

Differences in cognitive control

More recently, researchers have become interested in another source of individual variation and its relation to language processing, namely, the ability to resolve conflicting information. The language-processing landscape is one of rampant, cutthroat competition among linguistic representations, all clamoring for attention. Often the difference between successful understanding and utter cognitive mush comes down to being able to focus on the right representations while ignoring the others. This seems to be part of a general skill called cognitive control, also referred to as executive function. You might think of cognitive control as the managerial aspect of cognition, directing attention depending on specific mental goals and supervising behavioral responses to stimuli. You need it for much more than just language, and without it, it would be impossible to function competently in a complex environment; you wouldn’t be able to do things such as drive, follow a conversation in a crowded room, avoid swearing at your boss, or do any one of a thousand daily things that require you to ignore information or impulses that get in the way of your specific goals. As you might have noticed, the capacity for cognitive control seems to vary quite a bit among individuals. Some are able to push away intruding information or impulses very effectively, while others are more likely to be lured by them.

cognitive control (executive function) The goal-directed cognitive processes responsible for directing attention and supervising behavioral responses to stimuli.

BOX 9.5

The language experience of bookworms versus socialites

If you’re the kind of person who would rather stay home and read a good book than go out to a party, take comfort: your social life may not be that exciting, but there’s a good chance that you can outperform many of your peers in understanding sentences like the following:

Sam noticed that the fact that Andy had a new girlfriend bothered Freda.

That candidate will be hard to get many people to vote for.

The student who was scolded by the teacher finally finished the assignment.

That’s because these are structures that hardly anyone ever says, but that appear in text much more frequently. For instance, the last sentence in the example above is a fine specimen of a passive relative (the embedded clause is in the passive voice). Figure 9.10 shows the distribution of such examples across large collections of spoken and written language. As you can see, heavy readers will meet many more examples of this structure than even the most gregarious of people who rarely read.

Differences between spoken and written language can be found even in language directed at children. Jessica Montag and Maryellen MacDonald (2015) analyzed spoken and written language targeting children and found a much higher occurrence of both passive relatives and object relatives in children’s books than in spoken language—in fact, passive relatives were more frequent in their sample of children’s books than in the samples from Figure 9.10 of spoken language directed at adults.

There’s plenty of evidence that uncommon structures become easier to process with greater exposure (for a demonstration with passives, see Street Dąbrowska, 2010). Given the stark differences in sentence structures that appear in written versus spoken language, and the fact that people vary greatly in terms of how much text they consume, this is a factor that needs to be taken into account in any research on individual differences.

Figure 9.10 Frequency of occurrence of passive relatives across various written and spoken corpora (i.e., large samples of language). (Adapted from Roland et al., 2007, J. Mem. Lang. 57, 348.)

At one extreme of the spectrum lie those individuals who, because of a stroke or brain injury, have clearly identifiable brain damage to the prefrontal cortex area of the brain, an area that seems to be responsible for exercising cognitive control. Patients with damage to one specific area of the brain’s frontal lobes—the left inferior frontal gyrus (LIFG)—have a hard time resolving conflict between competing representations. One way this shows up is in their performance on a Stroop test, where they are told to name the color of the font that a word is printed in while ignoring the meaning of the word. For example, subjects would have to identify the font color for words that appear like this: BLUE, CHAIR, YELLOW. Needless to say, it’s not hard to name the font color for BLUE or even CHAIR. But to give the correct answer “red” when seeing YELLOW means having to suppress the interfering meaning of the word. People are slower and make more mistakes when the font color and meaning mismatch in this way, and patients with damage to the LIFG are especially slow and error-prone.

Stroop test Behavioral test in which subjects are required to name the color of the font that a word appears in while ignoring the (possibly conflicting) meaning of the word.

Language processing is often a lot like a Stroop test. Think about what you have to do to understand a garden path sentence, for example, or the meaning of the word ball in the sentence She decided not to attend the ball. In both cases, irrelevant meanings are activated and ultimately have to be squashed. A growing body of research suggests that all of these examples are related to cognitive control abilities. For instance, patients with LIFG damage have a hard time settling on the less-frequent meaning of an ambiguous word, or recovering from garden path sentences. Among people in the normal range, brain imaging studies have shown that the LIFG is especially active in just such situations, when people have to override a strong bias toward an interpretation that turns out to be wrong. And behavioral tests have shown that difficulty in retrieving the less common meaning of a word or recovering from a garden path sentence seem to be related to some measures of cognitive control (for a summary of these findings, see Novick et al., 2005). All of these findings point to there being some consistent aspect of a person’s cognitive profile that affects the ways in which that person experiences language and its tremendous potential for ambiguity.

WEB ACTIVITY 9.8

Measuring executive function A number of different tests can be used to measure executive function. This activity provides a demonstration of several of them.

https://oup-arc.com/access/content/sedivy-2e-student-resources/sedivy2e-chapter-9-web-activity-8

As luck would have it, human brains go through a good deal of their lives being somewhat short of managerial personnel. The prefrontal cortex is one of the last of the brain regions to fully mature, which explains why teenagers who are perfectly capable of mastering calculus or higher-order logic can torment their parents with their inability to suppress certain impulses or inhibitions. These facts of neuroanatomy have been used by experts to argue that criminal courts should not apply the same sentencing criteria to adolescents as they do to adults. And, even though it won’t land them in jail, you might expect that youngsters would also experience some trouble with linguistic ambiguity, and they do. John Trueswell and his colleagues (1999) tracked the eye movements of 5-year-olds listening to garden path sentences such as, “Put the frog on the napkin into the box.” There’s a temptation to treat on the napkin as attaching to the verb, and hence understand it as the intended destination for the frog, rather than as modifying the noun (as in the frog that’s on the napkin; see Figure 9.11).

Figure 9.11 An example of a visual display from a 1999 eye-tracking study conducted by Trueswell et al. The subjects were children, who viewed the display while being given ambiguous and unambiguous versions of the instruction “Put the frog (that’s) on the napkin in the box.” With the ambiguous instruction, in addition to showing evidence of confusion in their eye movements, children often performed incorrect actions, such as hopping the frog onto the napkin and then into the box. This suggests that they were unable to completely suppress the incorrect interpretation of the ambiguous sentence. (Adapted from Trueswell et al., 1999, Cognition 73, 89.)

When confronted with such sentences, adults usually experience a brief garden path effect, as evident from their eye movements, but then quickly recover and perform the correct action. But in what the researchers cutely dubbed the “kindergarten-path effect,” 5-year-olds experienced more devastating consequences of the ambiguity, performing the right action less than half of the time—for example, they might hop the frog onto the napkin, and then into the box, as if they were blending both interpretations rather than suppressing the incorrect one in favor of the right one. Not surprisingly, children at any given age also vary in their cognitive control skills; a more recent study led by Kristina Woodard (2016) found that performance on two tests of cognitive control was correlated with children’s ability to recover from such garden path sentences.

For kids and their immature prefrontal cortices, processing traps laid by potential ambiguity may not be limited to extreme garden path sentences; they can also lurk in mundane word recognition. Remember that all words—even those that don’t have a homophonous match—are temporarily ambiguous, so in the first few hundred milliseconds of the word log, both log and lock become activated (among others). In Chapter 8, we saw evidence of such eager activation in subjects’ eye movements to a visual scene—for example, when hearing log, people often look at a key because of its relationship to the cohort competitor lock. For normal adults, such eye movements are usually extremely fleeting, quickly dropping down to baseline right after enough phonetic information has rolled in to identify the target word. And subjects are almost never aware of these spurious eye movements, or that they briefly flirted with interpreting log as lock. But the competition seems to linger on a bit more for children: Yi Ting Huang and Jesse Snedeker (2010) found that in the same situation, 5-year-olds tended to persist in looking at the key for some time after hearing the disambiguating final consonant of the word log, and that they sometimes chose the key after being asked to “pick up the log.” Perhaps once lock has been activated, and hence its related word key, the irrelevant representations continued to reverberate.

WEB ACTIVITY 9.9

The kindergarten-path effect In this activity, you’ll see some video recordings of subjects’ eye movements and actions in response to ambiguity. You’ll be able to see how adults and children differ in how they resolve confusing garden path sentences.

https://oup-arc.com/access/content/sedivy-2e-student-resources/sedivy2e-chapter-9-web-activity-9

More relaxed cognitive control, though, may not be all bad. Some researchers (e.g., Sharon Thompson-Schill and colleagues, 2009) have actually suggested that a slow-maturing prefrontal cortex may have evolved because of its overall cognitive advantages. Perhaps less cognitive control helps children learn language by allowing them to easily generalize rules to new situations (and see Box 9.6 for a discussion of relaxed cognitive control in aging adults). Less top-down cognitive management may well help anyone think more creatively. When it comes to language, a bit less cognitive discipline could even lead to a deeper enjoyment of the aesthetic potential of language—I’m thinking of Virginia Woolf and her inability to hear the words “Passing Russell Square” without hearing the rustling sound of long skirts on floors (see p. 312). Who knows? Perhaps her keen awareness of the rich ambiguities inherent in all words may have had its roots in an artistic temperament and its failure to squash down all sorts of irrelevant but wonderfully interesting information.

BOX 9.6

How does aging affect sentence comprehension?

Can you expect to maintain your language processing powers as you get older? You might be pessimistic in light of research hinting at the importance of working memory and cognitive control—these abilities are known to decline dramatically with age. Even so, healthy older people rarely report having trouble with understanding language under normal conditions. And even though you’ve seen that older people may be less predictive in their processing, their ability to extract meaning even from difficult sentences remains fairly stable (Shafto & Tyler, 2014). How does such a complex mental activity stay robust in the face of declining abilities that contribute to it?

The answer may lie in the one great advantage that age has to offer: experience. There is no upper limit (that we know of) to the amount of knowledge you can accumulate over a lifetime, so people continue to build their knowledge of new words, develop more precise and detailed representations of words they’ve seen many times, hone their statistical knowledge of the syntactic patterns of their language, and generally acquire more knowledge about a variety of different topics. These gains may help offset the depressing list of disadvantages due to aging, including slowed processing, reduced working memory spans, and a weakened ability to resolve conflicting information. There are at least three ways in which experience might offset these effects. First, sharpened lexical representations may allow words to be recognized or read more efficiently, leaving more resources available for computation. Second, more accurate expectations about the structures of complex sentences may reduce the computational demands of processing such sentences. And third, a rich store of knowledge about the world may enhance older adults’ use of contextual and real-world knowledge to bolster their syntactic processing.

All this would suggest that the amount and quality of experience with language over your life could have a big impact—and over decades’ worth of language, this experience can be highly variable across individuals. In particular, regular exposure to written language—with its more varied and complex vocabulary and syntax—may be an important factor in preserving language comprehension abilities.

There is some support for this idea. Brennan Payne and his colleagues (2014) found that older adults with avid reading habits were more attuned to the statistical patterns of English in resolving syntactic ambiguities. A separate study (Payne et al., 2012) also found that older adults with more exposure to text were faster at reading complex words and were better able to recall sentences. More importantly, reading experience may have softened the effects of declining working memory: while participants with less reading experience showed a substantial difference in performance depending on their scores on a working memory task, differences in working memory were less relevant for the more avid readers.

The connection between literacy and language performance shows how enhanced knowledge of language might guard against the ravages of age. But there’s also evidence that knowledge of a specific topic can compensate for deteriorating working memory. Lisa Soederberg Miller (2009) had people between the ages of 18 and 85 read passages about cooking. She also assessed their knowledge of cooking and administered working memory tests. Overall, subjects with higher working memory scores read the passages more quickly and were better able to recall the information than those with lower working memory scores. But among the older adults, performance was enhanced by cooking expertise. Interestingly, this wasn’t the case among younger participants, suggesting that knowledge plays an especially important role in language processing as people get older.

Figure 9.12 Example sequences illustrating the n-back training task used by Hussey et al. (2017). (A) The low-conflict 4-back version of the task requires participants to identify targets that appeared 4 letters prior (marked here as YES). (B) The high-conflict 4-back version requires participants to identify targets that appeared 4 letters prior (marked here in red as YES) while disregarding distractor lures that appeared 2, 3, 5, or 6 letters prior (identified here in orange as NO). (Adapted from Hussey et al., 2017, J. Exp. Psychol. Learn. Mem. Cogn. 43, 23.)

Effects of cognitive training

Correlations between cognitive abilities and language behaviors provide clues that such abilities may be important for understanding language. But to get more direct evidence of their role, we need to look further. One approach is to see what happens if people improve abilities like cognitive control or working memory. Does this change the way they process sentences? The first step is to ask whether working memory or cognitive control are malleable—that is, can people improve with practice on tests that tap into these skills? Many studies show that indeed they can, opening the door to the question of whether such improvement translates into effects on language processing.

A study led by Erika Hussey (2017) posed just this question. These researchers used the so-called n-back task, which requires people to monitor the contents of memory in a very precise way. Subjects saw sequences of letters presented for half a second, 2 seconds apart, and were instructed to press one of two buttons to indicate whether the same letter (regardless of case) had appeared ⁿ letters back. Figure 9.12A shows what this might look like in a 4-back version; a 2-back version would be easier, and a 6-back version would be more difficult. Hussey and her colleagues were mainly interested in contrasting this version of the task with an even more difficult version that also requires the ability to suppress potentially interfering memory. In this high-conflict version, “lures” were inserted into the letter sequences; these were letters that had appeared previously, but not precisely the required number of letters back—for example, in a 4-back version, the lure might appear 2, 3, 5, or 6 letters back (see Figure 9.12B). In the high-conflict version, you not only have to correctly remember which letter appeared 4 letters back, but you also have to ignore the distracting influence of a letter that appeared close to, but not exactly, 4 letters back. In other words, this task recruits cognitive control in addition to memory span.

n-back task A test of working memory and cognitive control in which participants view sequences of symbols and have to maintain a certain number of previous letters symbols in memory and respond affirmatively when a symbol is repeated at a specified interval. The test can occur with lures (distractor symbols that repeat at intervals close to the specified one) or without lures. The version with lures increases the demands on cognitive control.

Hussey and her colleagues were interested in testing a very specific hypothesis: Given that garden path sentences appear to rely on cognitive control, they predicted that practice with the high-conflict ⁿ-back task would help subjects read garden path sentences more smoothly, but that the low-conflict version of the task, which doesn’t require people to resist the distraction of lures, would not have any effect. That is, they expected a reduced garden path effect (less difference in reading times between ambiguous and unambiguous sentences below), but only for subjects who had practice with the high-conflict task:

While the thief hid the jewelry that was elegant and expensive sparkled brightly. (difficult; temporarily ambiguous garden path sentence)

The jewelry that was elegant and expensive sparkled brightly while the thief hid. (easier; unambiguous sentence)

Moreover, they argued, not all complex sentences lean heavily on cognitive control, so these shouldn’t be affected by practice on the high-conflict task. Hence, they predicted that the training shouldn’t do anything to reduce the processing difficulty of sentences with object relative clauses as compared with subject relative clauses:

The farmer who the expert questioned promoted the product at the fair. (difficult; object relative clause)

The farmer who questioned the expert promoted the product at the fair. (easier; subject relative clause)

This was indeed the pattern they found after subjects spent a total of 8 hours practicing their designated training task, split into 30-minute sessions over 3 to 6 weeks.

This study supports the idea that resolving ambiguity draws on a more general ability to resolve conflicting information, and that strengthening this ability has measurable effects on reading garden path sentences. And, on the face of it, it seems to provide fodder for the claims of companies that sell computerized “brain-training” programs like Lumosity, Jungle Memory, or CogniFit—these companies suggest that practice with tasks similar to the one used by Hussey and her colleagues can produce real-world returns at work or school. The underlying logic is that abilities such as working memory or cognitive control are needed for a whole range of mental activities, so enhancing them will yield broad benefits. An easy analogy is that of doing weight training at the gym—doing exercises that target specific muscles will improve your performance in any sport or activity that uses those muscles.

But in reality, the results of these brain-training programs are fairly underwhelming. In a comprehensive review, Daniel Simons and his colleagues (2016) found virtually no evidence that this kind of training offers any benefits for everyday functioning. Why is this, and how does it square with the results of Hussey and her colleagues?

The gym analogy seems reasonable enough—brain-training companies are partly right when they compare the brain to “a muscle” that gets stronger through strenuous activity. Our brains do have a remarkable ability to adapt to experience. But the brain is hardly like a single muscle—in fact, the problem is that we just don’t know how many metaphorical “muscles” there are in the brain, or which activities use them. Lifting weights at the gym is worth your while because there’s a fairly limited number of skeletal muscles, and some of them are heavily used in a huge range of activities. But what if there are thousands of mental “muscles” each with fairly specialized functioning? Training a few of them wouldn’t really be worth your time—even if the effects of training did transfer to other activities, their overall contribution to your daily activities would be pretty minimal. And who has the time (or stamina) to train hundreds or thousands of skills on a computer?

Some cognitive skills seem to be highly specialized, with limited transfer across activities. For example, chess grandmasters are truly impressive at remembering the position of chess pieces on a board—but this doesn’t translate that well into being able to remember arrangements of other objects or even randomly placed chess pieces (Gobet & Simon, 1996). And a closer look at the results found by Hussey and her colleagues underscores just how specific their effects were: the low-conflict task, which exercises simple memory span, had no effect on reading any of the sentence types; and the effects of training on the high-conflict task were limited to resolving very strong garden path sentences (though it’s possible that training would affect other complex sentences that weren’t tested). As you saw in Section 9.3, really hard garden path sentences probably represent a very small proportion of the sentences we read. These results tell us something very interesting about the nature of ambiguity resolution—but how willing would you be to spend hours doing repetitive n-back tasks in order to smooth out your reading of garden path sentences?

In its current state, the brain-training industry is a cautionary tale about an intriguing, plausible hypothesis that got ahead of the evidence. It may ultimately turn out that some tasks do end up training highly general abilities that transfer to a number of important activities. But our knowledge of the brain and its functions is not yet at the point where we can say that with any confidence, and designing brain-training programs is a bit like trying to design a weight-training program without a good model of muscular anatomy. To add to the complexity, the analogous brain model could look very different at different points in the life span—for example, it could well turn out that training leads to better transfer for children because their brains haven’t yet specialized to the same degree as adults. In the meantime, it will take many painstaking studies like those of Hussey and her colleagues to map out which cognitive processes are shared among which mental tasks, and to locate the points of transfer and their limits.

LANGUAGE AT LARGE 9.2

A psycholinguist walks into a bar…

T ime for some jokes.

A man and a friend are playing golf at their local golf course. One of the guys is about to chip onto the green when he sees a long funeral procession on the road next to the course. He stops in mid-swing, takes off his golf cap, closes his eyes, and bows down in prayer. His friend says: “Wow, that is the most thoughtful and touching thing I have ever seen. You truly are a kind man.” The man then replies: “Yeah, well we were married 35 years.”

Two weasels are sitting on a barstool. One starts to insult the other one. He screams, “I slept with your mother!” The bar gets quiet as everyone listens to see what the other weasel will do. The first weasel yells again, “I SLEPT WITH YOUR MOTHER!” The other says, “Go home, Dad, you’re drunk.”

I want to die peacefully in my sleep like my grandfather. Not screaming in terror like his passengers.

A man put on a clean pair of socks every day of the week. By Friday he could hardly get his shoes on.

If you wanted to get analytical about humor (and why in the world wouldn’t you?), you might notice that these jokes create the eerily familiar cognitive sensation of leading you down one train of thought for a while and then, at the punch line, yanking you into a completely different mental frame of reference so that you have to reinterpret the entire situation in an unexpected way. You might think of it as the garden path approach to humor, with the punch line serving as the “disambiguating region.” Scholars who study humor often argue that such incongruous shifts are an important part of what makes something funny. The philosopher Immanuel Kant put it this way: “Laughter is an affection arising from a strained expectation being suddenly reduced to nothing.” There’s a striking parallel between the “strained expectation” you might feel in reading a garden path sentence, and the brief bewilderment that occurs before the Aha! moment in the joke.

For many jokes, the humor goes beyond merely being analogous to linguistic ambiguity—it relies on it:

A man walks into a restaurant and growls at the maitre d’: “Do you serve crabs here?” The maitre d’ responds, “We serve everyone here. Have a seat, sir.”

What do you call a fly with no wings? A walk.

There are two fish in a tank. The first fish says to the second fish, “How the hell do we drive this thing?”

Why did Ludwig van Beethoven kill his two ducks? They wouldn’t stop saying “Bach! Bach!”

Given that so many jokes hinge on recognizing and resolving conflicting interpretations, whether linguistic or otherwise, it wouldn’t be surprising to find that there’s some overlap in the neural machinery that’s involved in interpreting jokes and resolving ambiguities. We’ve seen that the frontal lobes of the brain play an important role in refereeing between competing interpretations, with a starring role for the region known as the left inferior frontal gyrus (LIFG). A brain-imaging study by Tristan Bekinschtein and colleagues (2011) found that this region was more active for jokes that rely on competing meanings than for ones that rely on absurdist thoughts or images:

Why did Cleopwa bathe in milk? Because she couldn’t find a cow tall enough for a shower.

But ambiguity alone isn’t enough to tickle the funny bone. If it were, this would also be a joke:

What was the problem with the other coat? It was hard to put on with the paint roller.

Jokes rely on ambiguity, plus a little magic. With run-of-the-mill ambiguity, the goal of the language processing system is to identify the correct interpretation and quickly and efficiently squash any competing ones. The magic in jokes probably has to do with the fact that the whole point is not to squash the competing meaning, but to keep both interpretations alive in order to enjoy the tension between them—the more interesting the tension, the better the joke. (And it’s precisely this tension that makes some unintended “crash blossoms” hilarious.) Becoming conscious of the reverberations between meanings is even more cognitively complex than basic ambiguity resolution. And indeed, the brain imaging study by Bekinschtein et al. found that the LIFG was especially active when people listened to jokes that played with ambiguity, when compared with non-funny sentences involving ambiguous words.

So there’s a great deal of enjoyment to be had from the fact that our minds activate multiple meanings at once, whether you’re a Virginia Woolf type who savors the useless multiplicity of the meanings of words, or just someone who likes to hear a good joke. If you need a good laugh, you might try visiting the LaughLab (http://richardwiseman.wordpress.com/books/psychology-of-humour/), a website in which psychologist Richard Wiseman documents his search for the funniest joke in the world. You’ll find many jokes there that rely on language processing working the way it does. Among the submissions, you’ll find one of my favorites. It hinges not on a garden path ambiguity, or even a pun due to outright lexical ambiguity, but on the more subtle lexical neighborhood activation effect, in which words that sound much like their targets are also accessed in the mind:

The monks were in the monastery copying those beautiful illuminated manuscripts. One young monk suggested that, since they’d been copying copies, it might be time to go back to the original and make sure that their copies were correct. The abbot agreed and sent the monk down into the cellar to examine the original. The monk was gone for a long time, and finally the abbot went to look for him. He found the monk in tears and asked what was wrong. Through his tears, the monk blurted out, “The word was celebrate!”

9.6 Questions to Contemplate

1. What is the connection between individuals’ scores on the reading span test and sentence processing, and why is this link not sufficient to favor the memory-based account over the experience-based account?

2. What’s a possible explanation for the fact that children have a harder time resolving linguistic ambiguity than adults?

3. If you were advising someone whether they should try to improve their reading comprehension by practicing on a task such as an n-back task, what evidence would you encourage them to consider?

GO TO

oup.com/us/sedivy2e

for web activities, further readings, research updates, and other features

DIGGING DEEPER

The great debate over the “bilingual advantage”

The previous section covered the question of whether specific training related to cognitive abilities can have a measurable effect on everyday activities like understanding language. Researchers have also asked whether the reverse is true—that is, is it possible that everyday language use can “train” the cognitive processes it recruits? Can the demands of sentence processing, for example, stretch the capacities of working memory?

Naturally, it’s not possible (or ethical) to create the sort of experiment where you randomly assign one group to become language users and another to learn something else—say painting—instead. But researchers have looked for naturally occurring groups of people who might be considered “super-users” of language to see whether they have an advantage over other language users. As a result, the spotlight has fallen on people who know and use more than one language.

What cognitive advantages might bilingual (or multilingual) people enjoy? The thing that makes the bilingual experience stand out is the need to constantly switch between two separate language systems. In Chapter 8, you learned that people who know more than one language mix them together in their minds, rather than storing and accessing each of them separately. For example, you saw that Russian-English bilinguals experienced competition from Russian words that sounded similar to the English words they were hearing. This means many words that become activated may need to be disregarded due to the fact that they belong to a language system that’s not relevant for the current exchange. Such heightened competition can make language processing slower and harder, and there’s evidence that bilinguals show less efficient word recognition compared with monolinguals (Rogers et al., 2006). They also experience interference across languages in speaking as well as understanding, and this is reflected in poorer performance in picture-naming tasks, even when the bilinguals are speaking in their first and dominant tongue (Ivanova & Costa, 2008).

But researchers have suggested that there’s an upside to bilingualism. Perhaps the regular exercise of wrestling with cross-language interference allows bilinguals to build up more muscular cognitive control abilities. In that case, we should find enhanced performance specifically on tasks that measure cognitive control.

A number of such claims have hit the journals, reporting on benefits throughout the life span. At the very young end of the life span, Ágnes Kovács and Jacques Mehler (2009) tested 7-month-old babies who were being raised in monolingual and bilingual households. They used a variant of a classic cognitive test known as the “A-not-B test,” in which infants see an object being hidden in one location, and then watch as it’s moved to a different location. In order to find the object in its new location, babies have to suppress their knowledge of the first location. This is a very basic cognitive control task that babies only accomplish reliably by age 18 months. Kovács and Mehler used a simplified version of the task in which children were first trained to look at one side of a screen, then had to learn to look at the other side before being rewarded with a nifty visual treat. The 7-month-old babies from bilingual households were better at redirecting their gaze to the new location; monolingual babies of the same age were more likely to get stuck in the pattern of looking at the old location. This is quite a dramatic demonstration, as babies at this age have yet to utter their first words!

At the other end of the life span, Ellen Bialystok and her colleagues (2004) measured the performance of adults aged 30 to 80 on a cognitive interference task known as the Simon task. In this task, a red or green square pops up on the screen, and people have to press a button on the right to indicate a green square, or one on the left to indicate a red square. The control condition of this test involves no spatial interference; the square shows up in a neutral location at the center of the screen. But in the “Simon” condition, the squares appear on either the left or right side of the screen. Sometimes the square is on the side that’s congruent with the correct button press (i.e., a red square appears on the left, requiring the subject to press the left button), while at other times the square appears on the side that’s incongruent with the button press (a red square appears on the right, requiring the subject to press the left button). In the incongruent trials, subjects have to ignore the square’s location in order to avoid its interference with their response. People usually respond more slowly to the incongruent trials than to the congruent ones, and the difference between them is taken as a measure of the degree of interference they experience.

Figure 9.13 presents the responses of subjects grouped by age. Figure 9.13A shows the control condition in which there’s no interference; you can see that the subjects show overall slower response times beginning in their 50s, but that the performance of bilingual and monolingual subjects is identical at all ages. But a different picture emerges from Figure 9.13B, which plots the degree of interference on the Simon trials. Here, the bilinguals outperform the monolinguals at all ages, and the difference becomes more pronounced as the subjects enter their 60s (and continues to grow).

Figure 9.13 Mean response times for subjects by age. (A) Response times are plotted for the control condition in which there’s no interference. (B) The degree of interference on Simon trials, calculated by subtracting the response times for congruent trials from the (longer) response times for incongruent trials. (Adapted from Bialystok et al., 2004, Psychol. Aging 19, 290.)

Bialystok and others have argued that strengthened cognitive control abilities have some important clinical implications for people suffering from Alzheimer’s disease. Along with memory loss, people with Alzheimer’s experience a dramatic loss of cognitive control. But some researchers suggest that bilingualism is a buffer against the effects of the disease, reporting that bilingualism delays the progression of symptoms, so even when they’re at a relatively advanced stage of the disease, bilingual patients typically function at a higher level than their monolingual counterparts (e.g., Alladi et al., 2013; Bialystok et al., 2007).

As you might guess, results like these generated an intense buzz of excitement, prompting many follow-up studies. Sadly, quite a few of the subsequent studies failed to find any bilingualism effects on cognitive control abilities. In fact, a recent survey of 152 studies found only a tiny effect of bilingualism on cognitive control abilities (Lehtonen et al., 2018). This state of affairs has led some researchers to argue that the so-called bilingual advantage is likely illusory.

What might explain the initial results then? Critics of studies showing positive effects have focused on a number of factors. One is the challenge of matching bilingual and monolingual populations in a way that would eliminate other possible explanations for differences in executive function. In some countries, people who speak more than one language do so because they belong to a fairly privileged class and have access to higher levels of education and travel; both education and socioeconomic status have been linked to differences in executive function. Or bilinguals may be more likely to be immigrants who were born elsewhere. In a country such as Canada, which requires most immigrants to rack up enough “points” to immigrate, showing proof of education, high-level work skills, and financial assets, the bilingual group may be endowed with a number of advantages not found among non-immigrant monolinguals. In fact, a Canadian study that controlled for ethnicity and socioeconomic status found no bilingual advantage for performance on the Simon task (Morton & Harper, 2007). In other countries, the immigrant population may have less education and lower socioeconomic status than the general population; this would make it harder to find evidence of a bilingual advantage even if it does exist.

Another difficulty is that it’s hard to interpret results that show stronger executive function in bilinguals—did the demands of bilingualism cause the enhanced performance? Or are people who have better executive function abilities more likely to become bilingual? (I know of several people who have told me that they never pursued foreign language study beyond mandatory courses in school because they were “no good at it” and were frustrated by the experience.) But scientists are notoriously persistent, and a number of researchers think it would be premature to abandon the search for a bilingual advantage. The challenge for them is to explain why so many studies have failed to show any effects.

Ellen Bialystok (2017) has argued that many of the failures to find effects have involved young adults, a group whose executive functioning abilities are very strong. This, she claims, makes it especially difficult to see effects of bilingualism on lab tasks, because these subjects are already near peak performance, leaving little room for improvement as a result of the bilingual experience; the effects may be much more pronounced among older participants or young children. Other researchers (e.g., Teubner-Rhodes et al., 2016) have emphasized that the experimental tasks need to be challenging enough for the participants who are involved. They note that a number of tasks that are used in studies may be too easy—over the course of the experiment, the monolingual group may gain enough practice with the task to catch up to the performance of bilinguals, so that even if bilinguals started at the peak without any need for practice, no overall difference between the groups would be found.

Another issue is that the term “bilingual” covers a huge and variable range of experiences. Learning a second language in childhood may have very different cognitive consequences than learning it adulthood. Some bilinguals may have a much weaker knowledge of one language than the other, so that the weaker language doesn’t interfere much with the stronger (and more commonly used) one. Some people may use both languages on any given day—even mixing them together with another bilingual conversational partner—whereas others may use their second language only when they travel to a different country, with strict separation between their contexts of use. All of these factors probably affect the extent to which bilinguals experience interference between their languages—and therefore the degree to which cognitive control is exerted.

A more productive approach might be to take into account these variations in experience. Degree of proficiency in a second language or the extent to which people mix their languages can be used as predictors for enhanced performance in executive function. For example, Julia Hofweber and her colleagues (2016) reported that German-English bilinguals who often mixed languages (resulting in sentences such as Wir haben friends gemacht mit dem shop owner) showed better performance on some executive function tasks than a group of bilinguals who rarely mixed languages this way (although the two groups also differed in other important ways).

Similarly, monolinguals take part in a broad range of cognitive activities, both linguistic and nonlinguistic. They may also be benefitting from other opportunities to exercise executive function—for example, Monika Janus and her colleagues (2016) reported that training in either a second-language or music program resulted in improvements in executive function. If a study includes monolingual participants who lead very rich cognitive lives, there may be little room to see additional effects of bilingualism.

Finally, as alluded to in the previous discussion of cognitive training, executive function is almost certainly not one ability, but several different abilities. Individuals who are given multiple tests of executive function show a surprising degree of inconsistency across these tasks, suggesting that these may be tapping into a number of different, only loosely related, abilities. If that’s true, and if bilingualism makes heavy use of only very specific subcomponents of executive function, then perhaps it’s not surprising that a large number of studies using a variety of different tasks have failed to find effects. Resolving the debate over the bilingual advantage is going to require greater precision in identifying the specific mental processes that go into cognitive control—and the widespread interest that’s been sparked by the hot debate over the bilingual advantage may inspire more researchers to devote their energies to this question.

It will be worth your while to follow this debate as it unfolds over the coming years. Not only will it give you an informed perspective from which to evaluate media coverage of this research, but it offers a good example of how scientific progress works in general. Evidence that something we thought was right may not be right after all is rarely the end of the story, but it is often a prompt to raise methodological standards and refine theories more precisely. You may find it interesting to ask your instructor whether they hold an opinion on this debate and why.

PROJECT

Find a journal article that reports a bilingual advantage effect for executive function. Drawing on the discussion in this section, provide a detailed critique of the paper. Design a hypothetical study to addresses some of its shortcomings. Be specific in your choice of participants, methods, assessment tasks and other measures, and provide a detailed rationale for your design.