CHAPTER 21
VERBAL UNCERTAINTY
21.1 INTRODUCTION
UNCERTAINTY management is key to the survival of animals who must plan for the best and worst contingencies, avoid unnecessary risks, and swiftly react to novel information that change the probability of future outcomes. Accordingly, many species other than humans show signs of efficient uncertainty management (Smith et al., 2003), including apes (Neldner et al., 2015), monkeys (Smith et al., 2013), and honey bees (Perry & Barron, 2013).
Humans, though, do not only manage their uncertainty: they also engage in intense collaboration that requires them to be able to communicate their uncertainty states to their conspecifics. People do not only assess uncertainty and risks, they communicate this uncertainty to others. They do so in a broad range of interpersonal contexts, for example teaching, advice-giving, persuasion attempts, or coordination of joint action. They do so in everyday life as well as in high-stakes domains such as health, finance, environmental risks, or intelligence analysis. And most importantly, they do so verbally.
The language of uncertainty largely predates modern inventions such as probability theory or decision analysis, together with their precise numerical tools. Clearly, people had to communicate uncertainty before the advent of Bayesian models, Kolmogorov axioms, and Markov chains, and they did so using phrases such as ‘this is likely’, ‘this is quite possible’ or ‘there is a small chance of that’. In fact, such phrases are still the preferred way to communicate uncertainty (Wallsten et al., 1993; Olson & Budescu, 1997), even in the age of big data and large computational power.
Because verbal uncertainty is so closely tied to decision-making, it has been intensively studied by experimental psychologists seeking to understand how people produce and interpret probability phrases, and how those phrases subsequently affect decision processes. And because the study of probability phrases has mostly been conducted by psychologists, the methods and the findings obtained in that field have not always been solidly tied to the concerns of linguists, or the methods used in experimental pragmatics. Our objective in this chapter is to offer an overview of the methods that psychologists use to study the verbal communication of uncertainty, as well as an overview of some key findings they have obtained, in the hope that we can offer an easy entry point for linguists wishing to take the experimental study of probability phrases into their own hands.
In the service of that objective, we will devote a large proportion of this chapter to reviewing experimental methods. Our choice was to decouple methods from findings. That is, rather than reviewing actual papers, our first section will sometimes exemplify experimental methods with simulated datasets constructed for that purpose. We hope that this strategy will help us clarify methodological concepts without requiring any prior theoretical knowledge from the reader. Our second section will capitalize on the first by reviewing some key results on the production, interpretation, and use of probability phrases, without stopping at each step to explain the specific methods that were used to generate these findings. We do not provide an exhaustive review of the field in a few thousand words, but we made sure to include references to seminal work as well as notable recent reviews for readers who wish to learn more about verbal probabilities.
21.2 OVERVIEW OF METHODS
Three methods are commonly used to study verbal probability expressions, and they investigate: (i) how people map verbal probability expressions onto a numerical probability space, (ii) how people map verbal probability expressions onto an outcome space, or (iii) how people select verbal probability expressions spontaneously or choose them from ordered or unordered lists. Whereas the first two approaches are typically used to study the listeners’ perspectives (and could be considered to be interpretation tasks), the third one is commonly used to study speakers’ preferences (and could be considered to be a production task). We will now elaborate on these three approaches. For a more detailed review of the empirical literature on verbal probability expressions, see Druzdzel (1989); D. Clark (1990); Theil (2002); Teigen & Brun (2003a).
21.2.1 Mapping verbal probabilities onto a numerical probability space
The most typical method of studying verbal probabilities is mapping them onto a numerical probability space. In other words, this method ‘translates’ verbal probabilities into ‘numbers’ expressing probability (e.g. ‘likely’ is translated as ‘a 70%’ probability). Therefore, this approach is sometimes called a translation approach (Teigen & Brun, 2003b). The common underlying assumption of this approach is that people can represent any verbal expression of uncertainty quantitatively and that these representations can account for the variability in the related judgements and decision-making (Budescu & Wallsten, 2011). For instance, if you believe that ‘a horse is likely to win’ means a 55 per cent probability of winning then you will be less eager to bet a lot of money on this horse than if you believe the expression means a 90 per cent probability of winning. Researchers commonly capture three numerical representations: (i) point numerical values, (ii) interval numerical values, and (iii) membership functions. These three representations allow us to capture (to a different extent) the degree of uncertainty associated with the expressions in people’s lexicons, variability of uncertainty, and vagueness associated with the expressions.
21.2.1.1 Point numerical values
In the case of the point numerical values, people represent the probabilistic meaning of a verbal probability expression with one number (e.g. ‘likely’ = ‘70%’). The authors of one of the first empirical studies on verbal probabilities asked participants to translate numerically a list of forty-one verbal probability expressions (Lichtenstein & Newman, 1967). In their study, participants were asked to select the probability value from an interval ranging from .01 to .99 that would most clearly reflect the degree of probability for each of the forty-one verbal probability expressions. Since then, several other methods of eliciting point estimates have been employed (Reagan et al., 1989; Hamm, 1991). Nowadays, this method commonly uses a probability scale ranging from 0 per cent to 100 per cent with increments of 1 per cent or 10 per cent (e.g. Clarke et al., 1992; Harris & Corner, 2011; Juanchich et al., 2012; Sirota & Juanchich, 2015). For example, in a study assessing the effect of outcome severity on probability perception, participants provided their numerical estimates of the sentence ‘the sea level will likely rise’ on a visual analogue scale ranging from 0: it is impossible to 100: it will definitely happen (Harris & Corner, 2011). The participants who were presented with more serious outcome consequences of the sea level rise provided (on average) a higher probability than those who were presented with less serious consequences of the sea level rise. Another example with different verbal anchors is the scale used in Sirota & Juanchich (2015). Participants estimated the probability communicated by a set of verbal probabilities (i.e. ‘in your opinion, what is the probability communicated by the expressions presented here?’). The participants positioned a cursor on a visual analogue 0–100 per cent probability scale (increments of 1) with three verbal/numerical probability anchors: 0 per cent, no chance, 50 per cent, and 100 per cent certain. In Figure 21.1, we can see the group mean point estimates for each probability expression (indicated by a circle), 95 per cent confidence intervals (indicated by the grey error bars) surrounding the point estimates, medians (indicated by the black bold horizontal line), as well as individual responses (small ticks on the scale) forming a density function (shaded area). We can see that, for example, the point representation of the expression ‘evenly possible’ is around the middle point of the scale and varies only minimally, whereas the expression ‘probable’ is around 70 per cent. We can also notice a large inter-individual variability in translating these expressions. For some people clearly ‘a very small probability’ can mean 20 per cent and for some others the same phrase could mean a 5 per cent probability. Group averages (and medians) nevertheless indicate that these expressions can be ordered and that they are quantitatively distinct from each other.
21.2.1.2 Numerical intervals
Sometimes, researchers are interested in knowing not only the typical numerical value of a certain verbal probability expression but also the lowest and the highest numerical value associated with it (thus creating a range of values). This might be useful to capture in certain situations. Indeed, the expression ‘likely’ could have the same typical value for two people (e.g. around 75%) but for the first person its meaning could be very wide (e.g. ranging from 51% to 99%), whereas for the other its numerical meaning could be very narrow (e.g. ranging from 70% to 80%). To capture this range, participants are asked to represent verbal probability expressions with more than one number. This method was used, for instance, in the cross-cultural study of Budescu et al. (2014), where participants from different countries read a set of probability sentences chosen from the Report of the Intergovernmental Panel on Climate Change (IPCC). For example, they read this sentence: ‘It is very likely that hot extremes, heat waves, and heavy precipitation events will continue to become more frequent.’ Then they provided the minimal, best, and maximal estimate of the numerical meaning of the expression ‘likely’. The individual values for the minimal, best, and maximal estimates for different predictions and for each country were compared with the numerical meaning that was intended by the IPCC (e.g. ‘likely’ should mean a probability between 66% and 100% according to the IPCC guidelines). The minimal and maximal estimates created ranges that were then mapped onto the IPCC expected ranges to show that only 18.5 per cent of the ranges overlapped with the IPCC expected ranges.
FIGURE 21.1. The point numerical estimates of seven probability expressions (adjusted from Sirota & Juanchich, 2015)
Note: The black lines represent the overall median probability per expression. The grey error bars represent 95% confidence intervals of the mean (the means are indicated by the circle). The polygons represent the density shape.
21.2.1.3 Membership functions
The most complex representation of the numerical meaning of verbal probability expressions is the membership function (Budescu et al., 2003; Karelitz & Budescu, 2004). One can think of the function as a rule that assigns to each element a number from the [0, 1] interval. The number indicates the degree to which that element is a member of (i.e. can convey) a particular probability category (Wallsten et al., 1986a). In other words, the membership function describes to what extent a verbal probability expression expresses the numerical probability categories 0, 0.1, 0.2,… 1. So, if a verbal expression is not perceived to convey a given probability category then it will be assigned a value close to 0, whereas if it is perceived as a prototype of the category then it will be assigned a value close to 1. For instance, an expression such as ‘almost impossible’ does not convey well a ‘100 per cent’ probability, so therefore it will be assigned a membership degree value close to 0; whereas its membership degree value will be approaching 1 for the ‘0 per cent’ probability category. By this means the membership function allows us to measure the fuzziness/vagueness of the terms (Wallsten et al., 1986a).
Prior research focused on two classes of elicitation of membership judgements: the Single Stimulus Method and the Multiple Stimuli Method (Budescu et al., 2003). In the Single Stimulus Method, participants were typically presented with one probability expression and one numerical probability at a time; they assessed to what extent the phrase described the probability on a Likert scale ranging from ‘not at all’ to ‘absolutely’. The same expression (e.g. ‘unlikely’, ‘likely’) was presented and assessed for all the probabilities from 0 per cent to 100 per cent by increments of 10; thus, the expression was assessed repeatedly eleven times. In the Multiple Stimuli Method, all eleven probabilities were presented simultaneously with the verbal expression and the participants indicated how well different probability values (e.g. 0%, 10%, 20%,… 100%) represented the verbal probability expression. Assessing probability membership by matching verbal probabilities with different numerical probability levels (e.g. 10%, 20%, 30%,…) is not the only elicitation technique available; alternative techniques are sometimes employed to elicit the membership judgements (e.g. lotteries or spinners; Wallsten et al., 1986a; Budescu et al., 2014). For example, in the pair-comparison technique (Wallsten et al., 1986a), participants saw a verbal probability and two spinners underneath; each spinner had a different proportion of white and black areas. Participants were asked to imagine a pointer and the probability that it would land over the white area if the spinner was spun randomly. The different proportions of the white area represented a different probability of a pointer landing in the area (Wallsten et al., 1986a). Participants were then asked to indicate (e.g. by moving the slider on a straight line), which spinner (and to what extent) better described the verbal probability. All membership functions (exemplified in Figure 21.2 with simulated data) can be described by three parameters, such as peak (the average probability of the highest membership value), range (Max–Min; where Min (Max) is the lowest (highest) probability with a non-zero membership), and asymmetry, also called skewness (the difference between the support of the function to the right and left of the peak: |Max – Peak| – |Min – Peak|). These parameters can be compared and analysed across experimental manipulations.
FIGURE 21.2. Membership functions for two verbal probability expressions: ‘unlikely’ and ‘likely’ (simulated data)
21.2.2 Mapping verbal probabilities onto an outcome distribution space
An alternative methodological approach to studying verbal probabilities is represented by the so-called Which Outcome approach (Juanchich et al., 2013b; Teigen, Juanchich, & Riege, 2013; Teigen, Juanchich, & Filkuková, 2014; Juanchich & Sirota, 2016). In the probability translation approach, researchers typically focused on the probability that an outcome will occur or not (e.g. it is likely that it will rain). In the Which Outcome approach, verbal probabilities are studied in association with the different levels/magnitudes of continuous outcomes (e.g. How much will it rain?). In the Which Outcome approach, participants are typically provided a prediction to complete and a range of outcomes. For example, they could be given the incomplete prediction: ‘it is likely that the sea level will rise…inches’ and a bar chart showing the frequency of sea level rise between 8 and 20 inches (see Figure 21.3). Standard categories were used to code the outcome chosen by participants to facilitate comparison of outcome selected across scenarios and experiments. (e.g. ‘out of range outcome’, ‘minimal outcome’).
FIGURE 21.3. How much will the sea level rise? Example of ‘Which Outcome’ vignette (Juanchich & Sirota, 2017)
21.2.3 Verbal probabilities, spontaneous production, and selection
There is a plethora of research in which participants select verbal probability expressions (on a scale) to measure subjective probability. However, here we are focusing on the research that is centred on understanding the use and interpretation of verbal probability expressions. Participants may generate whole sentences, complete unfinished sentences, or simply choose a verbal probability from unordered or ordered lists (see reviews by Druzdzel, 1989; Clark, 1990; Hamm, 1991). For instance, participants read a scenario describing a friend buying a used car and they were asked to communicate to him a precise numerical probability (e.g. ‘60%’) that his car will break down in the next few months (Juanchich & Sirota, 2013; Sirota & Juanchich, 2015). To communicate the probability, they were asked to use one of the verbal probability expressions depicted in Figure 21.1. Tactful speakers selected a verbal expression that conveyed a lower probability compared to plain speakers (Juanchich & Sirota, 2013). Similarly, in order to study the directionality of verbal probability expressions, Teigen & Brun (1995) studied how specific probability values of success (or failure) affected the selection of verbal probability expressions. Participants read a short scenario describing the success (failure) rate of an event (e.g. 20% of the students pass this exam) and then selected those expressions that sounded right for the given situation from a predefined (randomly presented) list.
21.3 OVERVIEW OF FINDINGS
We start by summarizing key results on the numerical meaning of verbal probabilities, paying special attention to the huge inter-individual differences in interpretation. Next, we explore some contextual factors that affect numerical interpretation. We then move on to the notion of directionality, a pragmatic dimension of verbal probabilities that impacts reasoning, decision-making and responsibility judgements independently of numerical interpretation. Finally, we review evidence about the way people use verbal probabilities, and in particular the outcomes that verbal probabilities are used to qualify.
21.3.1 Key results on numerical interpretation
Research on verbal probabilities thoroughly explored the degree of certainty they convey, as mapped on 0-100 per cent probability scales. This mapping represents the numerical meaning of verbal probabilities. Numerical probabilities are more readily comparable than verbal probabilities, and can be easily combined together or with other values that are important in decision-making, such as the subjective utilities of different potential outcomes. Accordingly, decision-making would be easier if verbal probability expressions could simply be translated into a numerical meaning. Unfortunately, this is hardly the case.
The previous section of this chapter described different methods used to elicit the numerical meaning of verbal probabilities, from simple single point estimates to membership functions. All those methods offer evidence that the numerical meaning of verbal probabilities is vague (Lichtenstein & Newman, 1967; Beyth-Marom, 1982; Clarke et al., 1992; Budescu et al., 2003).
The largest numerical meaning variation is observed between individuals. There is no between-subjects agreement on the numerical meaning of verbal probabilities, whatever the method chosen to assess it. For example, in one of the seminal works on the numerical translation of verbal probabilities (Lichtenstein & Newman, 1967), some of the participants reported that very likely meant a 45 per cent probability whereas others reported that it meant a 99 per cent probability. Similarly, unlikely could be interpreted as meaning a 1 per cent probability and up to a 45 per cent probability. The boundaries of the probability ranges are sometimes hard to believe. For example in Budescu et al. (2003), the numerical translation of impossible ranged from 2 to 75 per cent. This is a source of concern because people tend to underestimate the variability of the meaning of verbal probabilities (Brun & Teigen, 1988).
There are, however, some regularities in the way people interpret verbal probabilities. For example, a negative prefix (e.g. unlikely) conveys a lower numerical meaning than a lexical negation (e.g. not likely) (Reyna, 1981) and both these negative verbal probabilities are vaguer than their positive counterpart (e.g. likely) (Smithson et al., 2012). As shown in Table 21.1, seemingly symmetrical verbal probabilities (e.g. likely and unlikely) do not convey symmetrical numerical meanings (Beyth-Marom, 1982; Kong et al., 1986): the probability of a likely outcome and the probability of an unlikely outcome add up to less than 100 per cent (Beyth-Marom, 1982).
Table 21.1. Examples of numerical meaning for the verbal probabilities likely and unlikely, with and without modifier
In parallel, intensifiers (very, much, highly) modulate the numerical meaning of low and high verbal probabilities in opposite directions: they decrease the numerical meaning of low verbal probabilities (a very small chance) and increase the numerical meaning of higher verbal probabilities (very probable). This modulation is smaller for low verbal probabilities than for high verbal probabilities (Lichtenstein & Newman, 1967; Reyna, 1981). For example, very decreases the meaning of unlikely by nine percentage points whereas it increases the meaning of likely by fifteen percentage points (see Table 21.1). Verbal probabilities featuring intensifiers cover the extremes of the probability range, and are thus used to convey a narrower range of probabilities (Teigen & Brun, 1995). Accordingly, phrases conveying very low or very high probabilities are perceived as less ambiguous (O’Brien, 1989).
Even when these regularities are taken into account, the huge variability in the numerical interpretation of verbal probabilities is a very robust phenomenon. It has been replicated with a wide range of uncertainty phrases across languages, contexts, and samples. Evidence that verbal probabilities convey vague numerical meanings was, for example, found in British English (Juanchich et al., 2012) and American English (Budescu et al., 2003), as well as in Portuguese (Teixeira & Silva, 2009), Spanish (Cohn et al., 2009), Hebrew (Beyth-Marom, 1982), French (Juanchich et al., 2013a), Norwegian (Brun & Teigen, 1988), and Japanese (Honda & Yamagishi, 2006).
In the largest study conducted on verbal probabilities, Budescu et al. (2014) surveyed more than 10,000 people from twenty-five countries, covering a total of eighteen languages. Participants provided the numerical probability translation of eight climate change predictions such as: ‘Over the past 3,000 to 5,000 years, oscillations in global sea level on time-scales of 100 to 1,000 years are unlikely to have exceeded 0.3 to 0.5m’. This unlikely climate change event could be perceived as having a 0 per cent chance for one person and an 80 per cent chance for another. The authors also described the data as being regressive in the sense that low probabilities were overestimated and high probabilities were underestimated.
Most research on the numerical interpretation of verbal probabilities was done with psychology students, but recent research used more diverse samples accessed through online platforms such as Amazon Mechanical Turk (Juanchich et al., 2013a), or representative samples provided by other companies (Budescu et al., 2012). All samples tell the same story. The numerical meaning of verbal probabilities largely varies from one person to the next, whether the person is a student (Budescu et al., 2003), an accountant (Doupnik & Richter, 2004), an expert (Beyth-Marom, 1982) or a child (Mullet & Rivet, 1991). Findings focusing on samples with different professional backgrounds show similar patterns of variability. For example, a large between-subjects variability was found in samples of general practitioners (Brun & Teigen, 1988; O’Brien, 1989), military intelligence officers (Dhami et al., 2015), political forecasters (Beyth-Marom, 1982), and judges (McCauliff, 1982). Hence, expertise, or lack thereof, cannot explain variability. Evidence comparing samples with different levels of expertise support this point. For example, verbal probabilities translated by medical students and general practitioners showed similar numerical ranges (Kong et al., 1986).
Arguably, single point estimates of numerical meaning may vary because they are taken from a consensual range of acceptable meanings. For example, one person may translate a chance as 20 per cent, another as 30 per cent, but both would agree that 20-30 per cent is an acceptable range of meaning for the expression. People are indeed capable of providing ranges when interpreting verbal probabilities, and even degrees of membership for the different values within the range: for example, when they say that likely can mean 50 per cent, but would be even more appropriate for a 70 per cent probability (Beyth-Marom, 1982; Kong et al., 1986; Budescu et al., 2003). However, the shape and location of these membership functions also change from one person to the next. For example, people consider that beyond reasonable doubt is best translated as 96 per cent (Dhami, 2008), but the Multiple Stimuli Method yields a range of fifty-two for this phrase: this means that ‘beyond reasonable doubt’ can mean fifty-two different numerical probabilities on a scale from 0 to 100 per cent.
One could think that the variability in the numerical interpretation of verbal probability may be amplified when context is minimal or totally absent. Surprisingly, the opposite is true: verbal probabilities presented without context are interpreted more consistently than when they are presented in TV news or medical reports (Beyth-Marom, 1982; Brun & Teigen, 1988). In the work of Brun & Teigen (1988), participants judged the probability conveyed by nine verbal probabilities, either based on the written verbal probability out of context or based on sentences recorded from TV news. Results showed that for eight out of nine verbal probabilities, the variability was larger in the TV news condition than in the no-context condition.
Importantly, even though the numerical meaning of verbal probabilities is extremely vague, verbal probabilities can still be ranked in a broadly consistent manner, across samples and studies. For example, even though people cannot agree on the meaning of a chance, they can still agree that it conveys a lower numerical meaning than possible or likely (Beyth-Marom, 1982; Mullet & Rivet, 1991). The consistency in these rankings is especially good when discriminating between seven levels of probability (Beyth-Marom, 1982).
The outcome that is qualified by a probability phrase is an obvious candidate for influencing the numerical translation of the phrase. In this section, we consider two outcome characteristics that affect the interpretation of verbal probabilities: the outcome base rate and its desirability.
The base rate of an outcome, also known as incidence or prevalence in medicine, refers to how often it has occurred in the past, whether it was rare or frequent. The base rate is a strong positive predictor of probability perception (Wallsten et al., 1986b; Weber & Hilton, 1990). For example, ‘a small chance of rain in the UK’ is perceived as meaning a higher probability than ‘a small chance of rain in Spain’ because rain is more common in the UK than in Spain. Similarly, in Budescu et al.’s (2012) experiment, people who believe in climate change attribute a higher numerical meaning to a ‘likely sea level rise’ than people who do not. The desirability of the outcome can also affect the interpretation of the probability phrase that qualifies it, leading to what is commonly called the severity effect. One speaks of a severity effect when people give different numerical translations of the same probability phrase, as a function of the severity of the event it qualifies; for example, when events of increasing severity receive higher probabilities even though they are described with the same probability phrase. In a medical context, the severity effect means that two diagnoses or prognoses described as probable will be assigned different probabilities, if they are of different severities. For example, when assessing the prognoses ‘you will probably suffer from insomnia’ and ‘you will probably suffer from deafness’, people tend to assign a moderate probability to insomnia, and a high probability to deafness (Bonnefon & Villejoubert, 2006; Pighin & Bonnefon, 2011; Pighin et al., 2011). Similarly, in a forensic context, when given the predictions ‘the offender will possibly return to the crime scene’ and ‘the offender will possibly kill again’, people tend to assign a greater numerical probability to the offender killing again than to the offender returning to the crime scene (Villejoubert et al., 2008). Similarly again, in an environmental risk context people thought that a traffic accident described as ‘unlikely’ was more probable when it would release toxic uranium and kill thousands, than when it would cause a minor traffic delay (Harris & Corner, 2011).
Research identified two distinct but non-exclusive explanations of the severity effect: the asymmetric loss account and the politeness account. At the core of the asymmetric loss account is the idea that the more severe an outcome is, the worse it is to underestimate its probability, since such an underestimation could prompt a decision-maker to forgo protective measures. Accordingly, it might be rational for a decision-maker to overestimate the probability of severe outcomes, for which the expected cost of not taking protective measures outweighs their intrinsic cost. The severity effect would then reflect this rational miscalibration (Harris et al., 2009; Harris & Corner, 2011).
In parallel, Bonnefon & Villejoubert (2006) offered a politeness account of the severity effect, which was later extended to other quantifiers such as some (Bonnefon et al., 2009; Feeney & Bonnefon, 2013). The core idea of the politeness account is grounded in linguistic pragmatics. Usually, probability phrases such as possible receive a maximally informative interpretation. More precisely, listeners assume that if the speaker qualifies an event as possible, then she was not in a position to use a stronger phrase such as very likely. Accordingly, they infer that the probability of the event cannot be very high. But if the event is severe, scary, or particularly bad news, then listeners may assume that the speaker used the term possible as a way to sugarcoat the bad news, and not to convey genuine uncertainty. Accordingly, they would assign a high probability to the severe event. This account was supported by experimental data contrasting low-severity events (‘you will possibly suffer from insomnia’) and high-severity events (‘you will possibly suffer from deafness’). When the event was more severe, listeners were more likely to think that the speaker used the word possibly in order to be tactful, and this interpretation mediated the larger numerical probability that they assigned to the high-severity event. Later experiments confirmed that the perceived intention to be polite was a strong determinant of the severity effect, and that people spontaneously used probability phrases to soften bad news rather than convey genuine uncertainty (Juanchich et al., 2012; Sirota & Juanchich, 2012; 2015; Juanchich & Sirota, 2013).
This pragmatic account of the severity effect draws attention to the possibility of ‘polite misunderstandings’ (Bonnefon et al., 2011). The problem is the following: when we find ourselves in high-stakes situations where bad things are likely to happen, and where mistakes have severe consequences, we would want people to communicate their uncertainty as clearly as possible; but when people use uncertainty terms about bad outcomes and bad mistakes, we do not know whether they mean these terms genuinely, or if they use them as a politeness strategy. As a consequence, the very situations where we would like people to communicate uncertainty clearly are the same situations in which they start communicating uncertainty ambiguously. In recent years, this phenomenon has been studied in the context of medical tutoring, online instructional communication, and the assessment of medical residents (Bromme et al., 2012; Brummernhenrich & Jucks, 2013; Jucks, Päuler, & Brummernhenrich, 2014; Ginsburg et al., 2015). Data suggested that training communicators to eschew nonproductive politeness could help in some contexts, but that it also came at a cost, since they were evaluated less positively than tactful yet imprecise communicators (for more on politeness, see Holtgraves, Chapter 30 in this volume).
21.3.2 Directionality
Beyond their numerical meaning, verbal probabilities convey a key pragmatic signal: directionality. The directionality of verbal probabilities attracts the attention of recipients towards the occurrence (e.g. the outcome is likely) or non-occurrence (e.g. the outcome is not certain) of an outcome. This is akin to describing a glass as half-full or half empty: it contains the same amount of water, but the perspective on that amount changes.
Table 21.2 displays some verbal probabilities of positive and negative directionality. The numerical meaning of verbal probabilities varies widely but, in contrast, their directionality is stable. Each verbal probability is perceived as being either positive or negative and this judgement is made consistently across individuals, contexts, and grammatical variations. Directionality is not context-dependent and is not affected by modifiers such as very (Budescu et al., 2003). In contrast, numerical probabilities do not have a clear directionality (Teigen & Brun, 2000).
Some researchers proposed that directionality is simply a consequence of the membership function of verbal probabilities (i.e. location of the peak and skewness of the function) (Budescu et al., 2003). Others believe that it is a separate feature (Teigen & Brun, 1995, 1999; Moxey & Sanford, 2000; Honda & Yamagishi, 2006, 2009), and there is evidence in favour of both views (Budescu et al., 2003; Juanchich et al., 2010; Juanchich et al., 2013a). Rather than reviewing this literature, we will focus on three key issues: the overall preference for positive directionality, the determinants of directionality choice, and the impact of directionality on reasoning and decision-making.
Table 21.2. Examples of positive and negative verbal probabilities conveying a range of numerical probabilities
Overall, people have a preference for positive directionality (Juanchich et al., 2010). As shown in Table 21.2, positive directionality can fit a broad span of the probability scale whereas negative directionality is less flexible, more often selected for low to medium numerical probabilities (Reyna, 1981; Reagan et al., 1989; Teigen & Brun, 1995). The higher proportion of positive directionality phrases may be explained by markedness (Clark, 1969; Clark & Clark, 1977; Dressler et al., 1987). Unmarked words are more often used because marked terms apply to less frequent situations when a change in the status quo of the concept can apply. Markedness can take the form of negation prefixes, and hence most negative directionality phrases can be considered as marked (e.g. unlikely, uncertain, unsure), whereas the positive directionality phrases can be considered as unmarked (for further consideration on negation, see Tian & Breheny and Xiang, Chapters 12 and 26, respectively, in this volume). Marked words are also more complex, harder to process and are acquired later than their unmarked counterpart (Dressler et al., 1987), and the numerical meaning of negative directionality phrases is more variable (Smithson et al., 2012), potentially because of this processing difficulty.
The choice of a negative directionality phrase is affected by the perceived likelihood of the outcome. Low probabilities are more often described with negative directionality phrases (Juanchich & Villejoubert, 2009; Juanchich et al., 2013a). Directionality can also reflect trends, just as when a glass is described as half-empty because it was previously full (Sher & McKenzie, 2006). For example, to express a 40 per cent probability, people generally prefer a positive directionality phrase, but they select a negative directionality phrase when the probability was previously 70 per cent (Juanchich et al., 2010). Similarly, negative directionality phrases are more likely to be selected in order to contradict downward the view of a conversation partner, as in ‘it is not certain that it will rain’ in response to ‘it is likely it will rain’ (Juanchich et al., 2010).
Directionality affects the way people reason with probabilities, and the extent to which they demonstrate well-known biases such as the conjunction fallacy (Teigen & Brun, 1999). But most importantly, a speaker’s choice of directionality has large consequences for the decisions made by the listener. In that sense, directionality can be understood as a tool for speakers to nudge listeners toward a course of action. In a seminal paper, Teigen & Brun (1999) compared the hypothetical decisions made by listeners exposed to the same information may expressed with either positive or negative directionality. For example, how likely would you be to recommend a medical treatment, knowing that ‘it is quite uncertain’ that it will work, or knowing that ‘there is some possibility’ that it will work? Even though people interpret both statements as conveying a 30 per cent chance that the treatment will work, they are 91 per cent likely to recommend the treatment that will possibly work, and only 32 per cent likely to recommend the treatment that is quite uncertain to work. This effect reflects the fact that listeners may infer the preference of the speaker from the directionality of the sentence. Positive directionality means that the speaker herself would recommend the treatment, whereas negative directionality could be perceived as meaning that the speaker herself would be reluctant to recommend the treatment.
21.3.3 Continuous outcomes
So far we have only considered probability phrases that qualified binary outcomes, for example ‘it is possible that it will rain’ or ‘I will probably pass my exam’. But probability phrases can also qualify continuous outcomes, for example ‘it is possible that it will rain 4 mm’, or ‘I will probably score higher than 70/100 on my exam’ (for more on modified numerals, see Nouwen et al., Chapter 11 in this volume). Recent research has investigated the way people choose probability phrases when they qualify such continuous outcomes (Juanchich et al., 2013b; Teigen et al., 2013; Teigen et al., 2014; Juanchich & Sirota, 2016).
In a typical task, participants study a figure showing the frequency distribution of the different values of a continuous outcome. For example, they may study a histogram showing the observed duration of a hundred laptop batteries, as in Figure 21.4. This figure makes it clear that none of the batteries lasted four hours or more, that most of them lasted between two and three hours, and that very few lasted less than two hours. The task given to participants is to complete sentences such as ‘it is certain/possible/unlikely that the battery lasts for…hours’.
One might assume that participants would select durations that broadly match the probability conveyed by the probability phrase: for example, that they would pick a duration that occurs in 50 per cent of the cases for the phrase possible, or a duration that occurs in 20 per cent of the cases for unlikely. This is not the case, though. What participants do instead is to choose extreme, or even impossible durations (Figure 21.4). For example, they typically offer sentences like ‘it is unlikely that the battery lasts for 4 hours’: that is, they associate an out-of-range outcome to a phrase that typically describes events with a 20 per cent probability of occurrence.
These results are not tied to displaying frequency distributions as histograms, and can be obtained when the data are presented either in table or text form (Teigen et al., 2014). They can also be obtained in more consequential contacts than battery life, for example, sea level rise due to climate change (Juanchich & Sirota, 2017).
FIGURE 21.4. Distribution of quantities provided to participants in the outcome selection task and typical answer provided for different verbal probabilities (Teigen et al., 2014). Dotted lines show the usual numerical meaning of the verbal probabilities as derived from translation studies (e.g. Beyth-Marom, 1982)
The preference for predicting extreme outcomes is likely driven by their informativeness and by their objective frequency (Teigen et al., 2014). Indeed, skewed or U-shaped distributions showed that extreme outcomes are still preferred but to a lesser extent than when those extreme values are rare (Juanchich et al., 2013b). Furthermore, preferences for extreme outcomes can shift based on the information needs of the listeners. For example, when describing the possible price of a house to a seller, participants select the highest price, whereas when they described the possible price to a buyer, they selected the lowest price (Teigen et al., 2014). Nevertheless, the choice of out of range values is harder to explain. Why do people report that they interpret unlikely as meaning a 20 per cent probability, if they use that term to describe outcomes whose frequency is zero? And does that mean that people who hear that some event is unlikely will attribute it a probability of 20 per cent, while the communicator actually knew that the frequency of the event was zero? This miscommunication issue can have dramatic consequences, for example when it applies to predictions of the reach of a volcanic eruption (Jenkins et al., 2016). Using numerical probabilities alongside verbal ones can help people form predictions that do not unduly focus on extreme outcomes (Juanchich & Sirota, 2017).
21.4 CONCLUSION
Verbal uncertainty is a key component of interpersonal decision-making: whether to give advice, to coordinate, or to persuade, people must convey their state of uncertainty about relevant outcomes. Because quantitative evidence is rarely available in everyday life, verbal uncertainty is conveyed with the use of phrases such as ‘a little chance’ or ‘most probably’, rather than numerical expressions such as ‘a .09 probability’ or ‘a 23/79 chance’. Unfortunately, the meaning of probability phrases such as ‘probably’ is far from being consensual or stable. There are large differences in the interpretation of these phrases, from one individual to another, or from one context to another. To tackle this complexity, psychologists interested in judgement and decision-making have developed a complex toolbox of experimental measures, and amassed a rich trove of experimental results. We hope that this chapter can serve as an entry point for experimental linguists who may wish to familiarize themselves with these methods and results.