Published in November 1966, the ALPAC Report was a milestone in the history of machine translation: its influence was significant, but is now perhaps a bit overestimated. At the beginning of 1964, the funding agencies that had been financing machine translation programs in the United States1 commissioned a group of experts to create the report. Now known for having highlighted the failures of work conducted since the late 1940s, the report was clearly a follow-up to Bar-Hillel’s observations.
The ALPAC Report can easily be found online,2 and there have been several articles on the history and impact of the report (among others, see Hutchins, “ALPAC: The (In)famous Report,” 2003). Here we will discuss the content of the report and the research conducted up to the end of the 1980s, in the years following its publication.
The title of the report itself is “Languages and Machines: Computers in Translation and Linguistics.” The key focus of this short report was in fact translation needs: the usefulness of translation for relevant agencies—mostly the public sector and businesses related to security and defense; the report observes that the majority of requested translations are of negligible interest, and ultimately are either partially read or not read at all—and the costs associated with these translations. The discussion on machine translations takes up only a short five-page chapter.
The Automatic Language Processing Advisory Committee (ALPAC) was directed by John R. Pierce, an information and communication theory specialist (he had worked with Claude Shannon in particular; see chapter 5). In addition to Pierce, the committee was made up of linguists, artificial intelligence specialists, and a psychologist. None of the committee members were working on machine translation at the time of the report, though two of the members (David G. Hays and Anthony G. Oettinger) had previously been active in the field. The committee did, however, interview several machine translation specialists (Paul Garvin, Jules Mersel, and Gilbert King as representatives from private companies working on machine translation, as well as Winfred P. Lehmann from the University of Texas).
The introduction of the report mentioned two reasons that could justify financing the research out of public funds (with the exception of the National Science Foundation, the agencies financing the research in the United States were closely related to defense and intelligence agencies). Those two reasons were as follows:
The report indicated that research on machine translation clearly corresponded to the second reason (obtaining quick and effective methods at a lower cost in a relatively short period of time) and therefore proposed an evaluation of the field in this regard. This was obviously a major bias, in that failing to develop a practical and efficient solution within a short period of time did not demonstrate the uselessness of the research being conducted. Ultimately, the nature of the agencies funding the research constituted a major bias for the evaluation process.
The research teams also suffered the consequences of failing to fulfill the promises they had made since the beginning of research in the field. The 1954 demonstration (see chapter 5) suggested that a practical solution was within reach. Yet industrial attempts and public demonstrations at the end of the 1950s and from the beginning of the 1960s showed that they were far from finding a solution. In fact, this contradicts the discourse from several years earlier, when the groups suggested that machine translation could yield operational results within a few months.
We must therefore keep in mind that the report was above all geared toward evaluating the possibility of obtaining high-quality machine translation in the near future (FAHQMT, or fully automatic high-quality machine translation; see chapter 5). This gave a particular twist to the report, and later had a significant impact on the field. This perspective also explains why the first half of the report examined the large quantity of translations ordered by the agencies involved, the number of available translators, and the costs incurred. Upon reading the report, it is very clear that a practical issue was under evaluation, and that the main yardstick was cost! Research perspectives were the least of the report authors’ concerns.
In fact, the report concluded, in terms of costs, a human translator was more affordable than machine translation. At the time, human translators allowed for better and faster translations, as there was no need for additional editing (correcting a text translated entirely by machine often took longer than a direct translation carried out by an experienced translator). The report only considered translations from Russian to English, and, as a result, concluded that the need for Russian to English translation was limited. The largest “consumers” of Russian translation would do better to learn the language itself, the authors suggested. Incidentally, the report seemed overly optimistic, given that it suggested a few weeks was enough to acquire a good command of a foreign language!3
The report clearly shows that, in the mid-1960s, there was no need for machine translation. According to the report, this field had no practical interest given that there were no appropriate systems to carry out the task. The original text put it very bluntly: “There is no emergency in the field of translation. The problem is not to meet some nonexistent need through nonexistent machine translation.”
“There is no emergency in the field of translation. The problem is not to meet some nonexistent need through nonexistent machine translation.” [Alpac Report, 1966]
The report then addressed the more general question of funding machine translation. The report began with a fairly standard definition: machine translation “presumably means going by algorithm from machine-readable source text to useful target text, without recourse to human translation or editing.” The report immediately concluded that no type of automated system existed at the time of drafting the report and that no such system was conceivable in the near future.4 Georgetown’s system was specifically mentioned: after eight years of funding, the system was still unable to produce a proper translation. A professional translator still had to step in and correct the translation errors. The report emphasized that while machine translations most commonly produced a decipherable text, they were equally likely to contain mistranslations and errors. The more faults a translation contains, the more difficult it becomes to manipulate and correct the text.
To illustrate the point, the report included four translation results from Russian to English using four of the era’s machine translation systems. The translations were mediocre at best.
In his 1996 article, Hutchins recalled the notoriety of the ALPAC report, pointing out that its importance had probably been exaggerated. Research funding had already decreased at the beginning of the 1960s, a situation for which Bar-Hillel’s 1959 report was partially responsible. Consequently, the number of groups working in the field of machine translation in 1966 was much lower compared to 10 years earlier (Washington University and Michigan University, as well as Harvard, had stopped their research projects in 1962; Georgetown University, specifically mentioned in the report, had not received any financial support since 1961). Other projects were pursued after 1966, at Wayne State University and the University of Texas in particular (up until the 1970s in both cases). The report simply confirmed the decision to drastically cut back on financial support for the field of machine translation.
Hutchins also emphasized the bias of the report: it only took into account translations from Russian to English executed by American agencies, and it ignored the problems of multilingualism beyond this particular context. The nature of the report, in addition to its ambitions, needs to be examined as a whole. It was very clear that the automatic translation systems of the mid-1960s were not capable of directly solving industrial needs. Nevertheless, machine translation drew attention to many scientific issues that were hardly mentioned in the report. The report even amplified Bar-Hillel’s conclusions that a completely automatic translation system was not possible in the near future.
On a positive note, the ALPAC report did express interest in computer-assisted translations, an idea Bar-Hillel also supported. The report also pointed out, rather indirectly, the need for more fundamental research on the automatic analysis of languages. It should be noted, for example, that even Hays and Oettinger, members of the ALPAC committee, had stopped their research on automatic translations a few years earlier and instead focused on syntax and parsing. Thus, Oettinger’s 1963 report, entitled “The State of the Art of Automatic Language Translation: An Appraisal,” broadly recapped Bar-Hillel’s conclusions concerning automatic translation, but also revealed a clear interest for natural language processing.5
The period following the publication of the ALPAC report represented a break from research in the English-speaking world. Other countries continued to finance research teams, while the first commercial systems began to emerge. The technical innovations during this period were limited, following their abundance in the first decade.
The ALPAC report cemented the lack of funding in the United States in the field of automatic translation during the mid-1960s. In the United States, two groups nonetheless continued research on automatic translation (at the aforementioned Wayne State University and the University of Texas), but even there, the emphasis was on syntactic analysis to make it possible to develop rich transfer rules between languages. Other groups (like Oettinger’s group at Harvard, for example) completely abandoned automatic translation and turned toward syntactic analysis, which in some ways can be considered a logical continuation from the previous period.
Hutchins (2010) emphasizes that, contrary to the United States, numerous countries have to cope with a multilingual landscape, making it easier to justify the continuation of research in the field. Canada, in particular, opened a research center in Montreal in 1965, when the majority of American centers had already closed (in the 1970s, the center was called Traduction Automatique de l’Université de Montréal, or TAUM). The need to produce a large quantity of official documents in English and in French led to high costs, which created a strong incentive to launch research in the field. The group from Montreal quickly produced two important results: a formalism suitable for representing linguistic information, developed by Colmerauer (this formalism can be seen as a precursor of the Prolog programming language that has been since then very popular in computational linguistics and more generally in artificial intelligence) and, above all, probably the most well-known automatic translation system: TAUM-Météo (later referred to simply as Météo; see below).
In France, research continued in Grenoble, where the CETAG group (later known as CETA after the closing of the Parisian center; see chapter 5), under the direction of Vauquois during the 1960s, developed an original translation system in which syntactic relationships were represented in a language-independent logical formalism (though the system was not really interlingual, since it also used bilingual dictionaries). The research mainly focused on the translation of mathematical and physics texts from Russian to French. However, the system lacked flexibility: a problem at any level was enough to block the entire translation process. During the mid-1970s, Vauquois set off to develop a modular system, with the possibility of transferring linguistic information between two languages at different levels. This was to develop into the Ariane-78 system, and is reminiscent of the image of his triangle shown in chapter 3: an ideal translation would require a logical representation (i.e., the top of the triangle), but if an automatic translation system cannot reach this level of precision, a precise syntactic or semantic analysis is better than nothing.
In the same way that prevailing bilingualism drove Canada to finance a research center while the United States was turning away from machine translation, the need to produce translations between a growing number of languages within the European Union encouraged the European Commission to become interested in automatic translation in the 1970s. The European Union initially examined the first available commercial systems. That was how a company set up in the United States in 1968, Systran, came to present its system to the European Union in 1975. Systran then developed a prototype that integrated different languages from Europe, entering into a partnership agreement that continued throughout the 1980s. We will return to this topic when we examine Systran’s history in greater detail in chapter 14. Also at the end of the 1970s, and largely under the leadership of Vauquois, a major European research program was launched: the Eurotra project. Active from 1978 until 1992, the project emphasized the syntactic level of analysis more than the development of bilingual dictionaries. The goals of the project—initially to create an operational system—were progressively scaled down and never gave birth to a successful system. It mainly resulted in several prototypes and in the emergence of new collaborations between European research institutes. Elsewhere in the world, particularly China, Japan, and the Soviet Union, several centers were created and carried out their own research during this period.
The eventual emergence of parallel corpora (i.e., pairs of translated texts) led to the invention of new methods for automatic translation, pushing this research area in multiple new directions. This will be the topic of discussion in the following chapters.
Some of the previously mentioned research groups produced prototypes that led to commercial or operational systems.
Montreal’s research center developed the TAUM-Météo system in the 1970s, which then became simply Météo and was run by John Chandioux, an independent developer from the TAUM group. Each day, the system translated the weather forecasts in Canada into two languages, French and English, on behalf of Environment Canada. The forecasts concerned not only the country but also each of the provinces that produced several daily forecasts, resulting in a significant volume of translations. The system was operational from 1977 until 2002, translating several hundreds of thousands of weather forecasts in total—about 30 million words a year during the 1990s. Although the system was relatively classic in design, it was the first to show the possibilities of obtaining operational solutions in restricted domains. The quality of the translated texts was good: very little post-editing work was required, thus allowing for reliable, robust, and regular translations. This system played an important role in promoting machine translation, especially during a period when the field was suffering from a rather tarnished reputation.
Through the 1970s and 1980s, other research groups established partnerships with manufacturers to develop specific translation solutions. For example, during the 1980s, the University of Texas teamed up with Siemens to develop Metal, a translation system initially aimed at the German-English language pair, then gradually adapted to other languages. In Japan, most companies in the software and hardware industry launched projects to produce operational systems between Japanese and English, but some also focused on other Asian languages, such as Chinese and Korean. The semiautomatic translation of short technical texts and product leaflets (i.e., the translation by means of an automatic translation system whose results can be revised by hand) was a primary commercial objective.
A final point to mention is the emergence of the first companies specifically dedicated to automatic translation since the late 1960s. First and foremost was Systran, founded in 1968 by Peter Toma, a former member of the group at Georgetown. Thanks to contracts with American defense organizations and its commercial partnership with the European Union, Systran quickly acquired a unique status within the field (for more information, see the history of Systran, told in chapter 14). Another example is the Logos Corporation, established in 1970 with the support of the American Ministry of Defense for the purpose of translating texts from English to Vietnamese. The context of the Vietnam War suddenly led to a period of increased need for translation into Vietnamese. Logos gradually expanded the number of languages processed over several decades until it became Systran’s main competitor. The company closed in 2000, and only a translation program called OpenLogos remains; it is still available online as free software.
These companies demonstrated that there was a limited but real need for automatic translation. Translating texts (leaflets, manuals, etc.) into several languages is relatively complex and costly (it requires, for example, finding translators for different languages, making sure the translations remain up to date with product development, etc.). Small and medium-sized companies need to produce translations but cannot invest too much money in this. Hence, machine translation is often seen as a desirable technology from their point of view. Outside of this niche market, large public administrations and the defense and intelligence industries remained the primary clients of these companies. We will return to this topic at greater length when we take a look at the current machine translation market in chapter 14.