Article body

Let me begin by laying my cards on the table, briefly informing readers of my background and the biases that inevitably come with it. Trained as a linguist in the old Chomskyan school of generative linguistics, I joined the machine translation project at the Université de Montréal in 1977. From that point on, almost my entire professional life has been spent in research and development, both in MT and machine-aided translation (MAT), except for the few years when I earned my living as a French-to-English translator at the Canadian federal Translation Bureau. Today, I work as an independent consultant in machine translation, while continuing to translate, both for pay and for pleasure.

Hence, readers will not be too surprised to learn that the chapters that I found most interesting in this collection of articles are those that deal with MT and MAT. Chief among these are two articles by Mark Seligman, one on the evolving treatment of semantics in MT, the other, co-authored with Alex Waibel, on speech-to-speech translation. The first article is a lengthy and impressive historical overview of the role that semantics has (and has not) played in MT. Seligman opens on a philosophical note, picking up John Searle’s well-known Chinese room argument in which Searle contends that no computer program (not just MT) can ever operate with anything like a human understanding of the language it processes; all it can do is manipulate symbols. Seligman grudgingly accepts Searle’s general point, but only for those programs that operate without any explicit meaning representations. He then goes on to trace the role of semantics throughout MT’s long 70-year history, from which we learn that the great majority of MT systems have eschewed explicit semantics. Only at the end of his article does Seligman allude to a form of semantics that could potentially refute Searle’s argument: a perceptually grounded semantics in which the classes and categories employed by an MT system would be learned through artificial perception of the real world.

This is indeed an intriguing possibility, and given AI’s remarkable progress in recent years, it doesn’t appear entirely outlandish or far-fetched. My problem with Seligman’s position lies not so much in the feasibility of such an autonomous machine-learned semantics; rather, it has to do with its necessity. Simply put, neural machine translation (NMT) systems have become so good of late that one can’t help wonder how much of a difference a perceptually grounded explicit semantics could possibly make to these systems’ output quality.

As it turns out, there are several articles dealing with MT and MAT in this collection which, one might argue, appear to have been overtaken by the stunning progress made of late by neural MT[1]. The reordering techniques described by Masaaki Nagata in chapter 9 for MT between Japanese and English apply to the syntactic intermediate structures produced by statistical MT systems. As he himself recognizes at the end of his article, neural MT systems make no use of this kind of intermediate structure and have largely resolved the reordering problem that formerly plagued MT between these two very different languages. Even the pertinence of the EXPERT Project (described in Chapter 11), which set out to develop new hybrid data-driven approaches to translation, may need to be re-evaluated in the light of the quality of the output produced by today’s NMT systems. Much of the work in this EC-funded FP7 project focussed on improving what is presented as being the central component of the modern translator’s arsenal, that is, translation memories. And the project did propose sensible ways of enhancing the retrieval algorithms at the heart of a TM, as well as tackling the need for effective ways of automatically cleaning large TM databases, particularly those populated by data culled from the Web. However, there is little mention in this chapter of the crucial interplay of between translation memory and machine translation. Traditionally, the division of labour between the two has always been skewed in favour of TM: one resorted to a machine-generated translation only when no exact or approximate match could be found in the TM. Given the general quality of the output produced by previous generations of MT systems, the priority accorded to the human translations in the TM was certainly warranted. With the vastly improved quality of the output generated by current NMT systems, I would contend that that division of labour needs to be reviewed. Translators should be able to see the MT output for every sentence today, and for many texts recourse to the TM output may principally serve to verify or correct parts of the machine version.

Please don’t get me wrong. I am not claiming that the language automation problem has finally been solved, or that fully automatic high-quality translation of unrestricted texts—the holy grail—is at last upon us[2]. Nor do I mean to suggest that there is no longer any use for such tried and tested CAT tools like translation memories. What I am saying is this: the best neural MT systems are now capable of producing translations of surprisingly good quality for a wide range of run-of-the-mill texts, as well as for certain specialized texts for which they have been trained. In many situations, the quality of these “raw” translations may be sufficient for their intended use; although whenever that use is in any way potentially dangerous, delicate or compromising, the MT output still needs to be carefully revised by a qualified human translator. In many other situations, a grammatically correct, good-enough machine translation is simply not sufficient; and here, the MT output can serve as a first draft which a qualified translator can post-edit and improve upon. What I’m saying, in short, is that important changes have recently occurred in the translation automation landscape which require us to reassess the conventional view of the computer’s role in language translation.

What of the other articles in this collection? Space limitations do not allow me to comment on all of them, but I do want to consider the three case studies that are included in the volume, which the editors consider to be particularly important in that they exemplify what they call the “social turn” in empirical translation studies.

These chapters explored and demonstrated how advanced statistical methods can be deployed to construct empirical analytical instruments in order to enable socially oriented empirical translation research.

p. 255

And earlier on the same page:

Translation studies, especially the descriptive, empirical research branch, no longer needs to be limited to the documentation or description of the production, process and function of translation, when it can offer translation-based innovative, effective solutions to help address practical, pressing social issues.

p. 255

The pressing social issues addressed in the three case studies involve the communication of environmental information, the dissemination of drinking-water guidelines and the assessment of the readability of translated health-related documentation among “non-native” populations. No one would question the worthiness of these efforts. My problem with these three chapters has to do with the claims that are being made for their import to the field of translation studies. The proposed solutions are said to be “translation-based,” but in the first case study, only a cursory description is provided of the terminological data that is employed in the methodology, and that terminology in no way evolves in the course of its application[3]. On the other hand, pages and pages are devoted to describing the complex statistical mechanisms that are marshalled to arrive at conclusions that are (I’m sorry to say) not exactly earth-shattering. In one case, the country-based ranking of environmental performance derived from counts of environmental terms in various publications is found to generally correspond to the widely endorsed Environmental Performance Index[4]. In the second case, the authors discover that the mass media play a significant role in determining how drinking-water guidelines are disseminated in different industrial sectors. And in the third case study the authors conclude that the program they develop to automatically assess whether translated health notices are more or less difficult to read applies nearly as well to original health notices that aren’t translated. My problem, in a nutshell, is this: none of these three studies makes any contribution to our understanding of translation; nor does the sophisticated statistical machinery they describe contribute to facilitating the production of translations. Call me old-fashioned, but I still believe that the principal goal of any scientific field—and presumably that includes translation studies—is to push back the limits of our knowledge and increase our understanding of that field, not primarily to put itself at the service of various social causes, regardless of how worthy those may be.