We recently attended a lecture on “Artificial Intelligence and the limits of machine translation” by renowned Spanish-Argentine linguist Violeta Demonte, which was part of the Festival de las Ideas held in September in Madrid. The lecture was followed by a conversation with Spanish writer and journalist Marta Fernández, and a discussion with the audience.
A BRIEF HISTORY OF MACHINE TRANSLATION
Ms. Demonte opened by framing Machine translation as a subset of the larger field of Artificial Intelligence: a group of technologies that operate on human language, processing vast amounts of verbal data to conduct classifications, summarize, program chatbots, conduct documentary analysis and, of course, translate.
Even though translation is thousands of years old, Demonte traced the history of machine translation back to the 1950s.
The first two machine translation models were Rule-Based Machine Translation (RBMT), based on bilingual or multilingual dictionaries, as well as human-written syntactic rules, and Statistical Machine Translation (SMT), based on statistical algorithms. The first and second versions of Google Translate were respectively based on RBMT and SMT. RBMT was a costly failure; and, while SMT was a great improvement on RBMT, its results were still very poor.
Neural Machine Translation (NMT), introduced in 2015, was the great revolution in machine translation. NMT systems are modelled on neural networks, simulating the operation of the human brain. While still built on a statistical foundation, NMT employs vectors to group together sentences with similar meanings, using similarity principles, and make a decision as to the best translation choice. Systems are trained on trillions of equivalent sentences in two or more languages. The results of NMT are remarkable by comparison to the two previous models.
WILL MACHINE TRANSLATION MATERIALLY IMPACT THE TRANSLATION INDUSTRY?
An important distinction was made here between machine translation and computer-assisted translation (CAT). CAT tools, such as translation memories, are intended to optimize human translation by automating certain repetitive and formatting tasks and keeping a record of past translation choices. They are currently used by the vast majority of professional human translators. By contrast, machine translation is a process in which the translation itself is not carried out by a human being but by a piece of software.
As Ms. Demonte moved on to point out that machine translation has materially affected the translation industry is a fait accompli, rather than a possible future development. However, despite the merits of NMT, human translators have not been replaced. Actual use cases predominantly involve it as a first, “brute force” draft, particularly in certain fields (institutional, administrative, industrial), with human translators editing the output.
As to the pressing question of whether human translators might one day be fully replaced, Demonte opined that such development would be made more likely by a general impoverishment of human discourse than by technological prowess –both unlikely scenarios.
STRENGTHS AND WEAKNESSES OF NEURAL MACHINE TRANSLATION
Ms. Demonte then went on to discuss the strengths and weaknesses of neural machine translation.
First and foremost, NMT is fast, generating translations in real time. It is also cheap for users – albeit it involves an exorbitant computational expense which creates a steep barrier of entry, favouring technological companies.
Furthermore, neural models are scalable. General machine translators can be adjusted to specific domains, using smaller or larger corpora. They can also be extended or “transferred” for multilingual translations: some companies are even making their databases open-access.
Finally, NMT systems have significantly improved in terms of fluency – the feeling of naturalness and consistency conveyed by the translated text – and adequacy.
As to the weaknesses of NMT models, their results depend on the quantity and quality of the data on which they are trained. They are eminently susceptible to content bias: the content of the data determines the content of the output.
They also have difficulties with so-called “low-resource” languages, where data is scarce and English is used as a pivot. They thus tend to focus on “high-resource”, i.e. Western European languages.
As stated before, NMT models are very costly in terms of computation, resulting in high financial barriers of entry for would-be developers, as well as a considerable environmental footprint due to energy consumption.
NTM models, and Artificial Intelligence in general, lack the human competences required for translation, such as critical reading, cultural knowledge, documentation, specialist knowledge. Scaling a corpus to new languages or new domains, while possible, requires a huge effort, as words do not mean the same across all domains.
Finally, even in the case of high-resource languages like English, Spanish, German, or French, specific problems persist. Translation choices are systematically inconsistent in long texts (e.g. translating the same term alternately as living room and sitting room). “Hallucination” – the phenomenon whereby the model “makes up” translations – and errors are also common. Problems with certain grammatical structures, such as since in English and word order, recur across languages. The system, lacking cognition, fails to identify ambiguity, polysemy, metaphor, and nuance. Finally, machine translation is, almost always, overliteral and paraphrases rather than translating idioms and jargon.
THE CHARACTERISTICS OF LEGAL LANGUAGE
Ms. Demonte then discussed the characteristics of legal language and the problems they pose for machine translation.
From a linguistic perspective, legal discourse is intricate and archaic, with features that differentiate it from other specialised domains. One of its most distinctive traits is that its terminology relies on commonplace words that acquire technical meanings (e.g. causa in Spanish, liability in English). Thus, even though its purpose is precisely to avoid being ambiguous, it often is.
Legal language is also characterised by a peculiar syntax and style: an abundant use of the passive voice, long sentences, few punctuation signs, an abundance of anaphoric and cataphoric terms (e.g., hereto, thereof), syntactic breaks, and many Latinisms (e.g. prima facie, habeas corpus).
Ms. Demonte gave some examples of difficulties of machine translation when dealing with legal language. One of them is ambiguity in semitechnical terms: when, as stated, certain common terms acquire specialized meanings. E.g., causas de (“grounds for”) is often translated by machine translation systems as causes of; acción infundada (“unmeritorious proceedings”) is often translated as unfounded action.
Another common problem is that of ambiguity and false friends: e.g., complaint can be translated into Spanish as queja, denuncia, or demanda, depending on the context; liability can be translated into Spanish as responsabilidad, debilidad, or pasivos; domestic law is legislación interna, not legislación doméstica, etc.
Machine translation models also have difficulties with fixed collocations, terms that are most frequently found together in legal language, such as to file/lodge/bring an appeal, to enter into a contract. It tends to translate them directly rather than finding the appropriate terminology in the target language.
Ms. Demonte gave as an example the comparison between the translations of a section of the Mexican Civil Code by a professional human translator, by DeepL, and by Google Translate. While both machine translation systems rendered the phrase produzca provisionalmente sus efectos as the overly literal provisionally producing its effects, the human translator more accurately translated the phrase as having provisional consequences.
The consensus, according to Ms. Demonte, is that pure machine translation is inadequate for the translation of set phrases, legal formulae, and common words in a legal context: post-edition by a human appears to be acutely required.
THE PRIMING EFFECT
A significant part of specialist texts translated by machine translation systems require post-edition by a human being. The role of the translator has indeed extended to that of a proofreader, editor, and data mining expert, among other functions.
While machine translation post-editing (MTPE) seems to save time and costs, it gives rise to the priming effect: correcting a text generated by a machine has the psychological effect of inducing an acceptance of the translation “as is”, making editors less likely to spot errors and inaccuracies.
CONSEQUENCES AND ETHICAL ISSUES
Ms. Demonte closed her talk by discussing certain ethical issues pertaining to AI and machine translation.
As stated, because they come mainly from high-resource languages, training data are not representative of the global population, which results in content bias and disparity.
Furthermore, pertaining to the vast quantities of data used to train machine translation models, and AI systems in general, are trained on, there is the question of data ownership: to whom does the data belong? Existing machine translation systems rely on the work of many human translators, often without their consent. Should developers obtain consent to reuse a translation? Legally, the question is interesting, as machine translation models do not plagiarise, but rather “cook” the data they work with.
As regards the processing of such data, both machine translation and AI lack many human competencies that are required in translation. Should we accept this as a given?
Another issue is how machine translation models should be assessed. The evaluation metrics used by technology companies to assess the quality of machine translation have been strongly criticised by linguists due to their inaccuracy, lack of transparency, focus on certain questions, and lower quality than expert evaluations.
Ms. Demonte ended her presentation by posing the “apocalyptic” question of whether machine translation and AI will put an end to linguistic invention, resulting in a flattened, homogeneous language. Furthermore, will people stop learning languages if automatic translation is available at all times on your phone? (Her take: they will not).
Ms. Demonte’s was an informative, often illuminating talk on the nature, impact, benefits, and flaws of machine translation. Her lecture (in Spanish, with machine-translated English subtitles available) can be watched here: