Reflections on the history and present state of MT

What it is about

An essay on how machine translation developed from the 1950s. Early researchers, limited by hardware and by weak linguistic theory, aimed low: word-for-word or dictionary-based output, pre-editing and post-editing by humans, controlled languages and narrow domains, and improvement by repeated cycles of trial and feedback. The essay describes how the demonstrations of the 1950s led sponsors to expect good quality within a few years, and how the disappointment that followed shaped the field. It then follows the move to linguistic rule systems and the return of corpus methods, and it ends by asking whether the MT community can meet the needs of a global information society. The text gives no measurements.

Ideas

I1. State the expected quality with each release: for every release, publish the exact-match rate, chrF and the copy baseline on the held-out split, so that claims are bounded the way the early researchers bounded theirs. The release page in docs/ is the place.
I2. Keep the feedback cycle explicit: each cycle in the translator records what changed, the residual before and after, and the next target class, so that the iteration the paper describes can be audited.
I3. Test controlled input: a subset of English sentence types (short declaratives, fixed forms of address) that the reader must handle exactly. Report which types fail, the way early systems were restricted to a domain.

For Mova

The essay supports the design habit Mova already uses: small goals, measured cycles, and honest limits. The controlled-input idea gives a sub-test for tools/roundtrip. The paper's quality claims are general, so no number from it is used.

Topics: Machine translation

Cite

The paper itself, not this card. BibTeX from the library's record of the work.

@inproceedings{hutchins1995reflections,
  title = {{Reflections on the history and present state of MT}},
  author = {Hutchins, John},
  booktitle = {Proceedings of Machine Translation Summit V},
  year = {1995},
  url = {https://aclanthology.org/1995.mtsummit-1.6/}
}

References

7 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

There were of course dissenters from the dominant ‘perfectionism’. Researchers at Georgetown University and IBM were working towards the first operational systems, and they accepted the long-term limitations of MT in the production of usable translations. More influential was the well-known dissent of Bar-Hillel. In 1960, he published a survey of MT research at the time which was highly critical of the theory-based projects, particularly those investigating interlingua approaches, and which included his demonstration of the non-feasibility of fully automatic high quality translation (FAHQT) in principle. Instead, Bar-Hillel advocated the development of systems specifically designed on the basis of what he called ‘man-machine symbiosis’, a view which he had first proposed nearly ten years before when MT was still in its infancy (Bar-Hillel 1951). find
Nevertheless, the main thrust of research was based on the explicit or implicit assumption that the aim of MT must be fully automatic systems producing translations at least as good as those made by human translators. The current operational systems were regarded as temporary solutions to be superseded in the near future. There was virtually no serious consideration of how ‘less than perfect’ MT could be used effectively and economically in practice. Even more damaging was the almost total neglect of the expertise of professional translators, who naturally became anxious and antagonistic. They foresaw the loss of their jobs, since this is what many MT researchers themselves believed was inevitable. In these circumstances it is not surprising that the Automatic Language Processing Advisory Committee (ALPAC) set up by the US sponsors of research found that MT had failed by its own criteria, since by the mid 1960s there were clearly no fully automatic systems capable of good quality translation and there was little prospect of such systems in the near future. MT research had not looked at the economic use of existing ‘less than perfect’ systems, and it had disregarded the needs of translators for computer-based aids. find
While the ALPAC report brought to an end many MT projects, it did not banish the public perception of MT research as essentially the search for fully automatic solutions. The subsequent history of MT is in part the story of how these is this mistaken emphasis of the early years has had to be repaired and corrected. The neglect of the translation profession has been made good eventually by the provision of translation tools and translator workstations. MT research has turned increasingly to the development of realistic practical MT systems where the necessity for human involvement at different stages of the process is fully accepted as an integral component of their design architecture. And ‘pure’ MT research has by and large recognised its role within the broader contexts of commercial and industrial realities. Hence since the early 1970s development has continued in three main strands: computer-based tools for translators, operational MT systems involving human assistance in various ways, and ‘pure’ theoretical research towards the improvement of MT methods. find
Tools for translators It is now evident that MT as such is not appropriate for translators. They do not like subservience to a machine; they do not want to revise the poor quality of MT systems. What they want are sophisticated translation tools, e.g. translation workstations, which can make their work more productive without taking away the intellectual challenge of translation. What professional translators need are tools to assist them to translate: access to dictionaries and terminological databanks, multilingual word processing, management of glossaries and terminology resources, input and output communication (e.g. OCR scanners, electronic transmission, high-class printing). For these reasons, the most appropriate and successful developments of the last few years have been the translator workstations The development of translation tools became feasible only since the 1960s, firstly with the availability of real-time interactive computer environments, then the appearance of word processing in the 1970s and of microcomputers in the 1980s and, subsequently, networking and larger storage capacities. From the computational linguistics perspective, the development of translation tools and translation workstations is not as challenging as MT itself for those with 'perfectionist' inclinations, and indeed this research has taken place primarily in non-academic environments. find
With the appearance of workstations which have appreciably aided the day-to-day work of professional translators, there are clear signs that the previous antagonism of translators to the MT community in general is disappearing. These tools are seen to be the direct result of MT research. The most recent addition has been the translation memory' facility which enables the storage of and access to existing translations for later (partial) reuse or revision or as sources of example translations; and this facility derives directly from what was initially 'pure' research on bilingual text alignment within corpus-based MT (see below). Practical machine translation All current commercial and operational systems and probably most future ones produce output which must be edited (revised) if it is to be of publishable quality. Only if rough translation is acceptable for information-gathering purposes can the output of MT systems be left untouched by human revisers. It follows that commercial developers of MT must provide adequate facilities for the revision of texts. They must also stress to customers that MT does not and cannot produce translations acceptable without revision. It is a lesson which was learnt, painfully in some cases, during the 1980s and most, perhaps all, current MT developers and system vendors stress the imperfect nature of MT output. find
MT systems have been developed on the assumption that they would be used primarily by bilinguals. In practice, the post-editing of MT output has been given to people knowing both the source and the target languages, i.e. most often to professional translators. It is a practice which has damaged the image of MT among translators: they do not wish to be revisers of poor quality output from a machine. In more recent years, the training of bilingual staff specifically for post-editing has been advocated and successfully implemented in many MT operations. The lesson has been learnt that MT post-editing should not be imposed upon professional translators; it is better to train people specifically for this role. It is now widely accepted that MT works best in domain-specific and controlled environments. In this respect, MT developers have effectively taken up the themes and suggestions first propounded by the pioneers in the 1950s. Sublanguage systems were also an early proposal - in the form of microglossaries - and since the success of Meteo in the mid 1970s have remained the focus of research systems. The control of language was in the rather crude and easily dismissable form of highly simplified ‘model English’ (Dodd 1955). The idea was largely forgotten until the practical application of Systran by Xerox in the late 1970s. Other applications of controlled input followed with the available general-purpose systems. Now it is recognised that MT systems should be designed ab initio for controlled language, and a number of independent companies outside the academic MT research community have been doing so in recent years (e.g. Volmac). The largest current development is the Caterpillar project based on the research at Carnegie Mellon University. find
One of the most important lessons from the history of MT research is that there can be no quick results. The development of operational systems is a long term commitment. Even good working prototypes with reasonably large dictionaries and many years of testing cannot be easily scaled up for fully reliable operational installation. Consequently no commercial system can incorporate the very latest methodology and technology, and this fact should be made more clearly to potential purchasers. The emphasis should be on reliability and economic viability. The construction and compilation of dictionaries is essential to the success of any and all MT systems. But there are no easy solutions. The adaption of conventional dictionaries to the needs of MT is not trivial. Much attention was paid to dictionaries in the 1950s and 1960s - indeed the 'direct translation' approach required most selectional and structural information to be accessed from lexical entries. In the years following, problems of syntactic analysis and structural transfer dominated research and the lexicon was relatively neglected for many years. Since the mid 1980s, with the adoption of constraint-based and unification grammatical formalisms there has consequently been a more ‘lexicalist’ orientation, and the lexicon has become again the focus of much current MT research. It is to be hoped that this research will both accelerate the development of MT prototypes and improve the operational flexibility of commercial systems. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/hutchins-1995-reflections-history-present-state-mt/