Machine Translation for Open Science – MaTOS

MaTOS is a multidisciplinary project, bringing together teams from a variety of scientific backgrounds. The studies carried out are based, on the one hand, on methods for analyzing specialized corpora to build up inventories of terms and study their usage, whether on the scale of a document, a collection, or a specific time period.

They are also based on the development of state-of-the-art machine translation systems, capable of managing extended translation contexts (from several sentences to several paragraphs, and beyond) and of handling expert resources (specialized dictionaries, translation memories). These systems are specialized for several scientific fields.

The evaluation work is based on the one hand on state-of-the-art automatic metrics, in their most recent developments, integrating, for example, large-scale neural models; on the other hand, it relies on the mobilization of post-editors specialized in their scientific field, recruited to carry out revision tasks that would be representative of a typical scientific writing activity.

After two years, the project has first produced a set of reports documenting the state of the art, focusing notably on:

- human assessments of translation quality

- automatic evaluation of document translation

- computational architectures for document translation.

Various resources have also been collected, prepared, and formatted. These include terminologies for two specialized domains, as well as various monolingual and bilingual corpora, in particular long documents and their translations (for abstracts and full-text articles) for the same two domains.

Software developments have focused on three aspects:

- the development of tools to identify terms and their variants in corpora; these will be used to thoroughly document the spectrum of acceptable terminological variations in academic documents, and to evaluate the degree of unacceptable variation ;

- the study of methods for automatically proposing neologisms to translate emerging terms;

- the development of specialized MT systems for the translation of long scientific documents, based on both encoder/decoder architectures and large multilingual language models.

In terms of evaluation, two pilot studies involving the post-editing of automatically translated abstracts have been carried out with the involvement of specialized translators and members of the academic community, in anticipation of a larger-scale study.

This work has already resulted in a dozen publications, which are available on the project website, and a number of corpora, which are also distributed via the same channel.

Capitalizing on the resource-building work already completed or in the process of being finalized, the main prospects for the end of the project are as follows:

- analyze terminological variation in source texts and its correlates in human and machine translations;

- develop specialized MT systems that translate entire documents, integrate terminological constraints, and also produce coherent, cohesive texts that require minimal proofreading for publication ;

- develop automatic methods to characterize and evaluate the ability of MT systems to (a) correctly translate segments involving specialized terms; (b) produce terminologically consistent translations ;

- implement very large-scale evaluations of the systems and metrics thus designed, and using them in post-editing protocols involving large communities of users.