Machine translation: going beyond sentence by sentence

Researchers supported by the Swiss National Science Foundation (SNSF) have come up with a new approach to improving machine translation tools such as the famous Google Translate, which processes hundreds of billions of words daily.

The computer scientists and linguists were the first to show that it is possible to improve translation systems by forcing artificial intelligence to go beyond simple “sentence by sentence” techniques. Instead, the algorithms track information contained elsewhere in the text, an approach that is currently being studied the world over. The scientists will be revealing their latest results (*) on 3 April 2017 during an Association for Computational Linguistics conference in Valencia (Spain).

Translating without understanding

“Machine translation systems don’t really understand the meaning of a text,” explains Andrei Popescu-Belis, who heads up the project as well as the Natural Language Processing Group at the Idiap Research Institute in Martigny (VS). They render one language into another according to statistical rules. And in particular, they proceed sentence by sentence. But isolated sentences frequently don’t carry enough information about the context to ensure a correct translation. The systems need to be able to take into account information in other parts of the text.”

To demonstrate their approach, the researchers settled on the question of pronouns – words like “she” or “it,” which replace other elements in the text. Often, these elements are located outside the sentence being translated, which explains the high number of errors made by machine translation systems.

Popescu-Belis provides a simple example that easily trips up the most sophisticated systems: “Ma tante a acheté une excellente voiture. Elle n’est pas très jolie.” In English, Google Translate renders this pair of sentences as “My aunt bought an excellent car. But she is not very pretty.” The tool has translated “elle” into “she.” But because this pronoun is reserved for female persons, the English-speaking reader will have the impression that it is “my aunt” who “is not very pretty.”

The statistical trap

The system has made an error because it knows that the qualifier “not very pretty” applies more frequently to people than to objects. If the qualifier is replaced by “rusty” or “broken down”, the chances are greater that the pronoun “elle” will be correctly translated by “it”.

To obtain a relevant result, automatic translation would have to consider the information contained in the first sentence. That, in a nutshell, is the approach of the system developed by the Idiap researchers together with the Departments of Linguistics at the Universities of Geneva and Utrecht (Netherlands), and the Institute of Computational Linguistics at the University of Zurich.

The researchers employ machine learning tools. During each test, they introduce or remove hundreds of parameters, which the algorithms refine, until an improvement is noted. “Put broadly, we tell the system the number of preceding sentences that it must analyse, how it must analyse them and then we proceed to testing under real conditions.”

A recruitment pool for Google

The results are encouraging, says Popescu-Belis. In language pairs like French-English or Spanish-English, pronouns lead machine translation systems to make errors about half of the time. “By forcing the system to consider information across sentences, we have managed to reduce the error rate to 30%,” says Popescu-Belis.

The implications of this research go far beyond the question of pronouns: sequence of verb tenses, word choice and register all pose challenges whose solution depends largely on the entire text as opposed to isolated sentences.

The techniques developed by Popescu-Belis and his colleagues are not yet ready for off-the-shelf tools, but they are of interest to practitioners and others in the field. “Our work has shown the necessity of going beyond sentence-by-sentence machine translation. But most important, three young researchers involved in the project are now working in the area at Google Zurich. That just goes to show the interest our approach has generated.”

Contact

Andrei Popescu-Belis
Idiap Research Institute
Centre du Parc, CP 592
1920 Martigny
Phone.: +41 27 721 77 29
E-mail: andrei.popescu-belis@idiap.ch

* * * * *

Project “MODERN” in the SNSF project database > http://p3.snf.ch/project-147653

(*) N. Q. Luong and A. Popescu-Belis: Machine translation of Spanish personal and possessive pronouns using anaphora probabilities. Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics (EACL), Valencia, 5-7 April 2017. pdf > http://publications.idiap.ch/downloads/papers/2017/Luong_EACL_2017.pdf

(*) X. Pu, L. Mascarell and A. Popescu-Belis: Consistent Translation of Repeated Nouns using Syntactic and Semantic Cues. Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics (EACL), Valencia, 5-7 April 2017. pdf > http://publications.idiap.ch/downloads/papers/2017/Pu_EACL_2017.pdf

(*) L. Miculicich Werlen and A. Popescu-Belis: Using Coreference Links to Improve Spanish-to-English Machine Translation. Proceedings of the EACL Workshop on Coreference Resolution beyond OntoNotes (CORBON), Valencia, 4 April 2017. pdf > http://publications.idiap.ch/downloads/papers/2017/Werlen_CORBON_2017.pdf

(*) A. R. Gonzales and D. Tuggener: Co-reference Resolution of Elided Subjects and Possessive Pronouns in Spanish-English Statistical Machine Translation. Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics (EACL), Valencia, 5-7 April 2017. pdf > http://www.zora.uzh.ch/136447/1/CoReferenceAwareMT.pdf