Lost in Machine Translation: A Method to Reduce Meaning Loss

Proceedings of NAACL-HLT 2019, pages 437–441 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 437 Lost in Machine Translation: A Method to Reduce Meaning Loss Reuben Cohn-Gordon Stanford Noah D. Goodman Stanford reubencg@stanford.edu ngoodman@stanford.edu Abstract A desideratum of high-quality translation sys- tems is that they preserve meaning, in the sense that two sentences with different mean- ings should not translate to one and the same sentence in another language. However, state- of-the-art systems often fail in this regard, par- ticularly in cases where the source and tar- get languages partition the “meaning space” in different ways. For instance, “I cut my finger.” and “I cut my finger off.” describe different states of the world but are translated to French (by both Fairseq and Google Translate) as “Je me suis coup´e le doigt.”, which is ambiguous as to whether the finger is detached. More gen- erally, translation systems are typically many- to-one (non-injective) functions from source to target language, which in many cases results in important distinctions in meaning being lost in translation. Building on Bayesian models of informative utterance production, we present a method to define a less ambiguous trans- lation system in terms of an underlying pre- trained neural sequence-to-sequence model. This method increases injectivity, resulting in greater preservation of meaning as measured by improvement in cycle-consistency, with- out impeding translation quality (measured by BLEU score). 1 Many-to-One Translations Languages differ in what meaning distinctions they must mark explicitly. As such, translations risk mapping from a form in one language to a more ambiguous form in another. For example, the definite (1) and indefinite (2) both translate (under Fairseq and Google Translate) to (3) in French, which is ambiguous in definiteness. The animals run fast. (1) Animals run fast. (2) Les animaux courent vite (3) Figure 1: State-of-the-art neural image captioner SSNT 0 loses a meaning distinction which informative model SSNT-IP 1 preserves. Survey To evaluate the nature of this problem, we explored a corpus1 of 500 pairs of distinct En- glish sentences which map to a single German one (the evaluation language in section 2.3). We iden- tify a number of common causes for the many- to-one maps. Two frequent types of verbal dis- tinction lost when translating to German are tense (54 pairs, e.g. “...others {were, have been} intro- duced .”) and modality (16 pairs, e.g. “...prospects for this year {could, might} be better.”), where German “k¨onnen” can express both epistemic and ability modality, distinguished in English with “might” and “could” respectively. Owing to En- glish’s large vocabulary, lexical difference in verb (31 pairs, e.g. “arise” vs. “emerge” ), noun (56 pairs, e.g. “mystery” vs. “secret”), adjective (47 pairs, e.g. “unaffected” vs. “untouched”) or deic- tic/pronoun (32 pairs, usually “this” vs “that”) are also common. A large number of the pairs differ instead either orthographically, or in other ways that do not correspond to a clear semantic distinc- 1Generated by selecting short sentences from the Brown corpus (Kuˇcera and Francis, 1967), translating …

Content truncated to 500 words