End-to-End Lexically Constrained Machine Translation for Morphologically Rich Languages

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4019–4033 August 1–6, 2021. ©2021 Association for Computational Linguistics 4019 End-to-End Lexically Constrained Machine Translation for Morphologically Rich Languages Josef Jon and João Paulo Aires and Dušan Variš and Ondˇrej Bojar Charles University surname@mail.ufal.mff.cuni.cz Abstract Lexically constrained machine translation al- lows the user to manipulate the output sen- tence by enforcing the presence or absence of certain words and phrases. Although cur- rent approaches can enforce terms to appear in the translation, they often struggle to make the constraint word form agree with the rest of the generated output. Our manual analy- sis shows that 46% of the errors in the output of a baseline constrained model for English to Czech translation are related to agreement. We investigate mechanisms to allow neural ma- chine translation to infer the correct word in- flection given lemmatized constraints. In par- ticular, we focus on methods based on training the model with constraints provided as part of the input sequence. Our experiments on the English-Czech language pair show that this ap- proach improves the translation of constrained terms in both automatic and manual evaluation by reducing errors in agreement. Our approach thus eliminates inflection errors, without intro- ducing new errors or decreasing the overall quality of the translation. 1 Introduction In Neural Machine Translation (NMT), lexical con- straining (Song et al., 2019; Hokamp and Liu, 2017; Post and Vilar, 2018) involves changing the translation process in a way that desired terms ap- pear in the model’s output. Translation constraints are useful in domain adaptation, interactive ma- chine translation or named entities translation. Cur- rent approaches focus either on manipulating beam search decoding (Hokamp and Liu, 2017; Post and Vilar, 2018; Hu et al., 2019) or training an NMT model using constraints alongside the input (Dinu et al., 2019; Song et al., 2019; Chen et al., 2020). In inflected languages, constraints from both source and target sides may appear in numerous surface forms, which may result in errors during Likud party has merged with an even more hawkish lot under Avigdor Lieberman. Input (EN) No constraint translation (CS) Strana Likud se spojila s ještě jestřábím losem pod Avigdorem Liebermanem. Surface form model output (CS) Strana Likud se spojila s ještě radikální partou pod vedením Avigdora Liebermana Lemmatized model output (CS) Strana Likud se spojila s ještě radikálnější partií pod vedením Avigdora Liebermana. radikální radikální Figure 1: Comparison between constrained translations from English to Czech. translation. By enforcing the presence of a certain exact term on the target side, existing approaches fail to deal with word inflections. As we show, they preserve the surface form of the word provided as constraint regardless of the context. Morphologi- cally rich languages have multiple forms of each word, e.g. inflections to nouns. For satisfactory results in these languages, the constraint process- ing method needs to be capable of detecting any surface form on the source side …

Content truncated to 500 words