Neural Machine Translation with Inflected Lexicon

Neural Machine Translation with Inflected Lexicon Nowakowski Artur artur.nowakowski@amu.edu.pl Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Poznan, 61-614, Poland Jassem Krzysztof jassem@amu.edu.pl Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Poznan, 61-614, Poland Abstract The paper presents experiments in Neural Machine Translation with lexical constraints into a morphologically rich language. In particular, we introduce a method, based on constrained decoding, which handles the inflected forms of lexical entries and does not require any mod- ification to the training data or model architecture. To evaluate its effectiveness, we carry out experiments in two different scenarios: general and domain-specific. We compare our method with baseline translation, i.e. translation without lexical constraints, in terms of translation speed and translation quality. To evaluate how well the method handles the constraints, we propose new evaluation metrics which take into account the presence, placement, duplication and inflectional correctness of lexical terms in the output sentence. 1 Introduction The incorporation of an inflected lexicon into Neural Machine Translation (NMT) enables sys- tem developers to adapt the translation to specific domains, and users to adjust translations of phrases generated by the translation system. Phrase-Based Statistical Machine Translation (PB-SMT; Setiawan et al., 2005) provided control over system output, e.g. by using a domain-specific lexicon. The shift from phrase tables in PB-SMT to a continuous-valued representation of text in NMT has made it more difficult to incorporate lexical constraints into the translation process. The task of integrating the lexicon and a neural translator is even more challenging for highly morphological languages, when the lexical items should be correctly inflected in the output text. We carry out experiments for translation with inflected lexical constraints. As the target language of the translation we choose Polish, whose inflection is typical of the Slavic languages. The number of declination cases is six, and the verbal groups are inflected by tense, number, and person. In terms of correct inflection of the output, translation from English to Polish seems to be a more challenging task than translation in the other direction. Unlike in some preceding experiments, we require that the lexicon may be modified after the model training has been completed. We believe that in post-editing mode users expect the translation engine to immediately mirror their adjustments to the lexicon. 2 Related Work One of the first papers that addressed the incorporation of a lexicon into an NMT system was Arthur et al. (2016). The authors noticed that NMT systems tend to produce unexpected output for low-frequency words (such as names of countries). The solution proposed there consisted Proceedings of the 18th Biennial Machine Translation Summit Virtual USA, August 16 - 20, 2021, Volume 1: MT Research Track Page 282 in designing probability lexicons and combining them with probabilities calculated by an NMT model. Let us note that the motivation for that research was the avoidance of major translation errors, rather than domain adaptation. Anderson et al. (2017) introduced the concept of a Constrained Beam Search (CBS) in the task of picture captioning. …

Content truncated to 500 words