Machine Translation Evaluation beyond the Sentence Level

Contextual Handling in Neural Machine Translation: Look Behind, Ahead and on Both Sides Ruchit Agrawal(1,2), Marco Turchi(1), Matteo Negri(1) (1)Fondazione Bruno Kessler, Italy (2)University of Trento, Italy {ragrawal, turchi, negri}@fbk.eu Abstract A salient feature of Neural Machine Trans- lation (NMT) is the end-to-end nature of training employed, eschewing the need of separate components to model different linguistic phenomena. Rather, an NMT model learns to translate individual sen- tences from the labeled data itself. How- ever, traditional NMT methods trained on large parallel corpora with a one-to-one sentence mapping make an implicit as- sumption of sentence independence. This makes it challenging for current NMT sys- tems to model inter-sentential discourse phenomena. While recent research in this direction mainly leverages a single previous source sentence to model dis- course, this paper proposes the incorpora- tion of a context window spanning previ- ous as well as next sentences as source- side context and previously generated out- put as target-side context, using an effec- tive non-recurrent architecture based on self-attention. Experiments show improve- ment over non-contextual models as well as contextual methods using only previous context. 1 Introduction Neural Machine Translation (Kalchbrenner and Blunsom, 2013; Sutskever et al., 2014; Bahdanau et al., 2014; Cho et al., 2014) has consistently out- performed other MT paradigms across a range of domains, applications and training settings (Ben- tivogli et al., 2016; Castilho et al., 2017; Toral c⃝ 2018 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. and S´anchez-Cartagena, 2017), thereby emerging as the de facto standard in Machine Translation. NMT models are typically trained at the sentence level (Cho et al., 2014), whereby the probability of an output sentence given an input sentence is max- imized, implicitly making an assumption of sen- tence independence across the dataset. This works well for the translation of stand-alone sentences or datasets containing shuffled sentences, which are not connected with each other in terms of discur- sive dependencies. However, in real life situations, written text generally follows a sequential order featuring a number of cross-sentential phenomena. Additionally, speech-like texts (Bawden, 2017) exhibit the trait of contextual dependency and se- quentiality as well, often containing a greater num- ber of references that require a common knowl- edge ground and discourse understanding for cor- rect interpretation. Figure 1 shows an example of such inter-sentential dependencies. These de- pendencies are not fully leveraged by the majority of contemporary NMT models, owing to the treat- ment of sentences as independent units for transla- tion. In order to perform well on sequential texts, NMT models need access to extra information, which could serve as the disambiguating context for better translation. Recent work in this direc- tion (Zoph and Knight, 2016; Jean et al., 2017; Tiedemann and Scherrer, 2017; Bawden et al., 2017; Wang et al., 2017) has primarily focused on previous source-side context for disambigua- tion. Since all of these approaches utilize recurrent architectures, adding context comprising of more than a single previous sentence can be …

Content truncated to 500 words