How Much Context Span is Enough? Examining Context-Related Issues for Document-level MT

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3017–3025 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3017 How Much Context Span is Enough? Examining Context-Related Issues for Document-level MT Sheila Castilho Dublin City University ADAPT Centre/School of Computing sheila.castilho@adaptcentre.ie Abstract This paper analyses how much context span is necessary to solve different context-related issues, namely, reference, ellipsis, gender, number, lexical ambiguity, and terminology when translating from English into Portuguese. We use the DELA corpus, which consists of 60 documents and six different domains (subtitles, literary, news, reviews, medical, and legislation). We find that the shortest context span to disambiguate issues can appear in different positions in the document including preceding, following, global, world knowledge; and that the average length depends on the issue types as well as the domain. Additionally, we show that the standard approach of relying on only two preceding sentences as context might not be enough depending on the domain and issue types. Keywords: document-level, context span, machine translation 1. Introduction Recently, there has been a rise in Neural Machine Translation (NMT) research, and NMT is now widely used in a variety of fields, mainly due to advancements in neural models (Sutskever et al., 2014; Bahdanau et al., 2015; Vaswani et al., 2017). Consequently, increas- ing efforts have been made in order to include discourse into these systems (Wang, 2019; Lopes et al., 2020). Despite the fact that context is extensively used in translation and interpreting literature (Baker, 2006), it still lacks a specific definition for practical purposes, such as in a professional translator’s day-to-day work (Melby and Foster, 2010). For Melby and Foster (2010, p.3), context can be categorised into non-text (non- linguistic variables) and text (linguistic aspects). In this work, we examine the context span necessary to solve six different context-related issues, namely, ref- erence, ellipsis, gender, number, lexical ambiguity, and terminology. We adopt Melby and Foster (2010)’s view of context that is important to the analysis of transla- tions, and focus (i) on the co-text, i.e. the boundaries within the document translated, and (ii) in the non-text, where the name of the authors, speakers, and products have an effect on the translation. We use the DELA cor- pus (Castilho et al., 2021), a document-level corpus an- notated with the aforementioned context-aware issues when translating from English (EN) into Brazilian- Portuguese (PT-BR). It consists of 60 documents and six different domains (subtitles, literary, news, reviews, medical, and legislation). We find that the shortest con- text span to solve the issues can appear in various posi- tions in the document including preceding, following, global, and world knowledge contexts. Moreover, we find that the average context length depends on the is- sue types as well as the domain. 2. Related Work Even though the Machine Translation (MT) commu- nity has become more interested in discourse MT, the definition of what constitutes a document-level MT is still unclear (Castilho et al., 2020). Furthermore, it is also unclear …

Content truncated to 500 words