Word Sense-Aware Machine Translation: Including Senses as Contextual Features for Improved Translation Models

Word Sense-Aware Machine Translation: Including Senses as Contextual Features for Improved Translation Models Steven Neale1, Lu´ıs Gomes1, Eneko Agirre2, Oier Lopez de Lacalle2 and Ant´onio Branco1 1Department of Informatics, Faculty of Sciences University of Lisbon, Portugal {steven.neale, luis.gomes, antonio.branco}@di.fc.ul.pt 2IXA NLP Group University of the Basque Country, Spain {e.agirre, oier.lopezdelacalle}@ehu.eus Abstract Although it is commonly assumed that word sense disambiguation (WSD) should help to improve lexical choice and improve the quality of machine translation systems, how to successfully integrate word senses into such systems remains an unanswered question. Some successful approaches have involved reformulating either WSD or the word senses it produces, but work on using traditional word senses to improve machine translation have met with limited success. In this paper, we build upon previous work that experimented on including word senses as contextual features in maxent-based translation models. Training on a large, open-domain corpus (Europarl), we demonstrate that this aproach yields significant improvements in machine translation from English to Portuguese. Keywords: Machine translation, word senses, maxent-models 1. Introduction In natural language processing (NLP), word sense disam- biguation (WSD) refers to the process of determining the ‘sense’ or meaning of a word when used in a particular con- text (Agirre and Edmonds, 2006) – thus solving the com- mon problem of lexical ambiguity in language, where dif- ferent occurrences of a word token may have multiple dis- tinct meanings. To use a classic example, the word ‘bank’ could be interpreted in the sense of the financial institu- tion or as the slope of land at the side of a river. Some successful approaches to WSD in recent years have been ‘knowledge-based’, with classes of words stored in lexical ontologies such as WordNet (Fellbaum, 1998) where the collective meanings of open-class words (nouns, verbs, ad- jectives and adverbs) are grouped together as ‘synsets’. While it has long been assumed that an optimally suc- cessful MT system must incorporate some kind of WSD component (Carpuat and Wu, 2005), attempts to integrate WSD components into machine translation systems have met with mixed – and usually limited – success. Early at- tempts at ‘projecting’ word senses directly into a machine translation system (Carpuat and Wu, 2005) were followed by various complete reformulations of the disambiguation process (Carpuat and Wu, 2007; Xiong and Zhang, 2014) – some of which yielded small improvements in translation quality – but the question of whether pure word senses from traditional, knowledge-based WSD approaches can be use- ful for machine translation still remains. This paper provides further evidence to support our pre- vious work (Neale et al., 2015), which experimented in including the output from WSD as contextual features in maxent-based translation models in search of improved performance for machine translation from English to Por- tuguese. Training our transfer model using a much larger dataset – approximately 1.9 million English-Portuguese aligned sentences from Europarl – we find that the very small gains reported previously are now statistically signif- icant, confirming our original hypothesis that adding word senses provided by WSD …

Content truncated to 500 words