Contextual Modeling for Meeting Translation Using Unsupervised Word Sense Disambiguation

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1227–1235, Beijing, August 2010 Contextual Modeling for Meeting Translation Using Unsupervised Word Sense Disambiguation Yang Mei Department of Electrical Engineering University of Washington yangmei@u.washington.edu Katrin Kirchhoff Department of Electrical Engineering University of Washington katrin@ee.washington.edu Abstract In this paper we investigate the challenges of applying statistical machine translation to meeting conversations, with a particu- lar view towards analyzing the importance of modeling contextual factors such as the larger discourse context and topic/domain information on translation performance. We describe the collection of a small cor- pus of parallel meeting data, the develop- ment of a statistical machine translation system in the absence of genre-matched training data, and we present a quantita- tive analysis of translation errors result- ing from the lack of contextual modeling inherent in standard statistical machine translation systems. Finally, we demon- strate how the largest source of translation errors (lack of topic/domain knowledge) can be addressed by applying document- level, unsupervised word sense disam- biguation, resulting in performance im- provements over the baseline system. 1 Introduction Although statistical machine translation (SMT) has made great progress over the last decade, most SMT research has focused on the transla- tion of structured input data, such as newswire text or parliamentary proceedings. Spoken lan- guage translation has mostly concentrated on two- person dialogues, such as travel expressions or patient-provider interactions in the medical do- main. Recently, more advanced spoken-language data has been addressed, such as speeches (St¨uker et al., 2007), lectures (Waibel and F¨ugen, 2008), and broadcast conversations (Zheng et al., 2008). Problems for machine translation in these genres include the nature of spontaneous speech input (e.g. disfluencies, incomplete sentences, etc.) and the lack of high-quality training data. Data that match the desired type of spoken-language inter- action in topic, domain, and, most importantly, in style, can only be obtained by transcribing and translating conversations, which is a costly and time-consuming process. Finally, many spoken- language interactions, especially those involving more than two speakers, rely heavily on the par- ticipants’ shared contextual knowledge about the domain and topic of the discourse, relationships between speakers, objects in the real-world en- vironment, past interactions, etc. These are typ- ically not modelled in standard SMT systems. The problem of speech disfluencies has been addressed by disfluency removal techniques that are applied prior to translation (Rao et al., 2007; Wang et al., 2010). Training data sparsity has been addressed by adding data from out-of-domain re- sources (e.g. (Matusov et al., 2004; Hildebrandt et al., 2005; Wu et al., 2008)), exploiting com- parable rather than parallel corpora (Munteanu and Marcu, 2005), or paraphrasing techniques (Callison-Burch et al., 2006). The lack of con- textual modeling, by contrast, has so far not been investigated in depth, although it is a generally recognized problem in machine translation. Early attempts at modeling contextual information in machine translation include (Mima et al., 1998), where information about the role, rank and gen- der of speakers and listeners was utilized in a …

Content truncated to 500 words