Expectations of Word Sense in Parallel Corpora

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 621–625, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Expectations of Word Sense in Parallel Corpora Xuchen Yao, Benjamin Van Durme and Chris Callison-Burch Center for Language and Speech Processing, and HLTCOE Johns Hopkins University Abstract Given a parallel corpus, if two distinct words in language A, a1 and a2, are aligned to the same word b1 in language B, then this might signal that b1 is polysemous, or it might sig- nal a1 and a2 are synonyms. Both assump- tions with successful work have been put for- ward in the literature. We investigate these assumptions, along with other questions of word sense, by looking at sampled parallel sentences containing tokens of the same type in English, asking how often they mean the same thing when they are: 1. aligned to the same foreign type; and 2. aligned to different foreign types. Results for French-English and Chinese-English parallel corpora show simi- lar behavior: Synonymy is only very weakly the more prevalent scenario, where both cases regularly occur. 1 Introduction Parallel corpora have been used for both paraphrase induction and word sense disambiguation (WSD). Usually one of the following two assumptions is made for these tasks: 1. Polysemy If two different words in language A are aligned to the same word in language B, then the word in language B is polysemous. 2. Synonymy If two different words in language A are aligned to the same word in language B, then the two words in A are synonyms, and thus is not evidence of polysemy in B. Despite the alternate nature of these assumptions, both have associated articles in which a researcher claimed success. Under the polysemy assumption, Gale et al. (1992) used French translations as En- glish sense indicators in the task of WSD. For in- stance, for the English word duty, the French transla- tion droit was taken to signal its tax sense and devoir to signal its obligation sense. These French words were used as labels for different English senses. Similarly, in a cross-lingual WSD setting,1 Lefever et al. (2011) treated each English-foreign alignment as a so-called ParaSense, using it as a proxy for hu- man labeled training data. Under the synonymy assumption, Diab and Resnik (2002) did word sense tagging by grouping together all English words that are translated into the same French word and by further enforcing that the majority sense for these English words was pro- jected as the sense for the French word. Bannard and Callison-Burch (2005) applied the idea that French phrases aligned to the same English phrase are para- phrases in a system that induces paraphrases by piv- oting through aligned foreign phrases. Based on this, and other successful prior work, it seems neither of the assumptions must hold universally. Therefore we investigate how often we might expect one or the other to dominate: we sample polysemous words from wide-domain {French,Chinese}-English corpora, and …

Content truncated to 500 words