Translation Acquisition Using Synonym Sets

Proceedings of NAACL-HLT 2013, pages 655–660, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Translation Acquisition Using Synonym Sets Daniel Andrade Masaaki Tsuchida Takashi Onishi Kai Ishikawa Knowledge Discovery Research Laboratories, NEC Corporation, Nara, Japan {s-andrade@cj, m-tsuchida@cq, t-onishi@bq, k-ishikawa@dq}.jp.nec.com Abstract We propose a new method for translation ac- quisition which uses a set of synonyms to ac- quire translations from comparable corpora. The motivation is that, given a certain query term, it is often possible for a user to specify one or more synonyms. Using the resulting set of query terms has the advantage that we can overcome the problem that a single query term’s context vector does not always reliably represent a terms meaning due to the context vector’s sparsity. Our proposed method uses a weighted average of the synonyms’ context vectors, that is derived by inferring the mean vector of the von Mises-Fisher distribution. We evaluate our method, using the synsets from the cross-lingually aligned Japanese and English WordNet. The experiments show that our proposed method significantly improves translation accuracy when compared to a pre- vious method for smoothing context vectors. 1 Introduction Automatic translation acquisition is an important task for various applications. For example, finding term translations can be used to automatically up- date existing bilingual dictionaries, which are an in- dispensable resource for tasks such as cross-lingual information retrieval and text mining. Various previous research like (Rapp, 1999; Fung, 1998) has shown that it is possible to acquire word translations from comparable corpora. We suggest here an extension of this approach which uses several query terms instead of a single query term. A user who searches a translation for a query term that is not listed in an existing bilin- gual dictionary, might first try to find a synonym of that term. For example, the user might look up a synonym in a thesaurus1 or might use methods for automatic synonym acquisition like described in (Grefenstette, 1994). If the synonym is listed in the bilingual dictionary, we can consider the syn- onym’s translations as the translations of the query term. Otherwise, if the synonym is not listed in the dictionary either, we use the synonym together with the original query term to find a translation. We claim that using a set of synonymous query terms to find a translation is better than using a single query term. The reason is that a single query term’s context vector is, in general, unreliable due to spar- sity. For example, a low frequent query term tends to have many zero entries in its context vector. To mit- igate this problem it has been proposed to smooth a query’s context vector by its nearest neighbors (Pekar et al., 2006). However, nearest neighbors, which context vectors are close the query’s context vector, can have different meanings and therefore might introduce noise. The contributions of this paper are two-fold. First, we confirm experimentally that smoothing a query’s context vector with its synonyms leads in deed to higher translation accuracy, compared …

Content truncated to 500 words