Cross-Lingual Semantic Similarity of Words as the Similarity of Their Semantic Word Responses

Proceedings of NAACL-HLT 2013, pages 106–116, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Cross-Lingual Semantic Similarity of Words as the Similarity of Their Semantic Word Responses Ivan Vuli´c and Marie-Francine Moens Department of Computer Science KU Leuven Celestijnenlaan 200A Leuven, Belgium {ivan.vulic,marie-francine.moens}@cs.kuleuven.be Abstract We propose a new approach to identifying semantically similar words across languages. The approach is based on an idea that two words in different languages are similar if they are likely to generate similar words (which in- cludes both source and target language words) as their top semantic word responses. Se- mantic word responding is a concept from cognitive science which addresses detecting most likely words that humans output as free word associations given some cue word. The method consists of two main steps: (1) it uti- lizes a probabilistic multilingual topic model trained on comparable data to learn and quan- tify the semantic word responses, (2) it pro- vides ranked lists of similar words accord- ing to the similarity of their semantic word response vectors. We evaluate our approach in the task of bilingual lexicon extraction (BLE) for a variety of language pairs. We show that in the cross-lingual settings without any language pair dependent knowledge the response-based method of similarity is more robust and outperforms current state-of-the art methods that directly operate in the semantic space of latent cross-lingual concepts/topics. 1 Introduction Cross-lingual semantic word similarity addresses the task of detecting words that refer to similar se- mantic concepts and convey similar meanings across languages. It ultimately boils down to the automatic identification of translation pairs, that is, bilingual lexicon extraction (BLE). Such lexicons and seman- tically similar words serve as important resources in cross-lingual knowledge induction (e.g., Zhao et al. (2009)), statistical machine translation (Och and Ney, 2003) and cross-lingual information retrieval (Ballesteros and Croft, 1997; Levow et al., 2005). From parallel corpora, semantically similar words and bilingual lexicons are induced on the basis of word alignment models (Brown et al., 1993; Och and Ney, 2003). However, due to a relative scarce- ness of parallel texts for many language pairs and domains, there has been a recent growing interest in mining semantically similar words across languages on the basis of comparable data readily available on the Web (e.g., Wikipedia, news stories) (Haghighi et al., 2008; Hassan and Mihalcea, 2009; Vuli´c et al., 2011; Prochasson and Fung, 2011). Approaches to detecting semantic word similarity from comparable corpora are most commonly based on an idea known as the distributional hypothesis (Harris, 1954), which states that words with sim- ilar meanings are likely to appear in similar con- texts. Each word is typically represented by a high- dimensional vector in a feature vector space or a so- called semantic space, where the dimensions of the vector are its context features. The semantic similar- ity of two words, wS 1 given in the source language LS with vocabulary V S and wT 2 in the target lan- guage LT with vocabulary V T is then: …

Content truncated to 500 words