An English-Chinese Cross-lingual Word Semantic Similarity Measure Exploring Attributes and Relations

An English-Chinese Cross-lingual Word Semantic Similarity Measure Exploring Attributes and Relations ⋆ Lin Dai and Heyan Huang Department of Computer, Beijing Institute of Technology Beijing 100081, China. {dailiu,hhy63}dailiu@bit.edu.cn Abstract. Word semantic similarity measuring is a fundamental issue to many NLP appli- cations and the globalization has made an urgent request for cross-lingual word similarity measure. This paper proposed a word semantic similarity measure which is able to work in cross-lingual scenarios. Basically, a concept can be defined by a set of attributes. The basic idea of this work is to compute the similarity between words by exploring their attributes and relations. For a given word pair, we first compute similarities between their attributes by combining distance, depth and relation information. Then word similarity are computed through a combination scheme. The algorithm is implemented based on an English-Chinese bilingual ontology HowNet. Experiments show that the proposed algorithm results in high correlation against human judgments, which encourages its broad application in cross-lingual applications. Keywords: Word semantic similarity, Cross-lingual, Natural Language Processing, Com- puting linguistics. 1 Introduction Word semantic similarity measure plays a fundamental role in many natural language processing (NLP) applications such as information retrieval (IR), machine translation (MT) and word sense disambiguation (WSD). Many research efforts have been made in the past decades to improve the effectiveness of word semantic similarity measures. With the progress of the globalization, information emerged from a variety of languages to the Internet. Under this circumstance, cross- lingual applications such as machine translation, cross-lingual information retrieval, cross-lingual text categorization and clustering etc., become more and more attractive. Consequently, cross- lingual word semantic similarity measure becomes a meaningful research topic. Although there are some statistical and hybrid measures which have good overall perfor- mance (Li et al., 2003; Xia et al., 2011; Hassan and Mihalcea, 2009), the knowledge-based measures, to us, is still an important method. Even in a hybrid measure, the knowledge-based part is still a key part. Furthermore, if a large scale ontological knowledge base is available, the knowledge-based measures can be further improved if the knowledge base is comprehensively investigated and the similarity measures are elaborately designed. In this work, we “re-examine” merely knowledge based word similarity measure and a cross- lingual semantic similarity measure is proposed. As we know, concepts are language-independent and it is natural that each concept can be defined by several aspects, i.e., attributes. For example, the concept of the animal dog can be roughly defined by a kind-of attribute with value livestock and a modifier attribute with value domesticated. If all the words in this world are defined in this way, it is possible to compute the semantic similarity of any pair of words in spite of their ⋆ The authors would like to thank Prof. Z.D. Dong for his valuable advices. Copyright 2011 by Lin Dai and Heyan Huang 25th Pacific Asia Conference on Language, Information and Computation, pages 467–476 467 language. This is the basic idea of this work. In this work, this idea is implemented by …

Content truncated to 500 words