An Approach of Hybrid Hierarchical Structure for Word Similarity Computing by HowNet

International Joint Conference on Natural Language Processing, pages 927–931, Nagoya, Japan, 14-18 October 2013. An Approach of Hybrid Hierarchical Structure for Word Similarity Computing by HowNet Jiangming Liu Beijing Jiaotong Universtity jmliunlp@gmail.com Jinan Xu Beijing Jiaotong Universtity jaxu@bjtu.edu.cn Yujie Zhang Beijing Jiaotong Universtity yjzhang@bjtu.edu.cn Abstract Word similarity computing is an important and fundamental task in the field of natural lan- guage processing. Most of word similarity methods perform well in synonyms, but not well between words whose similarity is vague. To overcome this problem, this paper proposes an approach of hybrid hierarchical structure computing Chinese word similarity to achieve fine-grained similarity results with HowNet 2008. The experimental results prove that the method has a better effect on computing simi- larity of synonyms and antonyms including nouns, verbs and adjectives. Besides, it per- forms stably on standard data provided by SemEval 2012. 1 Introduction Word similarity computing plays an important role in various fields, such as Natural Language Understanding and Cognitive Science (Bunescu and Huang, 2010b; Mohler et al., 2011; Wang and Wan, 2011;). Moreover, it is a pivotal meth- od in Word Sense Disambiguation (WSD). Two main types of word similarity computing methods have been proposed. One is usually based on the thesaurus. The methods of this type utilize the structure of thesaurus (Liu and Li, 2002; Ge et al., 2010) with the advantages of preciseness and deep usage of word semantics, but a relatively complete semantic dictionary is required in order to ensure the presence of words in thesaurus. The other methods are based on large-scale corpora with some inevitable disad- vantages, such as the frequent need of large-scale corpora, noise, low search efficiency etc. (Nakov and Hearst, 2008). Therefore, it is fine to create a refined thesaurus with Internet resource or large- scale corpora (Morita et al., 2011; Navigli and Ponzetto, 2010; Davidov and Rappoport, 2010) as an interim for computing word similarity. WordNet is deemed to be very valuable the- saurus. Since Chinese that belongs to isolated language is different from English that belongs to inflected language and the complex Chinese grammar is highly ambiguous, computing Chi- nese words similarity is more difficult than Eng- lish under the same lack of systematic resource. HowNet is also a valuable bilingual knowledge thesaurus organized by Zhongdong Dong. HowNet uses a markup language called KDML to describe word’s concept which facili- tates computer processing (Li et al., 2012). A different semantic of one word has a different DEF description. DEF is defined by a number of sememes and the descriptions of semantic rela- tions between words. It is worth to mention that sememes are the most basic and the smallest units which cannot be easily divided (Liu and Li, 2002), and they are extracted from about six thousands of Chinese characters (Dong and Dong, 2006). An example of one DEF of saleslady can be described as a tree-like structure (Figure 1). The details of description in HowNet can be ac- cessed in the paper (Dong, 2002). In closely related works, …

Content truncated to 500 words