Research on concept-sememe tree and semantic relevance computation

Research on concept-sememe tree and semantic relevance computation GuiPing Zhang 1, Chao Yu1, DongFeng Cai1, Yan Song1, JingGuang Sun1 1 Natural Language Processing Laboratory, Shenyang Institute of Aeronautical Engineering P.O.box 118, No.52 North Huanghe Street, Shenyang, Liaoning, China, 110034. Email: zgp@ge-soft.com, yc089067@sina.com, cdf@ge-soft.com, mattsure@gmail.com , sunjingguang@gmail.com Abstract. In this paper, we parse the hierarchy relation of concepts in HowNet, with the parsing we design a concept-sememe tree structure, which can make it easy to understand the relations between the sememes in concepts. The tree structure can easily describe the relationship between sememes in concept and make it more convenient to process by computer. The steps of building the tree are also presented in this paper. We then discuss the relevance computation based on HowNet. The preliminary experiment shows the relevance computation method can achieve satisfying results. Keywords: HowNet; relevance; concept-sememe tree. 1 Introduction The computation of similarity and relevance between words has wide application in the fields of machine translation, sense disambiguation, IR etc. In some cases, similarity and relevance may be confused. Similarity between words refers to the feature of clustering while relevance refers to the extent of association [1]. The two words with similar semantic relationship may have similar associated words such as the words “医生(doctor)” and “护士(nurse)”, both have the same associated words as “医 院(hospital), 病人(patient), 打针(have an injection), 吃药(drug)” etc. However, associated words usually have no similarity such as the word pair “吃 (eat)” and “食物(food)” etc. Researches on computation of similarity and relevance are widely spread in the world. Recently, there are two semantic similarity computation methods based on HowNet: one is described by Professor Liu Qun in [2] and the other is described by professor Dong ZhenDong in [3]. In this paper, we describe how to build concept-sememe tree on the basis of decomposing word concept definition (DEF). The concept-sememe tree illustrates DEFs of the words and makes them easy to process by computer. 2 Concept similarity computation and the implementation of the concept- sememe tree Concept is a kind of description of word sense. Polysemous words have several different concepts. Because polysemous words contain several concepts, we must ascertain the DEFs of the polysemous words before we compute the semantic similarity of the word pairs. Thus the computation of the semantic similarity is, in fact, to compute the similarity of the concepts. 2.1 HowNet based concept similarity computation Number of the same concept node determines the similarity of the concept of the words. The more concept nodes two concepts share, the more similar the two concepts are. Here the concept node is a pair of “semantic role=sememe”. Same concept nodes refer to the same description of sememes in character 398 form and the same structure of these concepts. In practical computation, it is not easy to determine whether two concept nodes are with the same structure. In this paper we design the concept-sememe tree which translates concept DEF into tree structure, it can make us convenient to understand the hierarchy structure of …

Content truncated to 500 words