Chinese Semantic Class Learning from Web Based on Concept-Level Characteristics

Chinese Semantic Class Learning from Web Based on Concept-Level Characteristics Wenbo Pang a, Xiaozhong Fan a, Jiangde Yu b, and Yuxiang Jia c aSchool of Computer and Technology, Beijing Institute of Technology Beijing 100081, China hananpwb@126.com, {hananpwb, fxz}@bit.com bSchool of Computer and Information Engineering, Anyang Normal University Anyang 455000, China jangder yu@tom.com cInstitute of Computational Linguistics, Peking University Beijing 100871, China yxjia@pku.edu.cn Abstract. The automatic extraction of semantic class instance is a foundational work for many natural language processing applications. One of its crucial problems is how to val- idate whether a candidate instances is a true class member. Different from the common validation approaches based on the cooccurrence between instances, we present a novel ap- proach based on concept characteristics, including category features, interference semantic classes, and collective instance. Firstly, we analyze the common error instances produced by cooccurrence-based validation from the perspective of concept, and then utilize the concept characteristics to validate the candidate instances. We conduct experiments on eight seman- tic classes and achieved high accuracies and recall rates, especially on open semantic classes, such as fish and singer. Keywords: semantic class, candidate validation, concept characteristics 1 Introduction Semantic class learning is an important and well-studied task, which takes in a semantic class name as input (e.g. fruits) and automatically outputs its instances (e.g. apple, banana, orange, etc.). Although there are some existing semantic dictionaries, such as WordNet (Miller et al., 1990), they lack the coverage to handle large open domains or rapidly changing categories: Vieira and Poesio (2000) found that, only 56% of antecedent/anaphoric coreferent pairs in hyponymy relations in the WSJ were in WordNet. So, automatic semantic class learning has been the motivating force underlying many applications in lexical acquisition, information extraction, and the construction of semantic taxonomies. Many methods have been developed for automatic semantic class learning, under the rubrics of lexical acquisition, hyponym acquisition, semantic class identification, and web-based information extraction. Almost all of these approaches face the same crucial problem: how to validate whether a extracted instances is a true class member. Currently, the validation methods mainly are based on the co-occurrence between candidates (or true instances) (Kozareva et al., 2008; Kozareva et al., 2009; Wang and Cohen, 2009). This kind of validation has three shortages: 1) once some error instances are introduced in the bootstrapping process for some unavoidable reason, they possibly bring more and more error instances, which makes they get a high score in a common re-ranking algorithm; 2) can not reject Copyright 2009 by Wenbo Pang, Xiaozhong Fan, Jiangde Yu, and Yuxiang Jia 415 23rd Pacific Asia Conference on Language, Information and Computation, pages 415–424 the error instances caused by people’s usage habits or misunderstanding. For example, because � �(European Union) often appears with ��(America) and ��(Japan) in the hyponym pattern of countries, although �� is not a country, it is also accepted by cooccurrence-based validation; and 3) miss the instances that cooccur only with some specific instances, such as ����(Neon fish), which only cooccurs with other kind …

Content truncated to 500 words