Co-STAR: A Co-training Style Algorithm for Hyponymy Relation Acquisition from Structured and Unstructured Text

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 842–850, Beijing, August 2010 Co-STAR: A Co-training Style Algorithm for Hyponymy Relation Acquisition from Structured and Unstructured Text Jong-Hoon Oh, Ichiro Yamada, Kentaro Torisawa, and Stijn De Saeger Language Infrastructure Group, MASTAR Project, National Institute of Information and Communications Technology (NICT) {rovellia,iyamada,torisawa,stijn}@nict.go.jp Abstract This paper proposes a co-training style algorithm called Co-STAR that acquires hyponymy relations simultaneously from structured and unstructured text. In Co- STAR, two independent processes for hy- ponymy relation acquisition – one han- dling structured text and the other han- dling unstructured text – collaborate by re- peatedly exchanging the knowledge they acquired about hyponymy relations. Un- like conventional co-training, the two pro- cesses in Co-STAR are applied to dif- ferent source texts and training data. We show the effectiveness of this al- gorithm through experiments on large- scale hyponymy-relation acquisition from Japanese Wikipedia and Web texts. We also show that Co-STAR is robust against noisy training data. 1 Introduction Acquiring semantic knowledge, especially se- mantic relations between lexical terms, is re- garded as a crucial step in developing high-level natural language applications. This paper pro- poses Co-STAR (a Co-training STyle Algorithm for hyponymy Relation acquisition from struc- tured and unstructured text). Similar to co- training (Blum and Mitchell, 1998), two hy- ponymy relation extractors in Co-STAR, one for structured and the other for unstructured text, it- eratively collaborate to boost each other’s perfor- mance. Many algorithms have been developed to auto- matically acquire semantic relations from struc- tured and unstructured text. Because term pairs are encoded in structured and unstructured text in different styles, different kinds of evidence have been used for semantic relation acquisition: Evidence from unstructured text: lexico- syntactic patterns and distributional similar- ity (Ando et al., 2004; Hearst, 1992; Pantel et al., 2009; Snow et al., 2006; De Saeger et al., 2009; Van Durme and Pasca, 2008); Evidence from structured text: topic hierarchy, layout structure of documents, and HTML tags (Oh et al., 2009; Ravi and Pasca, 2008; Sumida and Torisawa, 2008; Shinzato and Torisawa, 2004). Recently, researchers have used both structured and unstructured text for semantic-relation acqui- sition, with the aim of exploiting such different kinds of evidence at the same time. They ei- ther tried to improve semantic relation acquisition by putting the different evidence together into a single classifier (Pennacchiotti and Pantel, 2009) or to improve the coverage of semantic relations by combining and ranking the semantic relations obtained from two source texts (Talukdar et al., 2008). In this paper we propose an algorithm called Co-STAR. The main contributions of this work can be summarized as follows. • Co-STAR is a semi-supervised learning method composed of two parallel and iter- ative processes over structured and unstruc- tured text. It was inspired by bilingual co- training, which is a framework for hyponymy relation acquisition from source texts in two languages (Oh et al., 2009). Like bilingual co-training, two processes in Co-STAR op- erate independently on structured text and unstructured …

Content truncated to 500 words