Chinese Hypernym-Hyponym Extraction from User Generated Categories

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1350–1361, Osaka, Japan, December 11-17 2016. Chinese Hypernym-Hyponym Extraction from User Generated Categories Chengyu Wang, Xiaofeng He⇤ School of Computer Science and Software Engineering East China Normal University, Shanghai, China chywang2013@gmail.com, xfhe@sei.ecnu.edu.cn Abstract Hypernym-hyponym (“is-a”) relations are key components in taxonomies, object hierarchies and knowledge graphs. While there is abundant research on is-a relation extraction in English, it still remains a challenge to identify such relations from Chinese knowledge sources accurately due to the flexibility of language expression. In this paper, we introduce a weakly supervised framework to extract Chinese is-a relations from user generated categories. It employs piece- wise linear projection models trained on a Chinese taxonomy and an iterative learning algorithm to update models incrementally. A pattern-based relation selection method is proposed to pre- vent “semantic drift” in the learning process using bi-criteria optimization. Experimental results illustrate that the proposed approach outperforms state-of-the-art methods. 1 Introduction A hypernym-hyponym (“is-a”) relation is a word/phrase pair (x, y) such that x is a hyponym of y. These relations are extensively employed in machine reading (Etzioni et al., 2011), query understanding (Hua et al., 2015) and other NLP tasks. The extraction of is-a relations is necessary to construct taxonomies for Web-scale knowledge graphs (Suchanek et al., 2007; Wu et al., 2012; Wang et al., 2015). In previous work, is-a relations were obtained by either using expert-compiled thesauri such as Word- Net (Miller, 1995), or harvested automatically from the Web. Since knowledge in thesauri is usually limited in quantity and variety, it is more prevalent to harvest is-a relations from online encyclopedias (Ponzetto and Strube, 2007), Web corpora (Wu et al., 2012), etc. Currently, a majority of existing meth- ods focus on syntactic, lexical and/or semantic analysis on English corpora, but most of these approaches are language dependent. It is not easy to apply methods for one language to knowledge sources in an- other language directly. For example, in Chinese, the word formation, grammar, semantics and tenses are flexible and more irregular. Thus patttern-based methods can only cover few linguistic circumstances. As pointed out by Li et al. (2013), the performance of syntactic analysis and named entity recognition on Chinese corpora still needs to be improved to support robust relation extraction. Furthermore, it is still difficult to use machine translation-based methods to extract such relations because there are great differences in word orders between English and Chinese (Cai et al., 2014). More recently, word embedding (or distributed word representation) has been empirically proved ef- fective in modeling some of the semantic relations between words by offsets of word vectors (Mikolov et al., 2013a; Mikolov et al., 2013b). The learning of word embeddings only requires shallow processing of a large text corpus. As Fu et al. (2014) suggest, the representation of is-a relations is more compli- cated than vector offsets. By studying the relations of word embeddings between hyponyms and their respective hypernyms, is-a relations can be identified by …

Content truncated to 500 words