Corpus-based Semantic Class Mining: Distributional vs. Pattern-Based Approaches

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 993–1001, Beijing, August 2010 Corpus-based Semantic Class Mining: Distributional vs. Pattern-Based Approaches Shuming Shi1 Huibin Zhang2* Xiaojie Yuan2 Ji-Rong Wen1 1 Microsoft Research Asia 2 Nankai University {shumings, jrwen}@microsoft.com zhanghuibin@126.com; yuanxj@nankai.edu.cn Abstract Main approaches to corpus-based seman- tic class mining include distributional similarity (DS) and pattern-based (PB). In this paper, we perform an empirical comparison of them, based on a publicly available dataset containing 500 million web pages, using various categories of queries. We further propose a frequency- based rule to select appropriate approach- es for different types of terms. 1 Introduction1 Computing the semantic relationship between terms, which has wide applications in natural language processing and web search, has been a hot topic nowadays. This paper focuses on cor- pus-based semantic class mining (Lin 1998; Pan- tel and Lin 2002; Pasca 2004; Shinzato and Torisawa, 2005; Ohshima, et al., 2006; Zhang et al., 2009), where peer terms (or coordinate terms) are discovered from a corpus. Existing approaches to semantic class mining could roughly be divided into two categories: distributional similarity (DS), and pattern-based (PB). The first type of work (Hindle, 1990; Lin 1998; Pantel and Lin 2002) is based on the distri- butional hypothesis (Harris, 1985), saying that terms occurring in analogous (lexical or syntactic) contexts tend to be similar. DS approaches basi- cally exploit second-order co-occurrences to dis- cover strongly associated concepts. In pattern- based approaches (Hearst 1992; Pasca 2004; Shinzato and Torisawa, 2005; Ohshima, et al., 2006; Zhang et al., 2009), patterns are applied to * Work done during an internship at Microsoft discover specific relationships between terms, from the general first-order co-occurrences. For example, “NP such as NP, NP…, and NP” is a popular and high-quality pattern for extracting peer terms (and also hyponyms). Besides the nat- ural language patterns, some HTML tag tree pat- terns (e.g., the drop down list) are also effective in semantic class mining. It is worth-noting that the word “pattern” also appears in some DS approaches (Pasca et al., 2006; Tanev and Magnini, 2006; Pennacchiotti and Pantel, 2009), to represent the context of a term or a term-pair, e.g., “(invent, subject-of)” for the term “Edison”, and “- starring -” for the term-pair “(The Terminal, Tom Hanks)”. Alt- hough “patterns” are utilized, we categorize them as DS approaches rather than PB, because they match the DS framework well. In this paper, PB only refers to the approaches that utilize patterns to exploit first-order co-occurrences. And the patterns in DS approaches are called contexts in the following part of this paper. Progress has been made and promising results have been reported in the past years for both DS and PB approaches. However, most previous re- search work (some exceptions are discussed in related work) involves solely one category of ap- proach. And there is little work studying the comparison of their performance for different types of terms (we use “term” to represent a sin- gle word or a phrase). In this …

Content truncated to 500 words