Text Patterns and Compression Models for Semantic Class Learning

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1376–1381, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Text Patterns and Compression Models for Semantic Class Learning Chung-Yao Chuang Institute of Information Science Academia Sinica, Taiwan cychuang@iis.sinica.edu.tw Yi-Hsun Lee Institute of Information Science Academia Sinica, Taiwan rog@iis.sinica.edu.tw Wen-Lian Hsu Institute of Information Science Academia Sinica, Taiwan hsu@iis.sinica.edu.tw Abstract This paper proposes a weakly-supervised approach for extracting instances of se- mantic classes. This method constructs simple wrappers automatically based on specified seed instances and uses a com- pression model to assess the contextual ev- idence of its extraction. By adopting this compression model, our approach can bet- ter avoid erroneous extractions in a noisy corpus such as the Web. The empiri- cal results show that our system performs quite consistently even when operating on a noisy text with a lot of possibly irrelevant documents. 1 Introduction Extracting instances of semantic classes is a fundamental task in natural language processing (NLP). Such a task aims to extract instances be- longing to a specific category such as acquiring Tom Hanks and Al Pacino into a list containing other actors. This kind of information serves as building blocks for various NLP tasks. For example, major search engines such as Yahoo and Google gather large amount of such classes (Pas¸ca, 2007; Chaudhuri et al., 2009) to bet- ter interpret queries and provide search sugges- tions. Other applications include ontology learn- ing (Cimiano et al., 2004), co-reference resolution (McCarthy and Lehnert, 1995) and advertisement matching (Chang et al., 2009). Most of the approaches for this task can be roughly classified into two categories: distribu- tional and pattern-based. The distributional ap- proaches use contextual similarity to model the instances of a given class. Following the distri- butional hypothesis (Harris, 1970), these methods take a small set of seed instances and generate new instances from noun phrases that are most similar to the seeds in terms of the distributions of sur- rounding words (Sarmento et al., 2007; Pantel et al., 2009). The pattern-based approaches use text patterns to extract instances of a given semantic class (Riloff and Shepherd, 1997; Riloff and Jones, 1999; Banko et al., 2007; Pas¸ca, 2007). The most representative study is the group of patterns pro- posed by Hearst (1992). For example, patterns like ‘X such as Y’ and ‘X including Y’ can be applied to extract instances from ‘actors such as Tom Hanks’ and ‘countries including Japan’. In these approaches, semantic classes are specified by providing small sets of seed instances or seed patterns such as (Kozareva et al., 2008) which utilized a single hyponym pattern combined with graph structure to extract semantic lexicons from the Web. In addition to natural language pat- terns, Wang and Cohen (2007) demonstrated an approach that learns the pattern of specific meta- structure of the document (e.g., tags in HTML) automatically from seed instances. In this paper, we propose a method similar to (Kozareva et al., 2008) and (Wang and Cohen, 2007) in …

Content truncated to 500 words