Clustering Clauses for High-Level Relation Detection: An Information-theoretic Approach

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 448–455, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Clustering Clauses for High-Level Relation Detection: An Information-theoretic Approach Samuel Brody School of Informatics University of Edinburgh s.brody@sms.ed.ac.uk Abstract Recently, there has been a rise of in- terest in unsupervised detection of high- level semantic relations involving com- plex units, such as phrases and whole sentences. Typically such approaches are faced with two main obstacles: data sparseness and correctly generalizing from the examples. In this work, we describe the Clustered Clause represen- tation, which utilizes information-based clustering and inter-sentence dependen- cies to create a simplified and generalized representation of the grammatical clause. We implement an algorithm which uses this representation to detect a predefined set of high-level relations, and demon- strate our model’s effectiveness in over- coming both the problems mentioned. 1 Introduction The semantic relationship between words, and the extraction of meaning from syntactic data has been one of the main points of research in the field of computational linguistics (see Sec- tion 5 and references therein). Until recently, the focus has remained largely either at the sin- gle word or sentence level (for instance: depen- dency extraction, word-to-word semantic simi- larity from syntax, etc.) or on relations between units at a very high context level such as the entire paragraph or document (e.g. categorizing documents by topic). Recently there have been several attempts to define frameworks for detecting and studying in- teractions at an intermediate context level, and involving whole clauses or sentences. Dagan et al. (2005) have emphasized the importance of detecting textual-entailment/implication be- tween two sentences, and its place as a key com- ponent in many real-world applications, such as Information Retrieval and Question Answering. When designing such a framework, one is faced with several obstacles. As we approach higher levels of complexity, the problem of defin- ing the basic units we study (e.g. words, sen- tences etc.) and the increasing amount of in- teractions make the task very difficult. In addi- tion, the data sparseness problem becomes more acute as the data units become more complex and have an increasing number of possible val- ues, despite the fact that many of these values have similar or identical meaning. In this paper we demonstrate an approach to solving the complexity and data sparse- ness problems in the task of detecting rela- tions between sentences or clauses. We present the Clustered Clause structure, which utilizes information-based clustering and dependencies within the sentence to create a simplified and generalized representation of the grammatical clause and is designed to overcome both these problems. The clustering method we employ is an inte- gral part of the model. We evaluate our clusters against semantic similarity measures defined on the human-annotated WordNet structure (Fell- baum, 1998). The results of these comparisons show that our cluster members are very similar semantically. We also define a high-level rela- tion detection task involving relations between clauses, evaluate our results, and demonstrate …

Content truncated to 500 words