Improving Semi-supervised Acquisition of Relation Extraction Patterns

Proceedings of the Workshop on Information Extraction Beyond The Document, pages 29–35, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Improving Semi-Supervised Acquisition of Relation Extraction Patterns Mark A. Greenwood and Mark Stevenson Department of Computer Science University of Sheffield Sheffield, S1 4DP, UK {m.greenwood,marks}@dcs.shef.ac.uk Abstract This paper presents a novel approach to the semi-supervised learning of Informa- tion Extraction patterns. The method makes use of more complex patterns than previous approaches and determines their similarity using a measure inspired by re- cent work using kernel methods (Culotta and Sorensen, 2004). Experiments show that the proposed similarity measure out- performs a previously reported measure based on cosine similarity when used to perform binary relation extraction. 1 Introduction A recent approach to Information Extraction (IE) is to make use of machine learning algorithms which allow systems to be rapidly developed or adapted to new extraction problems. This reduces the need for manual development which is a major bottleneck in the development of IE technologies and can be extremely time consuming (e.g. Riloff (1996)). A number of machine learning approaches have recently been applied. One is the use of itera- tive learning algorithms to infer extraction patterns from a small number of seed examples (Yangar- ber et al., 2000; Stevenson and Greenwood, 2005). These approaches use dependency analysis as the basis of IE patterns. Training text is parsed and a set of candidate patterns extracted. These patterns are then compared against the seeds with the most similar being selected and added to the seed set. (Optionally, a user may verify the patterns at this point.) The process is then repeated with the re- maining patterns being compared to the enlarged seed set. The process continues until a suitable set of patterns has been learned. These approaches require only a small number of example extraction patterns which greatly reduces the effort required to develop IE systems. While it has been found that these approaches are capable of learning useful IE patterns from a handful of examples (Yangarber et al., 2000; Stevenson and Greenwood, 2005) they are limited by the use of basic extraction patterns: SVO tu- ples. The patterns used by these systems are de- fined as a verb and its direct subject and/or object. They could then only extract a limited set of re- lations; those expressed using a verb and its di- rect arguments. For example, these patterns could identify the relation between Jones and Smith in the sentence “Jones replaced Smith”. However, no pattern consisting of a verb and its arguments could be constructed which could identify the same relation in “Jones was named as Smith’s suc- cessor.” Others have suggested alternative approaches for generating extraction patterns from depen- dency trees, each of which allows a particular part of the dependency analysis to act as an extraction pattern. For example, Sudo et al. (2003) used pat- terns consisting of a path from a verb to any of its descendents (direct or indirect) while Bunescu and Mooney (2005) suggest the shortest …

Content truncated to 500 words