A Semantic Approach to IE Pattern Induction

Proceedings of the 43rd Annual Meeting of the ACL, pages 379–386, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics A Semantic Approach to IE Pattern Induction Mark Stevenson and Mark A. Greenwood Department of Computer Science University of Sheffield Sheffield, S1 4DP, UK marks,m.greenwood@dcs.shef.ac.uk Abstract This paper presents a novel algorithm for the acquisition of Information Extraction patterns. The approach makes the assump- tion that useful patterns will have simi- lar meanings to those already identified as relevant. Patterns are compared using a variation of the standard vector space model in which information from an on- tology is used to capture semantic sim- ilarity. Evaluation shows this algorithm performs well when compared with a previously reported document-centric ap- proach. 1 Introduction Developing systems which can be easily adapted to new domains with the minimum of human interven- tion is a major challenge in Information Extraction (IE). Early IE systems were based on knowledge en- gineering approaches but suffered from a knowledge acquisition bottleneck. For example, Lehnert et al. (1992) reported that their system required around 1,500 person-hours of expert labour to modify for a new extraction task. One approach to this problem is to use machine learning to automatically learn the domain-specific information required to port a sys- tem (Riloff, 1996). Yangarber et al. (2000) proposed an algorithm for learning extraction patterns for a small number of examples which greatly reduced the burden on the application developer and reduced the knowledge acquisition bottleneck. Weakly supervised algorithms, which bootstrap from a small number of examples, have the advan- tage of requiring only small amounts of annotated data, which is often difficult and time-consuming to produce. However, this also means that there are fewer examples of the patterns to be learned, making the learning task more challenging. Pro- viding the learning algorithm with access to addi- tional knowledge can compensate for the limited number of annotated examples. This paper presents a novel weakly supervised algorithm for IE pattern induction which makes use of the WordNet ontology (Fellbaum, 1998). Extraction patterns are potentially useful for many language processing tasks, including question an- swering and the identification of lexical relations (such as meronomy and hyponymy). In addition, IE patterns encode the different ways in which a piece of information can be expressed in text. For exam- ple, “Acme Inc. fired Jones”, “Acme Inc. let Jones go”, and “Jones was given notice by his employers, Acme Inc.” are all ways of expressing the same fact. Consequently the generation of extraction patterns is pertinent to paraphrase identification which is cen- tral to many language processing problems. We begin by describing the general process of pat- tern induction and an existing approach, based on the distribution of patterns in a corpus (Section 2). We then introduce a new algorithm which makes use of WordNet to generalise extraction patterns (Sec- tion 3) and describe an implementation (Section 4). Two evaluation regimes are described; one based on the identification of relevant documents and another which aims to identify …

Content truncated to 500 words