Relational Learning of Pattern-Match Rules for Information Extraction

Relational Learning of Pattern-Match Rules for Information Extraction Mary Elaine Califf and Raymond J. Mooney Department of Computer Sciences University of Texas at Austin Austin, TX 78712 {mecaliff ,mooney}@cs. utexas, edu Abstract Information extraction systems process natural language documents and locate a specific set of relevant items. Given the recent success of empirical or corpus- based approaches in other areas of natu- ral language processing, machine learning has the potential to significantly aid the development of these knowledge-intensive systems. This paper presents a system, RAPmrt, that takes pairs of documents and filled templates and induces pattern-match rules that directly extract fillers for the slots in the template. The learning al- gorithm incorporates techniques from sev- eral inductive logic programming systems and learns unbounded patterns that in- clude constraints on the words and part- of-speech tags surrounding the filler. En- couraging results are presented on learn- ing to extract information from com- puter job postings from the newsgroup misc. jobs. offered. 1 Introduction An increasing amount of information is available in the form of electronic documents. The need to in- telligently process such texts makes information ex- traction (IE), the task of locating specific pieces of data from a natural language document, a particu- larly useful sub-area of natural language processing (NLP). In recognition of their significance, IE sys- tems have been the focus of DARPA's MUC program (Lehnert and Sundheim, 1991). Unfortunately, IE systems are difficult and time-consuming to build and the resulting systems generally contain highly domain-specific components, making them difficult to port to new domains. Recently, several researchers have begun to ap- ply learning methods to the construction of IE sys- tems (McCarthy and Lehnert, 1995, Soderland et ah, 1995, Soderland et al., 1996, Riloff, 1993, Riloff, 1996, Kim and Moldovan, 1995, Huffman, 1996). Several symbolic and statistical methods have been employed, but learning is generally used to construct only part of a larger IE system. Our system, RAPIER (Robust Automated Production of Information Ex- traction Rules), learns rules for the complete IE task. The resulting rules extract the desired items directly from documents without prior parsing or subsequent processing. Using only a corpus of doc- uments paired with filled templates, RAPIER learns unbounded Eliza-like patterns (Weizenbaum, 1966) that utilize limited syntactic information, such as the output of a part-of-speech tagger. Induced pat- terns can also easily incorporate semantic class infor- mation, such as that provided by WordNet (Miller et al., 1993). The learning algorithm was inspired by several Inductive Logic Programming (ILP) sys- tems and primarily consists of a specific-to-general (bottom-up) search for patterns that characterize slot-fillers and their surrounding context. The remainder of the paper is organized as follows. Section 2 presents background material on IE and re- lational learning. Section 3 describes RAPIEK's rule representation and learning algorithm. Section 4 presents and analyzes results obtained on extracting information from messages posted to the newsgroup mist.jobs.offered. Section 5 discusses related work in applying learning to IE, Section 6 suggests areas for future research, and Section 7 …

Content truncated to 500 words