Proper Name Classification in an Information Extraction Toolset

l l I II / / / / / / / / l Proper Name Classification in an Information Extraction Toolset Peter Wallis, Edmund Yuen, and Greg Chase Information Technology Division, DSTO, Box 1500 Salisbury, Sth Aust. 5108 {Peter.Wallis,Edmund.Yuen,Greg.Chase}@dsto.defence.gov.au Abstract Applied discourse analysis is a hot topic in In- formation Retrieval (IR) and the related field of Information Extraction (IE). Although in- teresting observations about discourse can be made "by hand," applications require large quantities of data about language -- data which is rather uninteresting. This paper in- vestigates using statistical analysis over a body of text to suggest new rules for recognizing named entities. 1 Introduction Understanding human languages on any sort of scale is a knowledge intensive task. This paper de- scribes a corpus based approach to gathering lan- guage data in the shallow parts of the NLP pond. Information retrieval is a popular application for re- searchers interested in applied NLP, but the problem of improving retrieval effectiveness appears to be in- tractable (Smeaton, 1992; Wallis, 1995). One helpful technique is tagging the proper names in text. Tag- ging and classifying (e.g. Is "Washington" a place or a person?) the named entities and co-references to them (she, he, the company) in text is also a pri- mary concern in systems for information extraction (DARPA, 1995). Information extraction (IE) is a well defined task; the aim being to extract data from free text, and put it in a more structured format. The IE task is not only well defined, it has application and is hence often seen as a prime example of language engineer- ing, where the aim is to explicitly solve a problem rather than to understand the nature of language. IE systems have typically only been successful in narrow domains with significant effort required to move and existing information extraction system to a new problem domain. One approach is to use tools in a development environment that assists the lan- guage engineer to create a new information extrac- tion system from pre-exisiting components. The DSTO Fact Extractor Workbench pro- vides the tools to create re-usable text skimming components, called fact extractors, that perform IE on a (very) limited domain. These components can be used directly to find things like dates and the names of companies including co-references, or they can be assembled to create larger fact extractors that skim text for more abstract entities such as company mergers. The workbench provides different views of the do- main text to assist in the development process. As an example, the language engineer might be inter- ested in seeing how the word "bought" is used in the domain o.f interest. A "grep"-like tool allows him or her to view all and only those sentences con- taining "bought". Naturally more complex patterns are possible incorporating previously developed fact extractors in the pattern. This paper discusses an extension to the corpus viewing tool set that assists the language engineer to find words, called selector terms, that may aid in the classification …

Content truncated to 500 words