Using Information Extraction to Build a Directory of Conference Announcements

We describe an application of information extraction for building a directory of announcements of scientific conferences. We employ a cascaded finite-state transducer to identify possible conference names, titles, dates, locations and URLs in a conference announcement. In order to cope with agrammatical text that is typical for conference announcements, our system uses orthographic features of the text and a domain-specific tag set, rather than general purpose part-of-speech tags. Extraction accuracy is improved by recognizing other entities in the text that are not extracted but could be confused with slot values. A scoring scheme based on some simple heuristics is used to select among multiple extraction candidates. We also present an evaluation of our system.