|
Two Machine Tractable Dictionaries for Information Retrieval
Peter Wallis
Royal Melbourne Institute of Technology
When: Friday, 17th June 1994
Time: 4:00pm
Where: Microsoft Institute, Conference Room 1 (next to Bistro)
Abstract:
Information retrieval (IR) systems use computers to find useful texts or documents given a description of a user's interests. IR systems would work better, we believe, if they found texts which _partially_paraphrase_ the users description. A system we are developing matches paraphrases by converting texts to representations constructed from a limited set of semantic features and then matching representations. The applied nature of the IR problem dictates that the set of features, and the mechanism for converting texts to representations, must be extensive enough to deal with ``live'' text and not just a few tailored examples. An online version of the Longman Dictionary of Contemporary English (LDOCE) is used to overcome the information bottle-neck and provide a mapping from words in texts to a rough set of semantic features.
As is often noted, LDOCE is not particularly systematic. In this talk two methods are presented for improving the suitability of definitions for the IR task, and some preliminary results are given. The first method is that described by Wilks et al (Machine Translation 5, 1990), the second is that proposed by Wallis (PACLING, Vancouver, 1993). Ideally these methods would be compared by incorporating the resulting machine _tractable_ dictionary into an IR system and comparing retrieval performance. However, recent studies indicate there are problems with the usual methods of assessment, and that proper evaluation is expensive. An alternative is presented which we believe is a reasonable test of the information content of the new dictionaries.
SPEAKER:
Peter Wallis is a PhD candidate at the Royal Melbourne Institute of Technology. His research interests are in information retrieval and machine translation.
Enquiries: Maria Milosavljevic 9850 6345 mariam@mpce.mq.edu.au
| Last modified: July 1997 |