SP500207 TIPSTER Panel -- DR LINK's Linguistic-Conceptual Approach to Document Detection chapter E. Liddy S. Myaeng National Institute of Standards and Technology Donna K. Harman Conceptual Graph Matcher Testing As the first prototype, the CG Matcher was implemented with a set of scoring heuristics whose theoretical basis is on the Dempster-Shafer theory of evidence. The score for a document is computed progressively from node-level evidence through multiple stages to generate a score for text units of different granularities. In order to test the feasibility of the prototype module, we have run it with manually generated OGs for twenty documents and five topic statements for which we have relevance judgments. While it is premature to draw any conclusions on the efficacy of the matching algorithm and the heuristics, mainly due to the size of the data set and the stage of the development, the results were encouraging and have provided us with much insight on how the scoring heuristics need to be tuned. Conclusions This paper on the DR-LINK system should be considered a report on a work in progress, since we did not have a fully devebped system at the time of the TREC testing. However, we do believe that the three system components which were tested perform quite respectably, given their innovativeness. Continued development and feedback from the TREC results will provide much more refined versions of these system modules. In addition, two system modules remain to be developed and the full system, which is quite synergistic in its approach to achieving its goals, remains to be integrated and tested as a full system. The Relation Concept Detector and Conceptual Graph Generator modules are being implemented in tandem and, when completed, will make the DR-LINK system fully operational. Full system testing will be conducted for the eighteen month TIPSTER testing. In the interim, our goal in this paper has been to describe the five unique modules which comprise DR-LINK and which, in combination, promise to provide a full system which has the necessary filtering power to make later processing more accurate and the depth of linguistic processing required to provide real conceptual level matching and retrieval. References Halliday, M. A. K. & Hasan, R. (1976). Cohesion in English. London, Longmans. Liddy, E.D. & Paik, W. (1992). Statistically-Guided word sense disambiguation. In Prnceedinos of MAI Fall Symposium Series: Probabilistic aooroaches to natural language. Menlo Park, CA: AAAI. Liddy, E.D., Paik, W., Mcvearry, K. & Yu, E. (In press). Automatic discourse-level structuring of newspaper texts: Empirical testing of a model. Liddy, E.D., Paik, W. & Woelfel, J. (1992). Use of subject field codes from a machine-readable dictionary for automatic classification of documents. Proceedings of 3rd ASIS Classification Research Workshop. Meteer, M., Schwartz, R. & Weischedel, R. (1991). POST: Using probabilities in language processing. Proceedings of the Twelfth International Conference on Artificial Intelligence. Sydney, Australia. Myaeng, S. H. (1992) Using conceptual graphs for information retrieval: a framework for representation and flexible inferencing. Proceedings of Symposium on Document Analysis and Information Retrieval, Las Vegas, March 16-18. Myaeng, S. H. & Khoo, C. (1992). On uncertainty handling in plausible reasoning with conceptual graphs. Proceedings of 7th Workshop on Conceptual Graphs, Las Cruces, NM, July, 1992. 128