SP500207 Query Improvement in INformation Retrieval Using Genetic Algorithms - A Report on the Experiments of the TREC Project chapter J. Yang R. Korfhage E. Rasmussen National Institute of Standards and Technology Donna K. Harman 10. Further Studies In dealing with these large databases, our algorithm shows many interesting phenomena which we could not deal with due to the short time period of TREC, and which could form the agenda for further studies. (1) The document term weights Although we use the vector space model mentioned in Section 3, unweighted keyterms are assigned to describe the documents. We have mentioned the two factors based on which we made the decision. Moreover, the databases consist of a variety of subjects which may make the general term weights less reliable than in most experimental databases, which deal with only a narrow area. While our Cranfield study supports this decision, the effect of weighted and unweighted document keywords in large databases should be further studied. (2) Window size for document retrieval Although we established the maximum size of the retrieved document set as 40, it would be interesting to set the window size smaller than 40 at the beginning of the process and gradually increase the size after the query individuals move toward convergence. Various thresholds and cutoff points are often arbitrarily set. Further experimentation is planned to investigate the effect of changes in these values. (3) Document classification In this experiment we found many cases in which different query individuals retrieved different documents. This may arise because the database contains relevant documents that belong to different classes, that can be retrieved only by using different queries. This fact would never be discovered by systems that rely on a single query. Our present algorithm aims at convergence to a single query form. We need a mechanism which will detect the development of distinct classes, and automatically separate the query into two or more distinct queries. Similar ideas, not based on genetic algorithms, have been discussed in document classification (Worona, 1971) and in query splitting (Borodin, et. al., 1971). (4) Term addition and deletion In our feedback process, no terms were added to or removed from the original query. This may cause lower recall. However, the determination of which terms should be added or removed is not an easy task (e.g., Harman, 1988, 1992). It would be interesting to see how this problem can be handled in the genetic process. Reference Aalbersberg, I.J. (1992). Incremental relevance feedback. Proceedings of the lSth Annual International SIGIR Conference, pp 11-22. Copenhagen, Denmark (June, 1992). 56