Okapi at TREC-4

  • Stephen Robertson ,
  • S. Walker ,
  • M. M. Hancock-Beaulieu ,
  • M. Gatford ,
  • A. Payne

The Fourth Text REtrieval Conference (TREC-4) |

Published by Gaithersburg, MD: NIST

Publication

The Okapi Basic Search System is a set-oriented, ranked-output system designed primarily for probabilistic-type retrieval of textual material using inverted indexes. City University have submitted interactive runs in all the previous TRECs (Text REtrieval Conferences), with fairly undistinguished results. This time, the main emphasis has been on the development of an entirely new interactive ad-hoc search system. On the non-interactive side, there has been further work on methods of selecting routing terms, while the automatic ad-hoc searching was done in more or less the same way as for TREC-3, but, in view of the very brief topic statements, a few runs were also done with manually edited queries. The interactive track at TREC is beginning to bear fruit. What we need now is extensive diagnostic analyses, to fill out the evidence provided by the summary results. The routing task continues to show that relatively heavy computation based on the training set can produce good results.