Stephen Tomlinson and Tom Blackwell. In E. M. Voorhees and D. K. Harman, editors, Proceedings of the Ninth Text REtrieval Conference (TREC-9), Gaithersburg, Maryland, November 2000. NIST Special Publication 500-249.
Hummingbird submitted ranked result sets for the Main Web Task (10GB of web data) and Large Web Task (100GB) of the TREC-9 Web Track, and for Stage 2 of the TREC-9 Query Track (43 variations of 50 queries). SearchServer's Intuitive Searching™ produced the highest Precision@5 score (averaged over 50 web queries) of all Title-only runs submitted to the Main Web Task. SearchServer's approximate text searching and linguistic expansion each increased average precision for web queries by 5%. Enabling SearchServer's document length normalization increased average precision for web queries by 10-30% and for long queries by 100%. Squaring the importance of the inverse document frequency (relevance method 'V2:4') increased average precision in the query track by 5%. Blind query expansion decreased average precision of highly relevants for web queries by almost 15%; the same method was neutral when counting all relevants the same.
Last Updated: 2003 Feb 25
Comments are welcome at comments@stephent.com.