IIT at TREC 2003, Task Classification and Document Structrure for Known-Item Search

This year’s TREC 2003 web task incorporated two retrieval tasks into a single set of experiments for Known-Item retrieval. We hypothesized that not all retrieval tasks should use the same retrieval approach when a single search entry point is used. We applied task classifiers on top of traditional web retrieval approaches. Our traditional retrieval is based on fusion of result sets generated by query runs over independent parts of the document structure. Our task classifiers combine query term analysis with known information resources and URL depth. This approach to task classification shows promise: our classified runs improved overall MRR effectiveness over our traditional retrieval results by ~10%; provided an MRR of .665; ranked 87% of relevant results in the top 10; correctly ranked the #1result 56% of the time. 67% of the queries performed above the average, and 49% above the median.