A Study of Term Proximity and Document Weighting Normalization in Pseudo Relevance Feedback - UIUC at TREC 2009 Million Query Track
ILLINOIS UNIV AT URBANA-CHAMPAIGN DEPT OF COMPUTER SCIENCE
Pagination or Media Count:
In this paper, we report our experiments in the TREC 2009 Million Query Track. Our first line of study is on proximity-based feedback, in which we propose a positional relevance model PRM to exploit term proximity evidence so as to assign more weights to expansion words that are closer to query words in feedback documents. The second line of study is to improve the weighting of feedback documents in the relevance model by using a regression-based method to approximate the probability of relevance and thus the name RegRM. In the third line of study, we test a supervised approach for query classification. Besides, we also evaluate a selective pseudo feedback strategy which stops pseudo feedback for precision-oriented queries and only uses it for recall-oriented ones. The proposed PRM has shown clear improvements over the relevance model for pseudo feedback, suggesting that capturing the term proximity heuristic appropriately could lead to a better feedback model. RegRM performs as well as relevance model, but no noticeable improvement is observed. Unfortunately, the proposed query classification methods appear to not work well. The results also show that the proposed selective pseudo feedback may not work well, since precision-oriented queries can also benefit from pseudo feedback, though not as much as recall-oriented queries.
- Operations Research