Pseudo-Relevance Feedback and Title Re-Ranking for Chinese Information Retrieval
Robert W. P. Luk, Kam‐Fai Wong · NTCIR · 2004
In our formal runs, we have experimented with the retrieval based on character-based indexing and hybrid term indexing because these are more distinct types of indexing for better pooling. We confirmed that character-based indexing did not produce relatively good retrieval effectiveness. We have also experimented with three new pseudo-relevance feedback (PRF) methods. These new methods were able to achieve slightly better performance compared with our original PRF method for short queries at the same time using only 60 expanded terms instead of 140 expanded terms, thereby reducing the retrieval time by abo ut 30 percentage points. We have experimented with a novel reranking strategy, called title re-ranking. This strategy rewards documents which title terms match with the terms in the title query. Title re-ranking is able to improve the performance but it h urts the performance of long queries when PRF is used with it. For title queries, our best MAP achieved was 24.9% using both title re-ranking and PRF based on rigid relevance judgment. For long queries using our original PRF, our best MAP was 26.0%.