Modeling Rich Interactions in Session Search - Georgetown University at TREC 2014 Session Track
Abstract
This year we participate in the TREC Session Track Task 1. We adopt the Query Change Model (QCM) weighted QCM, re-ranking, clustering, and error analysis in our approaches. The QCM retrieval model is employed to combine all queries in a session. QCM allows documents that are relevant to any query in a session to appear in the final retrieval list. Weighted QCM combines queries unevenly based on a prediction of query quality. It is based on the following intuition: if a query does not bring any document that leads to a SAT-Click from the user, it suggests that this query is poorly formed. Our re-ranking module is based on implicit feedback from the user; in this case the SAT-Clicked documents. The module boosts a document's ranking position if it has been SAT-Clicked in the session or in other sessions that share similar search topics. We apply K-means clustering algorithm to detect which sessions share similar search topics. Each unique term is one dimension of the vector and is weighted by its idf. We also apply session error analysis in RL3. From the query log, we first identify sessions with similar topics by clustering, then we use SAT-Clicks from most sessions to re-rank the documents for the sessions that the algorithm predicts as poorly issued sessions, i.e. more difficult session due to ill-form queries. Combining above approaches, we achieve a 20.9% nDCG@10 increment and a 13.0% P@10 increment from RL1 to RL2, and with utilization of the whole log data, we achieve a 4% nDCG@10 increment and a 0.5% P@10 increment from RL2 to RL3.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 2014
- Accession Number
- ADA618634
Entities
People
- Hui Yang
- Jiyun Luo
- Xuchu Dong
Organizations
- Georgetown University