PITT at TREC 2011 session track
|PITT at TREC 2011 session track|
|Author(s)||Jiang J., Han S., Wu J., He D.|
|Published in||NIST Special Publication|
|Keyword(s)||Query language model, Relevance feedback, Session, TREC (Extra: Ad Hoc retrieval, Clickthrough data, Dependence model, Evaluation results, Improve performance, Model informations, On currents, Pseudo relevance feedback, Query language model, Query model, Relevance feedback, Search performance, Search sessions, Session, TREC, Wikipedia, Computational linguistics, Mixing, Query languages, Information retrieval)|
|Article||BASE, CiteSeerX, Google Scholar|
|Web||Ask, Bing, Google (PDF), Yahoo!|
|Download and mirrors|
|Local copy||Not available|
|Remote mirror(s)||Not available|
|Export and share|
|BibTeX, CSV, RDF, JSON|
|Browse properties · List of conference papers|
In this paper, we introduce our approaches for TREC 2011 session track. Our approaches focus on combining different query language models to model information needs in a search session. In RL1 stage, we build ad hoc retrieval system using sequential dependence model (SDM) on current query. In RL2 stage, we build query language models by combining SDM features (e.g. single term, ordered phrase, and unordered phrase) in both current query and previous queries in the session, which can significantly improve search performance. In RL3 and RL4, we combine query model in RL2 with two different pseudo-relevance feedback query models: in RL3, we use top ranked Wikipedia documents from RL2's results as pseudo-relevant documents; in RL4, snippets of the documents clicked by users in a search session are used. Our evaluation results indicate: texts of previous queries in a session are effective resources for estimating query models and improving search performance; mixing query model in RL2 with the query model estimated using click-through data (in RL4) can improve performance in evaluation setting that considers all subtopics, but no improvement is observed in evaluation setting that considers the only subtopic of current query; our methods of mixing query model in RL2 with query model in RL3 did not improve search performance over RL2 in any of the two evaluation settings.
- This section requires expansion. Please, help!
Probably, this publication is cited by others, but there are no articles available for them in WikiPapers.