Modeling Search Engine’s Explorations in Dynamic Search: An Ontological Perspective
Yang, Grace Hui
Dynamic search is an information retrieval task, in which information systems retrieve documents for a user’s multiple queries. Each query starts a search iteration and aims to fulfill part of the user’s information need. Modeling search engine’s explorations in dynamic search serves to help search engines explore in the information space, retrieve relevant documents and fulfill the user’s information need. Previous work utilizes ad hoc retrieval models, such as language model to retrieve documents. However, such an approach treats each search iteration independently, fails to realize that each query reflects one facet on the user’s complex information need and closely connects to each other. Another approach employs topic modeling, such as Latent Dirichlet Allocation (LDA) to fulfill the user’s information need. In each iteration, the approach discovers potential topics of the user’s information need, and diversifies the search result by retrieving documents covering these topics. This thesis proposes to structure the user’s information need as an ontology (a topic hierarchy for knowledge representation), and to utilize topic transitions on the ontology to model search engine’s explorations in dynamic search. The ontology presents a clear landscape for search engine’s explorations and improves the effectiveness and efficiency of the user’s information seeking. The ontology can be obtained through extra resources, such as Wikipedia, or built on top of topic construction algorithms, such as nomothetic concept hierarchy construction method. In this thesis, we presume the ontology is presented to the search engine and focus on how the search engine efficiently achieves topic transitions on the ontology. Analogizing the search engine’s explorations on an ontology to a robot’s explorations in a world, we model the search engine’s explorations in dynamic search as a Reinforcement Learning (RL) problem and aim to learn a policy to optimize the topic transitions. We apply Multi- Armed Bandit (MAB) and Partially Observable Markov Decision Process (POMDP) to learn the search engine’s policy. We evaluate the model using the most recent Text REtrieval Conference Dynamic Domain track (TREC DD 2015) datasets. The result shows that our model is highly effective.
MetadataShow full item record
Showing items related by title, author, creator and subject.