Providing an effective mechanism for personal information retrieval is important for many applications, and requires different techniques than have been developed for general web search. This thesis focuses on developing retrieval models and representations for personal search, and on designing evaluation frameworks that can be used to demonstrate retrieval effectiveness in a personal environment.
From the retrieval model perspective, personal information can be viewed as a collection of multiple document types each of which has unique metadata. Based on this perspective, we propose a retrieval model that exploits document metadata and multi-type structure. Proposed retrieval models were found to be effective in other structured document collections, such as movies and job descriptions.
Associative browsing is another search method that can complement keyword search. To support this type of search, we propose a method for building an association graph representation by combining multiple similarity measures based on a user's click patterns. We also present a learning techniques for refining the graph structure based on user's clicks.
Evaluating these methods is particularly challenging for personal information due to privacy issues. This thesis introduces a set of techniques that enables realistic and repeatable evaluation of techniques for personal information retrieval. In particular, we describe techniques for simulating test collections and show that game-based user studies can collect more realistic usage data with relatively small cost.
Identifer | oai:union.ndltd.org:UMASS/oai:scholarworks.umass.edu:open_access_dissertations-1650 |
Date | 01 September 2012 |
Creators | Kim, Jinyoung |
Publisher | ScholarWorks@UMass Amherst |
Source Sets | University of Massachusetts, Amherst |
Detected Language | English |
Type | text |
Format | application/pdf |
Source | Open Access Dissertations |
Page generated in 0.001 seconds