Entity-relation search: context pattern driven relation ranking

Zhang, Tianxiao

Permalink

https://hdl.handle.net/2142/90846 Copy

Description

Title

Entity-relation search: context pattern driven relation ranking

Author(s)

Zhang, Tianxiao

Issue Date

2016-04-28

Director of Research (if dissertation) or Advisor (if thesis)

Chang, Kevin C.

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Entity Relation Search
• Context Pattern
• PubMed Data

Abstract

A traditional page link-based search system is not adequate for users intending to query data efficiently. For instance, emergent phenomena reveal that some entity-based search engines, such as EntityRank, directly return answers (target entities) to users instead of web pages. Most of the time, however, compared to searching for interested entities, users more often focus on relationships among entities. To our knowledge, there is only one web search system that automatically extracts relations from massive unstructured corpora. This system is referred to as OpenIE, which indeed brings us one step closer to an entity relation-based system. Nevertheless, its system extracts only direct relations between a pair of entities and ranks simply by occurrence frequency. The monotone pattern extraction, adopted in their relation phrase extraction model, provides high quality entity relations but also fail to return many potential true relations in the corpus, which has been explained in Section 4.2 and affects recall significantly shown in 5.3. In addition, it is difficult for users to retrieve their interested and true relations from massive relation candidate set without a qualified ranking model. Consequently, there still exists a gap between the system and users for retrieving entity relations efficiently by a simple query. To assist users to find their interested relations efficiently, this thesis specifically focuses on the core challenges of the ranking model. Naturally, the quality of each relation candidate is largely relevant to its context. Thus, to evaluate various conditions, a novel idea of context patterns driven ranking has been introduced. After evaluating our online prototype on millions of PubMed medical abstracts, we show that our system performs better than the OpenIE system on both precision and recall. Note that this rich and novel system is the product of a collaborative team effort comprised of the following members: Zequn Zhang, Jiarui Xu, and Varun Berry, and supervised by Professor Kevin Chang.

Graduation Semester

2016-05

Type of Resource

text

Permalink

http://hdl.handle.net/2142/90846

Copyright and License Information

Copyright 2016 Tianxiao Zhang

Owning Collections