Annotation-free location mention mining from text corpora

Liu, Tingcong

Permalink

https://hdl.handle.net/2142/125538 Copy

Description

Title

Annotation-free location mention mining from text corpora

Author(s)

Liu, Tingcong

Issue Date

2024-06-27

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

Han, Jiawei

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)

• Text Mining
• Pre-trained Language Model

Language

eng

Abstract

This thesis provides a novel framework to extract location mentions from text corpora. Location mention mining plays an important role at analyzing and extracting structured knowledge from real-world text corpora like news and social media. Existing methods mainly rely on NER models or semantic parsers to extract locations but suffer from the following problems: (a) Entities tagged by NER models as LOC or GPE may not represent locations in the context. For example, in the sentence S1: “Ukraine forces are approaching Russia-held Kherson”, only “Kherson” is the true location mention although “Ukraine” and “Russia” are also of GPE type; and (b) A semantic parser cannot recognize locations in verb phrases. In S1, although “Kherson” refers to a location, it cannot be extracted as a locative argument by a semantic parser because it does not follow a preposition. This thesis defines a new task, location mention mining, aiming at extracting from a corpus all the mentions corresponding to real-world locations based on the context, and propose an annotation-free method, LocMine, which (1) constructs location-indicative term repositories using a background corpus and a knowledge base, (2) extracts and mines context-free location mentions based on the repositories, and (3) classifies context-dependent location mentions with pre-trained language models. This thesis provides extensive experiments and case studies showing that LocMine achieves the best performance among all the compared methods in terms of the ability to mine a complete set of location mentions from real-world corpora.

Graduation Semester

2024-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/125538

Copyright and License Information

Copyright 2024 Tingcong Liu

Owning Collections