A unified framework to identify and extract uncertainty cues, holders, and scopes in one fell-swoop

Al-Sabbagh, Rania Mostafa

Permalink

https://hdl.handle.net/2142/78621 Copy

Description

Title

A unified framework to identify and extract uncertainty cues, holders, and scopes in one fell-swoop

Author(s)

Al-Sabbagh, Rania Mostafa

Issue Date

2015-04-17

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

• Girju, Roxana
• Diesner, Jana

Doctoral Committee Chair(s)

Girju, Roxana

Committee Member(s)

• Benmamoun, Elabbas
• Hockenmaier , John

Department of Study

Linguistics

Discipline

Linguistics

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Date of Ingest

2015-07-22T22:33:25Z

Keyword(s)

• Computational Semantics
• Semitic Languages
• Uncertainty
• Social Media Analysis

Abstract

Uncertainty refers to the language aspects that express hypotheses and speculations where propositions are held as (un)certain, (im)probable, or (im)possible. Automatic uncertainty analysis is crucial for several Natural Language Processing (NLP) applications that need to distinguish between factual (i.e. certain) and nonfactual (i.e. negated or uncertain) information. Typically, a comprehensive automatic uncertainty analyzer has three machine learning models for uncertainty detection, attribution, and scope extraction. To-date, and to the best of my knowledge, current research on uncertainty automatic analysis has only focused on uncertainty attribution and scope extraction, and has typically tackled each task with a different machine learning approach. Furthermore, current research on uncertainty automatic analysis has been restricted to specific languages, particularly English, and to specific linguistic genres, including biomedical and newswire texts, Wikipedia articles, and product reviews. In this research project, I attempt to address the aforementioned limitations of current research on automatic uncertainty analysis. First, I develop a machine learning model for uncertainty attribution, the task typically neglected in automatic uncertainty analysis. Second, I propose a unified framework to identify and extract uncertainty cues, holders, and scopes in one-fell swoop by casting each task as a supervised token sequence labeling problem. Third, I choose to work on the Arabic language, in contrast to English, the most commonly studied language in the literature of automatic uncertainty analysis. Finally, I work on the understudied linguistic genre of tweets. This research project results in a novel NLP tool, i.e., a comprehensive automatic uncertainty analyzer for Arabic tweets, with a practical impact on NLP applications that rely on uncertainty automatic analysis. The tool yields an F1 score of 0.759, averaged across its three machine learning models. Furthermore, through this research, the research community and I gain insights into (1) the challenges presented by Arabic as an agglutinative morphologically-rich language with a flexible word order, in contrast to English; (2) the challenges of the linguistic genre of tweets for uncertainty automatic analysis; and (3) the type of challenges that my proposed unified framework successfully addresses and boosts performance for.

Graduation Semester

2015-5

Type of Resource

text

Permalink

http://hdl.handle.net/2142/78621

Copyright and License Information

Copyright 2015 Rania Mostafa Al-Sabbagh

Owning Collections