Date of Award

May 2016

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

School of Information Studies

Advisor(s)

Howard R. Turtle

Second Advisor

Elizabeth D. Liddy

Keywords

emotion categories, fine-grained emotion classification, machine learning, microblog text, sentiment analysis, Twitter

Subject Categories

Social and Behavioral Sciences

Abstract

Automatic emotion detection in text is concerned with using natural language processing techniques to recognize emotions expressed in written discourse. Endowing computers with the ability to recognize emotions in a particular kind of text, microblogs, has important applications in sentiment analysis and affective computing. In order to build computational models that can recognize the emotions represented in tweets we need to identify a set of suitable emotion categories. Prior work has mainly focused on building computational models for only a small set of six basic emotions (happiness, sadness, fear, anger, disgust, and surprise). This thesis describes a taxonomy of 28 emotion categories, an expansion of these six basic emotions, developed inductively from data. This set of 28 emotion categories represents a set of fine-grained emotion categories that are representative of the range of emotions expressed in tweets, microblog posts on Twitter.

The ability of humans to recognize these fine-grained emotion categories is characterized using inter-annotator reliability measures based on annotations provided by expert and novice annotators. A set of 15,553 human-annotated tweets form a gold standard corpus, EmoTweet-28. For each emotion category, we have extracted a set of linguistic cues (i.e., punctuation marks, emoticons, emojis, abbreviated forms, interjections, lemmas, hashtags and collocations) that can serve as salient indicators for that emotion category.

We evaluated the performance of automatic classification techniques on the set of 28 emotion categories through a series of experiments using several classifier and feature combinations. Our results shows that it is feasible to extend machine learning classification to fine-grained emotion detection in tweets (i.e., as many as 28 emotion categories) with results that are comparable to state-of-the-art classifiers that detect six to eight basic emotions in text. Classifiers using features extracted from the linguistic cues associated with each category equal or better the performance of conventional corpus-based and lexicon-based features for fine-grained emotion classification.

This thesis makes an important theoretical contribution in the development of a taxonomy of emotion in text. In addition, this research also makes several practical contributions, particularly in the creation of language resources (i.e., corpus and lexicon) and machine learning models for fine-grained emotion detection in text.

Access

Open Access

Share

COinS