Degree Name

Doctor of Philosophy


School of Computing and Information Technology


The size of data shared over the Internet today is gigantic. A big bulk of it comes from postings on social networking sites such as Twitter and Facebook. Some of it also comes from online news sites such as CNN and The Onion. This type of data is very good for data analysis since they are very personalized and specific. For years, researchers in academia and various industries have been analyzing this type of data. The purpose includes product marketing, event monitoring, and trend analysis. The highest usage for this type of analysis is to find out the sentiments of the public about a certain topic or product. This field is called sentiment analysis. The writers of such posts have no obligation to stick to only literal language. They also have the freedom to use figurative language in their publications. Hence, online posts can be categorized into two: Literal and Figurative. Literal posts contain words or sentences that are direct or straight to the point. On the contrary, figurative posts contain words, phrases, or sentences that carry different meanings than usual. This could flip the whole polarity of a given post. Due to this nature, it can jeopardize sentiment analysis works that focus primarily on the polarity of the posts. This makes figurative language one of the biggest problems in sentiment analysis. Hence, detecting it would be crucial and significant. However, the study of figurative language detection is non-trivial. There have been many existing works that tried to execute the task of detecting figurative language correctly, with different methodologies used. The results are impressive but still can be improved. This thesis offers a new way to solve this problem. There are essentially seven commonly used figurative language categories: sarcasm, metaphor, satire, irony, simile, humor, and hyperbole. This thesis focuses on three categories. The thesis aims to understand the contextual meaning behind the three figurative language categories, using a combination of deep learning architecture with manually extracted features and explore the use of well know machine learning classifiers for the detection tasks. In the process, it also aims to describe a descending list of features according to the importance. The deep learning architecture used in this work is Convolutional Neural Network, which is combined with manually extracted features that are carefully chosen based on the literature and understanding of each figurative language. The findings of this work clearly showed improvement in the evaluation metrics when compared to existing works in the same domain. This happens in all of the figurative language categories, proving the framework’s possession of quality.

FoR codes (2008)

0801 ARTIFICIAL INTELLIGENCE AND IMAGE PROCESSING, 080107 Natural Language Processing, 080109 Pattern Recognition and Data Mining



Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.