Date of Graduation

Summer 2019

Degree

Master of Natural and Applied Science in Computer Science

Department

Computer Science

Committee Chair

Anthony Clark

Keywords

social media, qualitative study, image, text, multimodality, sarcasm detection, attention model

Subject Categories

Artificial Intelligence and Robotics | Graphics and Human Computer Interfaces | Other Computer Sciences | Social Media | Software Engineering

Abstract

In recent times, a major share of human communication takes place online. The main reason being the ease of communication on social networking sites (SNSs). Due to the variety and large number of users, SNSs have drawn the attention of the computer science (CS) community, particularly the affective computing (also known as emotional AI), information retrieval, natural language processing, and data mining groups. Researchers are trying to make computers understand the nuances of human communication including sentiment and sarcasm. Emotion or sentiment detection requires more insights about the communication than it does for factual information retrieval. Sarcasm detection is particularly more difficult than categorizing sentiment. Because, in sarcasm, the intended meaning of the expression by the user is opposite to the literal meaning. Because of its complex nature, it is often difficult even for human to detect sarcasm without proper context. However, people on social media succeed in detecting sarcasm despite interacting with strangers across the world. That motivates us to investigate the human process of detecting sarcasm on social media where abundant context information is often unavailable and the group of users communicating with each other are rarely well-acquainted. We have conducted a qualitative study to examine the patterns of users conveying sarcasm on social media. Whereas most sarcasm detection systems deal in word-by-word basis to accomplish their goal, we focused on the holistic sentiment conveyed by the post. We argue that utilization of word-level information will limit the systems performance to the domain of the dataset used to train the system and might not perform well for non-English language. As an endeavor to make our system less dependent on text data, we proposed a multimodal approach for sarcasm detection. We showed the applicability of images and reaction emoticons as other sources of hints about the sentiment of the post. Our research showed the superior results from a multimodal approach when compared to a unimodal approach. Multimodal sarcasm detection systems, as the one presented in this research, with the inclusion of more modes or sources of data might lead to a better sarcasm detection model.

Copyright

© Dipto Das

Available for download on Friday, May 01, 2020

Open Access

Share

COinS