Supervisor

Marina Iantorno

Programme

MSc in Data Analytics

Subject

Computer Science

Department

Computer Science

Abstract

Emojis have become an integral element of online communications, serving as a powerful, under-utilised resource for enhancing pragmatic understanding in NLP. Previous works have highlighted their potential for improvement of more complex tasks such as the identification of figurative literary devices including sarcasm due to their role in conveying tone within text. However present state-of-the-art does not include the consideration of emoji or adequately address sarcastic markers such as sentiment incongruence. This work aims to integrate these concepts to generate more robust solutions for sarcasm detection leveraging enhanced pragmatic features from both emoji and text tokens. This was achieved by establishing methodologies for sentiment feature extraction from emojis and a depth statistical evaluation of the features which characterise sarcastic text on Twitter. Current convention for generation of training data which implements weak-labelling using hashtags or keywords was evaluated against a human-annotated baseline; postulated validity concerns were verified where statistical evaluation found the content features deviated significantly from the baseline, highlighting potential validity concerns for many prominent works on the topic to date. Organic labelled sarcastic tweets containing emojis were crowd sourced by means of a survey to ensure valid outcomes for the sarcasm detection model. Given an established importance of both semantic and sentiment information, a novel sentiment-aware attention mechanism was constructed to enhance pattern recognition, balancing core features of sarcastic text: sentiment incongruence and context. This work establishes a framework for emoji feature extraction; a key roadblock cited in literature for their use in NLP tasks. The proposed sarcasm detection pipeline successfully facilitates the task using a GRU neural network with sentiment-aware attention, at an accuracy of 73% and promising indications regarding model robustness as part of a framework which is easily scalable for the inclusion of any future emojis released. Both enhanced sentiment information to supplement context in addition to consideration of the emoji were found to improve outcomes for the task.

Date of Award

Winter 2023

Full Publication Date

November 2023

Access Rights

open access

Document Type

Capstone Project

Resource Type

thesis

Download

Share

COinS