MOHAMAD AMIRUL HAZIQ BIN AHMAD AFFZAN
Rohana Mahmud
ASSOCIATE PROF. DR. SHEENA KAUR A/P JASWANT SINGH
Sentiment analysis traditionally relies on bag-of-words methods to mathematically represent documents based on word frequency (Atanu Dey, 2018). However, these approaches overlook semantic connections within a document, treating all clauses equally and failing to capture the importance of specific subordinate clauses. To address this limitation, a discourse-aware sentiment analysis method is needed, capable of recognizing the salience of individual subordinate clauses and differentiating the relevance of sentences based on their function.
The application of theories like Rhetorical Structure Theory (RST), as demonstrated by (Mathias Kraus, 2019), offers a framework for representing documents based on their discourse structure. By expanding representation learning techniques to incorporate the complete discourse tree, including relation types, tree depth, and hierarchy labels, their work advances the field of discourse-level sentiment analysis.
This project aims to extend the research conducted by (Mathias Kraus, 2019) by applying and evaluating their discourse-aware sentiment analysis solution in the context of scams. To achieve this, a scam-specific dataset capturing the language of pressure and coercion used by scammers will be collected. Additionally, a domain-specific lexicon will be developed to identify words related to pressure and coercion, addressing the limitations of existing lexicons such as SentiWordNet 3.0.
With full implementation, the project aims to achieve the following objectives: 1) To identify the linguistic cues and features associated with pressure and coercive language based on data sources such as websites and online forums within a span of 2 semesters. 2) To design and develop a discourse-level sentiment analyzer model that capable of identifying pressure and coercive language used by online fraudsters on digital communication platforms within a span of 2 semesters. 3) To evaluate the discourse-level sentiment analyzer model by the end of the project.