DWAEF: a deep weighted average ensemble framework harnessing novel indicators for sarcasm detection1

Abstract

Sarcasm is a linguistic phenomenon often indicating a disparity between literal and inferred meanings. Due to its complexity, it is typically difficult to discern it within an online text message. Consequently, in recent years sarcasm detection has received considerable attention from both academia and industry. Nevertheless, the majority of current approaches simply model low-level indicators of sarcasm in various machine learning algorithms. This paper aims to present sarcasm in a new light by utilizing novel indicators in a deep weighted average ensemble-based framework (DWAEF). The novel indicators pertain to exploiting the presence of simile and metaphor in text and detecting the subtle shift in tone at a sentence’s structural level. A graph neural network (GNN) structure is implemented to detect the presence of simile, bidirectional encoder representations from transformers (BERT) embeddings are exploited to detect metaphorical instances and fuzzy logic is employed to account for the shift of tone. To account for the existence of sarcasm, the DWAEF integrates the inputs from the novel indicators. The performance of the framework is evaluated on a self-curated dataset of online text messages. A comparative report between the results acquired using primitive features and those obtained using a combination of primitive features and proposed indicators is provided. The highest accuracy of 92% was achieved after applying DWAEF, the proposed framework which combines the primitive features and novel indicators together as compared to 78.58% obtained using Support Vector Machine (SVM) which was the lowest among all classifiers.