Pairwise Learning Approach Using Siamese Neural Network For Contextual Sarcasm Detection

The uprising trends of microblogging platforms has led to a shift in formulation of speech, with individuals opting for more attention-grabbing techniques, oftentimes resorting to sarcasm. Sarcasm, a form of irony used to convey contempt or humor, necessitates contextual understanding for detection. This work introduces a pairwise learning approach for context-based sarcasm detection on microblogging platforms. The proposed approach utilizes a variant of the siamese neural network trained on conversation-response pairs. Pre-trained BERT-Large is employed to extract features from input pairs, followed by separate branches of LSTM network. Outputs of them were merged before feeding it to the multi-layer LSTM and hidden layers. Unlike traditional siamese models, the intermediate outputs are merged and processed further to serve the fine-tuning purpose and eventually it aims to classify input data as sarcastic or non-sarcastic. Through extensive experimental evaluation conducted on Twitter and Reddit datasets, we achieved notable F1-scores of 0.73 and 0.66, respectively. Although our results demonstrate the system's efficacy, they did not surpass existing state-of-the-art systems. While the proposed approach provides a promising solution for sarcasm detection in microblogging platforms, further investigations are needed to address the existing limitations and optimize the system’s performance. This research contributes to the understanding of attention-grabbing techniques in Natural Language Processing (NLP) and lays the groundwork for future advancements in sarcasm detection.

I. Introduction

Contextual sarcasm detection involves identifying sarcastic remarks within a given set of data by considering the contextual cues surrounding them. The primary focus of this research is to detect sarcasm by leveraging contextual information derived from conversational data. Additionally, we will highlight the specific approach we will employ to tackle this problem, namely pairwise learning. Our primary objective is to address the intricate challenge of detecting sarcasm in conversation threads on microblogging platforms. Our aim is to develop an innovative pairwise learning model capable of discerning whether the response to a conversation is sarcastic or not, by closely considering the contextual information of the entire exchange. We seek to establish a strong connection between the background context of a conversation and the subsequent response it elicits, empowering our model to effectively classify pairs of context and response as either sarcastic or non-sarcastic.