Leveraging Emotion-specific Features to Improve Transformer Performance for Emotion Classification

TL;DR

This paper enhances transformer-based emotion classification by integrating emotion-specific features and ensembling, achieving improved accuracy and macro F1 scores on news article essays.

Contribution

The paper introduces the use of emotion-specific representations and ensembling techniques to boost transformer performance in emotion classification tasks.

Findings

Achieved 0.619 accuracy on emotion classification

Attained 0.520 macro F1 score

Outperformed baseline transformer models

Abstract

This paper describes the approach to the Emotion Classification shared task held at WASSA 2022 by team PVGs AI Club. This Track 2 sub-task focuses on building models which can predict a multi-class emotion label based on essays from news articles where a person, group or another entity is affected. Baseline transformer models have been demonstrating good results on sequence classification tasks, and we aim to improve this performance with the help of ensembling techniques, and by leveraging two variations of emotion-specific representations. We observe better results than our baseline models and achieve an accuracy of 0.619 and a macro F1 score of 0.520 on the emotion classification task.