LAStER: Lexical Data Augmentation with Stacking Ensembles for Robust Emotion Detection in Low-Resource Corpora

Abstract

Emotion classification is a crucial task in Natural Language Processing (NLP) with applications in sentiment analysis, social media monitoring, and user modeling. Conventional augmentation methods often fail to identify nuanced emotional cues. We propose Lexical Data Augmentation with Stacking Ensembles for Robust Emotion Detection (LAStER), which integrates emotional lexicons with ensemble learning techniques. We augmented the ISEAR and the English datasets from SemEval-2025 Task 11, with explicit emotional indicators from NRC Emotion Lexicon. Our findings indicate that LAStER significantly enhances performance in multi-label contexts, with the stacking configuration utilized on the SemEval dataset attaining a subset accuracy of 0.229 and a macro F1-score of 0.324. The stacking configuration utilized for the single-label attained an accuracy of 0.350 and an F1-score of 0.182. The results demonstrate that the integration of lexical augmentation with stacking ensembles is an efficacious method for emotion identification, and suggest that LAStER operates competently with multi-label datasets.