Enhancing Text-Based Emotion Detection in Turkish and English: A Sentence-Level Enrichment Approach Using BERT and DistilBERT

This study introduces a novel approach to emotion detection in text by focusing on sentence-level emotion enrichment using transformer-based models, specifically BERT and DistilBERT. In recent studies, it has been suggested that vector representations enriched with additional emotion information be used rather than solely relying on pre-trained model-based vectors in emotion detection tasks. Unlike these methods enriching word embeddings with emotional lexicon words, our approach enriches entire sentence vectors with emotional sentence representations and lexicon word vectors which have been refined by adding emotional information. This aims to capture the nuanced emotional context of sentences, addressing the limitations of existing word-based approaches. Three emotion-enrichment methods are employed and evaluated across two languages, Turkish and English. Evaluation metrics include in-category similarity and classification performance using various machine learning classifiers. Our results show that sentence-level enrichment yields better results compared to vector representations produced by original pre-trained models, resulting in significant improvements in in-category similarity and classification performance. Additionally, the success of emotionally enriched sentence representations has been compared with different sentence representation methods and has shown better classification performance. Our study highlights the potential of sentence-level emotion enrichment in advancing emotion detection systems and contributes valuable insights for future research in emotion processing tasks.