Enriching Transformer-Based Embeddings for Emotion Identification in an Agglutinative Language: Turkish

Text-based emotion detection is an important and expanding research area due to the increasing accessibility of written data via the Internet and social media. Vector space models, such as semantic and contextual methods, are frequently used in many domains in natural language processing. Currently, to improve performance in emotion/sentiment detection studies, a new research area has emerged, which involves adding extra emotion information (emotion enrichment) to these models. Furthermore, as emotion depends on multiple parameters, the success of enrichment may vary based on different languages. In this study, we applied two emotion-enrichment methods on emerging transformer-based models [bidirectional encoder representations from transformers (BERT), a robustly optimized BERT pretraining approach, a distilled version of BERT, and efficiently learning an encoder that classifies token replacements accurately] and a traditional semantic model (Word2Vec) (as a baseline) on the Turkish (a highly agglutinative language) dataset. The performance was analyzed with classification models and cosine-similarity metrics.