Regresar

UAE at EmoSPeech–IberLEF2024: Integrating Text and Audio Features with SVM for Emotion Detection

Abstract:

Automatic emotion recognition (AER) has long been a significant challenge and is becoming increasingly important in various fields such as health, psychology, social sciences, and marketing. The EmoSPeech shared task at IberLEF 2024 aims to advance AER by addressing classification challenges, including feature selection for emotion discrimination, the scarcity of real-world multimodal datasets, and the complexity of combining different features. This task includes two subtasks: text-based AER and multimodal AER, emphasizing the novel aspect of multimodal AER by evaluating language models on authentic datasets. This paper presents the contributions of the UAE team to both subtasks. For Task 1, we used text embeddings from the pre-trained language model BETO and classified emotions using the SVM algorithm, achieving an M-F1 score of 0.51, outperforming the baseline and ranking 9th. For Task 2, we extended this approach by incorporating audio features from the Wav2Vec 2.0 model, resulting in an M-F1 score of 0.56 and a ranking of 7th. These results outperformed the baseline, demonstrating that audio features complement text features and improve the performance of the unimodal model.