Sentence embedding based emotion recognition from text data
Manabu Ito, Konstantin Markov · 2022
Automatic emotion recognition from text is an important task in the field of natural language processing (NLP) with applications in data mining, e-learning, information filtering systems, human-computer interaction, and internet services. Recent advances in machine learning and deep neural networks have boosted the NLP systems' performance tremendously. The availability of large pre-trained language models has simplified the feature extraction and facilitated the building of systems from small amounts of data by transfer learning. In this study, we investigate and compare various methods of sentence embedding including simple embedding matrix as well as sophisticated models such as BERT. The recognition back-end consists of standard SVR or Feed-Forward DNN regressors. In our experiments, we used the EmoBank corpus where each sentence is labeled with Valence-Arousal scores. The results clearly show the benefits of the transfer learning and fine-tuning of pre-trained models with respect to classical model training from scratch.