論文ID: TJSKE-D-23-00027
This paper proposes a system to estimate the emotions of Japanese words by considering the context in which they are used. It is expected to be used for realizing a dialog system that can understand and empathize with users. The proposed system uses word embeddings from a pre-trained language model: BERT to create a 10-dimensional emotion vector that expresses 10 different emotions. A 3-layer neural network with an intermediate layer with 400 nodes is used for training. Text data for learning does not need to be labeled for emotions. The proposed system automatically extracts emotional words to give its surrounding words information about the emotional words. The vocabulary of the proposed system is much larger than existing dictionary-like methods. We carried out extensive evaluation experiments and the results show that the proposed system can estimate context-sensitive word emotion.