Hierarchical classification for dialogue act estimation: Addressing data imbalance in Japanese conversation corpus

抄録

In this study, we propose a hierarchical classification approach based on deep learning to address imbalanced data in dialogue act estimation. Using the CEJC corpus, we classified 20 dialogue acts across six levels, employing Word2Vec for feature extraction and SVM for classification. The proposed approach achieved a 73.07% accuracy and a weighted F1-score of 70.39 in testing. Levels with distinctive dialogue acts showed high accuracy, whereas overlapping features or insufficient data posed challenges. These results demonstrate the effectiveness of hierarchical classification for dialogue act estimation and highlight areas for further improvement, particularly in handling data imbalance and enhancing feature representation.