LDEB -- Label Digitization with Emotion Binarization and Machine Learning for Emotion Recognition in Conversational Dialogues
This addresses the problem of entangled emotions in dialogues for conversational AI, but it appears incremental as it builds on existing techniques like text normalization and encoding.
The paper tackled emotion recognition in conversational dialogues by proposing LDEB, a method using label digitization and emotion binarization to disentangle nested emotions, resulting in models achieving up to 78% accuracy and 76% precision.
Emotion recognition in conversations (ERC) is vital to the advancements of conversational AI and its applications. Therefore, the development of an automated ERC model using the concepts of machine learning (ML) would be beneficial. However, the conversational dialogues present a unique problem where each dialogue depicts nested emotions that entangle the association between the emotional feature descriptors and emotion type (or label). This entanglement that can be multiplied with the presence of data paucity is an obstacle for a ML model. To overcome this problem, we proposed a novel approach called Label Digitization with Emotion Binarization (LDEB) that disentangles the twists by utilizing the text normalization and 7-bit digital encoding techniques and constructs a meaningful feature space for a ML model to be trained. We also utilized the publicly available dataset called the FETA-DailyDialog dataset for feature learning and developed a hierarchical ERC model using random forest (RF) and artificial neural network (ANN) classifiers. Simulations showed that the ANN-based ERC model was able to predict emotion with the best accuracy and precision scores of about 74% and 76%, respectively. Simulations also showed that the ANN-model could reach a training accuracy score of about 98% with 60 epochs. On the other hand, the RF-based ERC model was able to predict emotions with the best accuracy and precision scores of about 78% and 75%, respectively.