🔗 Thesis Link: https://tdr.lib.ntu.edu.tw/handle/123456789/91309
Affective states profoundly influence human behaviors, motivations, and decisions, making them a crucial factor to consider in dialogue systems aimed at simulating or predicting human reactions. To improve the conversational experience and user satisfaction in dialogue systems, prediction of users’ affective states is essential. Existing research primarily focuses on recognizing affective states within dialogue history, neglecting the proactive forecasting of upcoming affective states. However, the ability to forecast upcoming affective states proactively can enable dialogue systems to adjust responses in advance.
Therefore, in this research, we concentrate on the task of Sentiment Forecasting in Dialogue and propose a multi-task learning model by incorporating sentiment recognition and dialogue act recognition within dialogue history sequence and upcoming dialogue act forecasting as auxiliary tasks. We also develop a novel mechanism to dynamically adjust the importance of each task during training. Experimental results demonstrate the effectiveness of our model in capturing diverse sentiment-related information and learning better sentiment representations, leading to improved sentiment forecasting performance, surpassing existing state-of-the-art methods.
Additionally, to enhance real-world applicability, we collect a new dialogue dataset simulating common dialogue scenarios and conduct domain transfer experiments, further validating the efficacy of our proposed domain transfer methods. Our research emphasizes the significance of multi-task learning and domain transfer in sentiment forecasting tasks, providing a foundation for developing more sophisticated sentiment analysis techniques, improving sentiment understanding in dialogue systems, and enhancing user experiences.
📌 Keywords: Dialogue System, Sentiment Analysis, Sentiment Forecasting in Dialogue, Multi-task Learning, Transfer Learning, Domain Adaptation