Fine-tuning Models for Final Disagreement Anticipation in Negotiation Mid-Dialogues

抄録

Negotiation is an essential form of social interaction; however, reaching a reasonable agreement in negotiations is difficult. This is especially due to the complexity of the issues and backgrounds of the negotiations. In addition, when the interests of the negotiators conflict, the negotiation may break down. Therefore, machine learning and natural language processing must be used to support negotiations in real life. However, few studies have reported the use of such approaches. In this study, we propose the task of anticipating final disagreements in negotiation mid-dialogues. By anticipating final disagreements in negotiations, we can support human negotiations by understanding the elements necessary for negotiations to reach an agreement and building a system that supports negotiations to reach an agreement. We aim to reduce the training time by fine-tuning a dialogue act-based GRU model and a text-based BERT pretrained on a large-scale dataset with different datasets compared to training on the target dataset alone. The results show that fine-tuning improves the predictive performance of machine learning models that detect disagreements in the early stages of negotiations, when the input process dialogue is in the middle of the negotiation. A comparative experiment between a dialogue act-based GRU model and a text-based BERT shows that the dialogue act-based GRU model is more accurate than the text-based BERT model. We also visualize the final attention layer of the BERT model for further analysis.