Agreement Tracking for Multi-Issue Negotiation Dialogues

Anonymous

Abstract: Automated negotiation support systems seek to help human negotiators reach more favorable outcomes. When supporting multi-issue negotiations, it is crucial for these support systems to accurately track the agreements reached by the participants in real-time (e.g., an employer and a candidate negotiating over multiple issues such as salary, hours, and promotions before finalizing a job offer). However, existing task formulations are either geared towards differing dialogue paradigms (e.g., dialogue state tracking is aimed at task-oriented dialogues) or generate a single, unstructured output at the end of the dialogue (e.g., meeting summarization). We introduce the novel task of agreement tracking for two-party multi-issue negotiations, in which the goal is to continuously track the agreements over a \textit{structured} state space. Due to the absence of large-scale corpora with turn-level annotations in this domain, we propose a simple, but strong initial baseline for our task based on transfer-learning a T5 model from the dialogue state tracking task on the MultiWOZ 2.4 corpus of task-oriented dialogues. Additionally, we also study the sample-efficiency of our approach by running experiments on smaller fractions of the training data. Our results demonstrate the challenging nature of the agreement tracking task and the need for more data-efficient approaches to solve it.

Paper Type: long

Research Area: Dialogue and Interactive Systems