Agreement Tracking for Multi-Issue Negotiation Dialogues

Amogh Mannekote, Bonnie J. Dorr, Kristy Elizabeth Boyer

University of Florida

“Automated negotiation support systems aim to help human negotiators reach more favorable outcomes in multi-issue negotiations (e.g., an employer and a candidate negotiating over issues such as salary, hours, and promotions before a job offer). To be successful, these systems must accurately track agreements reached by participants in real-time. Existing approaches either focus on task-oriented dialogues or produce unstructured outputs, rendering them unsuitable for this objective. Our work introduces the novel task of agreement tracking for two-party multi-issue negotiations, which requires continuous monitoring of agreements within a structured state space. To address the scarcity of annotated corpora with realistic multi-issue negotiation dialogues, we use GPT-3 to build GPT-NEGOCHAT, a synthesized dataset that we make publicly available.

Negotiation dialogues are common in both adversarial and collaborative contexts. However, a long line of foundational research in psychology and business has established that in general, humans tend to be poor negotiators, often failing to maximize favorable outcomes.

We demonstrate a deficiency in existing methods for three closely-related tasks: building negotiation dialogue agents, summarization of meetings, and dialogue state tracking for task-oriented dialogues.”

Dialogue State Tracking (DST): The goal of DST in task-oriented dialogue is to extract users’ goals by inferring the values for a predefined set of keys (commonly referred to as “slots”). As we shall see later, we re-purpose this notion of a slot to track agreement over a single issue in our multi-issue negotiation dialogue setup.

Dialogue state tracking has been a long-standing task in task-oriented dialogue systems literature (Williams et al., 2016; Jacqmin et al., 2022; Zhao et al., 2021; Rastogi et al., 2020). Although there has been extensive work in recent years to improve the ability of task-oriented DST models to generalize to unseen domains with zero-shot (Lin et al., 2021b; Campagna et al., 2020) and fewshot (Wu et al., 2019) models, these are still limited to form-filling dialogues (e.g., restaurant reservation and hotel booking). Moreover, these models are challenged by agreement tracking, as it requires explicit agreement from both interlocutors, unlike DST that estimate the goals of a single interlocutor. Hence, the question of how effectively state-of-the-art DST techniques can be used for state tracking in vastly different dialogue paradigms (such as negotiation) remains open-ended.