CaSiNo: A Corpus of Campsite Negotiation Dialogues for Automatic Negotiation Systems

TL;DR

CaSiNo provides a large, realistic corpus of free-form campsite negotiations to advance automatic negotiation systems. It introduces a 1030-dialogue dataset with three goods, random priority orders, and nine strategy annotations; it analyzes correlations between dialogue strategies and outcomes and develops a multi-task strategy-prediction model with in-domain pre-training. The results show prosocial strategies correlate with higher satisfaction and mutual gains, and that multi-task learning substantially boosts strategy recognition for skewed labels. The dataset, annotations, and code are released to support research on human-machine negotiation and feedback-driven training of negotiation agents.

Abstract

Automated systems that negotiate with humans have broad applications in pedagogy and conversational AI. To advance the development of practical negotiation systems, we present CaSiNo: a novel corpus of over a thousand negotiation dialogues in English. Participants take the role of campsite neighbors and negotiate for food, water, and firewood packages for their upcoming trip. Our design results in diverse and linguistically rich negotiations while maintaining a tractable, closed-domain environment. Inspired by the literature in human-human negotiations, we annotate persuasion strategies and perform correlation analysis to understand how the dialogue behaviors are associated with the negotiation performance. We further propose and evaluate a multi-task framework to recognize these strategies in a given utterance. We find that multi-task learning substantially improves the performance for all strategy labels, especially for the ones that are the most skewed. We release the dataset, annotations, and the code to propel future work in human-machine negotiations: https://github.com/kushalchawla/CaSiNo

Figures (6)

• Figure 1: Architecture for multi-task strategy prediction. + represents element-wise summation.
• Figure 2: Visualizing task-specific self-attention layers for two examples from the test dataset for the first cv fold. The heatmap shows the attention scores for each token in the utterance for corresponding strategy labels.
• Figure 3: Screenshots from the data collection interface: Task Preview. This is a brief task description which the MTurkers see before signing up for our data collection task.
• Figure 4: Screenshots from the data collection interface: Participant On-boarding.
• Figure 5: Screenshots from the data collection interface: Chat Interface.