Effective human communication in social settings is contingent on recognizing the subtle signals encoded in conversational exchange. However, inferring such social signals is challenging for most dialogue systems, especially when faced with a new task or setting. We introduce Social Scaffolds, a rationale-generation framework for generalization in social understanding tasks. Our framework uses LLMs to generate three types of social signals or rationales that reflect the perspectives of the speaker, listener, and the general worldview. We conduct a comprehensive set of experiments spanning 150 cross-task scenarios wherein we first pre-train a model on a given source task (say detecting persuasion strategies), and subsequently deploy it for a target task (say identifying implicit hate speech). Our results show that providing language models with these rationales facilitates conversational understanding in both instruction-tuned and in-context learning settings; we find significant gains when we incorporate the social rationales alongside the utterance text as part of the input. Particularly, rationales modeling the speaker's intentions yield the largest generalization gains (34%) across tasks. Our analysis also reveals that the generated rationales share low similarity with each other and the corresponding utterance, thereby capturing distinct concepts. They are also designed to be task-agnostic such that the rationale category with greatest impact depends on the task. Our framework shows the promise of pragmatics-oriented data augmentation for social understanding and generalization.
Track: Main Track
Keywords: Dialogue understanding, cross-task generalization, machine-generated rationales
TL;DR: We propose a framework that shows the promise of pragmatics-oriented data augmentation for social understanding and generalization.
Abstract:
Submission Number: 8
Loading