Commonsense reasoning about social dynamics in text

Abstract

When humans interact with each other (e.g., having conversations, sharing stories, etc.), they are able to reason more deeply about social implications in order to better understand each other and have more productive interactions. For example, when hearing someone else discuss a personal story, most people are able to think about the consequences of various events, anticipate the feelings of their conversation partner, and respond accordingly. Reasoning about social relationships in text is natural for most people, but is challenging for natural language processing models, in part because these relationships are often subtle, nuanced, and implicit. Training models for this type of inference is additionally challenging due to a lack of designated tasks, resources, and modelling frameworks specifically designed for this type of social commonsense reasoning. We approach this problem by designing new focused tasks and resources specifically aimed towards types of social reasoning. We also introduce new modeling frameworks to learn to integrate social inferences with downstream tasks such as story and dialogue generation. First, we investigate reasoning about social dynamics of characters and actions within stories. We create a new benchmark for reasoning about character mental state based on story events. We demonstrate that this type of reasoning is challenging even for state-of-the-art language understanding models. We also introduce plot dynamics as part of a new modeling framework for story generation. Our results indicate that tracking plot state and integrating discourse features are beneficial for writing tighter narratives. We also explore two types of reasoning about a speaker (e.g., a writer of a piece of text, a conversation partner, or so on) based on what they have said or written. We present connotation frames, a novel formalism for measuring connotative relationships implicit in the text that imply the writer’s underlying message. We create a connotation frames lexicon, which may be useful in tasks like detecting implied stance, bias, or subtle meaning intended by a writer. Lastly, we investigate reasoning about a speaker in the dialogue setting by exploring the challenges of creating empathetic responses to a conversation partner. We introduce the task of empathetic response generation and a new dataset for training dialogue models to generate responses that are more empathetic and socially aware of a conversation partner’s feelings.