ANR JCJC ERIANA
Event-centric Reasoning for Interpreting everydAy NArratives
Making sense of everyday narratives is a very difficult task, requiring a deep understanding of language and a broad knowledge of the world. While a “human reader” can rely on high-level reasoning to draw conclusions, current models largely lack this ability. The majority of existing approaches to language understanding, especially those based on end-to-end neural models, focus primarily on performing low-level (sentence-level) forms of reasoning to accomplish tasks. However, if we are to move forward, we need high-level reasoning capabilities that combine common sense knowledge in an efficient manner. This project aims to develop an approach to high-level reasoning about everyday stories. Reasoning about narratives is event-centered, which inherently relies on event possibilities (aspects or expectations) and interactions (e.g. causality or temporal) between events. We will first develop a neuro-symbolic representation framework for narratives that will allow encoding events as well as their affordances, and the interactions between them in an interpretable way. Based on such a framework, we will then develop deep generative models leading to the discovery of knowledge that can be considered as symbolic rules. Once learned, these rules will be combined using neural models to implement high-level reasoning methods, in the same way that rules are combined in a logic formalism. By engaging in practice throughout the project, several methods, references and applications will be developed.