In this talk I will discuss our recent work on developing neuro-symbolic models for natural language inference. The model combines distributed representation with natural logic to select and modify the intermediate symbolic reasoning steps in order to discover reward-earning operations using reinforcement learning. The model leverages external knowledge to alleviate spurious reasoning paths and increase training efficiency. The proposed method demonstrates a good capability in monotonicity inference, systematic generalization, and interpretability. I will also discuss how to leverage natural logic to develop adversarial attacks to examine if existing models possess the desired property of reasoning.

Recording: https://drive.google.com/file/d/1WBIR0huvPvXPo4423jffGiKAN-_7kGZh/view?usp=share_link