Reasoning and Structured Explanations in Natural Language via Analogical and Neural Learning

Neves Ribeiro, Danilo

doi:https://doi.org/10.21985/n2-88hy-5139

Work

Reasoning and Structured Explanations in Natural Language via Analogical and Neural Learning

Public

Download PDF

Download All Files (.zip)

Performing complex reasoning has been a long-standing challenge in artificial intelligence (AI).This thesis describes a class of AI systems designed to reason, extract knowledge, and answer questions on various domains such as process understanding, elementary science, and math word problems. Our approach differs from traditional logical reasoning systems since we work directly over the natural language description of problems, thus bypassing the need to manually create formal representations. The proposed systems rely on two architectures: the Companion Cognitive Architecture and Transformer Language Models. At their core, these architectures use analogical and neural learning to reason and extract patterns from the input text. Experiments on multiple language tasks show that our methods outperform strong baselines and can make predictions from just a few training examples. In addition, we study how such AI systems can generate chains of reasoning that explainhow known facts can be used to reach conclusions and answer questions. Called structured explanations, these chains of reasoning contain multi-premise textual entailments, where intermediate conclusions are used by subsequent entailment steps. This reasoning and explanation approach differs from existing work on natural language inference or multi-hop question answering, where reasoning is either single-step or single-premise. In particular, these structured explanations are shown to alleviate opaqueness issues of neural language models and have the potential to help humans validate and gain more trust in the output of these systems.

Creator