Application of Neuro-Symbolic Reasoning in Natural Language Processing

Abstract

Commonsense reasoning is a crucial part of human intelligence. It has been a goal in the study of artificial intelligence to develop machines mimicking commonsense reasoning at par with humans. The deep learning models have advanced in many natural language processing tasks. However, they lack logical reasoning ability. Symbolic artificial intelligence has the potential reasoning ability wherein logic is used to reason efficiently. This led to the rise of neuro-symbolic reasoning, which is an integration of symbolic logic with deep learning. In this paper, we have experimented with the BERT, RoBERTa and GPT-J language models for question answering task to check their reasoning ability and found that they lack the reasoning ability. This paper compares the performance of the BERT, RoBERTa and GPT-J question-answering models with the neuro-symbolic reasoning model using logical neural networks. It is found that the neuro-symbolic reasoning model has the reasoning ability like humans, which infers that the neuro-symbolic reasoning model can effectively handle logical reasoning and is a key to the future of artificial intelligence.