Multi-hop commonsense knowledge injection framework for zero-shot commonsense question answering

Abstract

Zero-shot commonsense question answering (QA) task is to evaluate the general reasoning ability of the language model without training on the specific datasets. The existing zero-shot framework transforms triples within the commonsense knowledge graphs (KGs) into QA-format samples, serving as a pre-training data source to integrate commonsense knowledge into the language model. However, this approach still faces the following challenges: 1) The model trained from synthetic QA generated from triples lacks the multi-hop commonsense knowledge required for handling complex QA problems. 2) Ambiguity caused by confusing commonsense knowledge within synthetic QA, making it challenging for models to discern semantically similar entities. To address the above problem, we propose a novel Multi-hop Commonsense Knowledge Injection Framework (MCKI). Specifically, we draw inspiration from human complex reasoning thinking and further propose a synthetic multi-hop commonsense QA generation method. Meanwhile, we introduce negative samples with high confusion in synthetic QA, and then use contrastive learning to improve the model’s ability to distinguish similar commonsense knowledge. Extensive experiments on five commonsense question answering benchmarks demonstrate that our framework achieves state-of-the-art performance, surpassing existing methods, including large language models like GPT3.5 and ChatGPT.