G2DiaR: Enhancing Commonsense Reasoning of LLMs with Graph-to-Dialogue & Reasoning

Knowledge-rich dialogues are crucial for assessing and enhancing the reasoning capabilities of Large Language Models (LLMs). Detailed knowledge hints are incorporated in prompts to generate knowledge-rich dialogues. However, the intricate interplay among dialogue, context, and diverse knowledge facets is often overlooked. This causes LLMs to exhibit basic commonsense understanding but struggle with commonsense reasoning. To tackle these challenges, we proposed a graph-to-dialogue & reasoning generation method G2DiaR. G2DiaR can controllably convert the knowledge path in the graph into a dialogue, and can also generate reasoning that is faithful to the dialogue based on the neighbor nodes of the path nodes. We leverage G2DiaR to create G2DiaD, a challenging dataset for dialogue-based reasoning in multiple-choice question-answering (MCQA), derived from the ATOMIC commonsense graph. G2DiaD comprises 6,492 QA pairs, each offering inference-based answers supported by or conflicting with reasoning. Experimental results demonstrate that training on G2DiaD leads to an average performance improvement of 6.08% across various commonsense reasoning benchmarks, surpassing retrieval methods (0.60% improvement) and training on the CICERO dataset (−4.78% decline). G2DiaR can generate dialogue data rich in commonsense, relation and reasoning, enhancing the reasoning and generalization capabilities of LLMs.