Knowledge Graph Prompting for Complex Domain Question Answering

Large Language Models (LLMs) have achieved widespread success in various Natural Language Processing (NLP) tasks, especially in the Question Answering (QA) task. Despite their impressive performance, they often struggle with issues such as hallucinations and lack of domain knowledge, limiting their ability to accurately reasoning within specialized fields. Knowledge Graphs (KGs), as structured sources of factual knowledge, offer a valuable resource to address these issues. However, existing KG-augmented LLM methods mainly focus on the open-domain multi-hop QA task. Notably, they fail to account for other types of com plex questions, such as comparing and parallel questions. In this paper, we propose a novel method KGPCQA, which introduces a joint fine-tuning task for extracting topic entities and predicting relation paths to enhance domain knowledge understanding of the model. This improves reasoning path retrieval and enables the model to effectively answer do main questions using questions and pruned knowledge. Experiments on the operator field dataset prove that our method outperforms existing advanced approaches, achieving Hit@1 accuracy and F1_score of 95.2 and 90.8, respectively.