TriRAG: Enhancing Retrieval-Augmented Generation Method with Triple-Based Knowledge Graphs for Improved Question Answering

Large language models have a wide range of applications, but they still need to be supplemented for specific domain problems. So we proposed the TriRAG method, an innovative enhancement to the traditional Retrieval-Augmented Generation method, which integrates a structured knowledge graph of semantic triples to improve the performance of large language models on multiple-choice question-answering tasks. TriRAG uses the knowledge graph triples obtained from the text, embeds them into vectors, and retains the most useful triples for question-answering tasks by calculating similarities. By replacing the conventional text-based retrieval with a triple-based retrieval approach, TriRAG provides a more precise and efficient way for information retrieval, which enhances the accuracy of responses from large language models in multiple-choice question-answering tasks. We evaluate the efficacy of TriRAG using the Textbook Question Answering dataset, demonstrating improvements over traditional Retrieval-Augmented Generation methods across several leading large language models including Gemma, Llama, and ChatGPT variants. The results of our experiments and ablation studies confirm that our triple-based systems' retrieval accuracy and processing efficiency improve model performance.