CEDA-TQA: Context Enhancement and Domain Adaptation Method for Textbook QA Based on LLM and RAG

Textbook Question Answering (TQA) is a specific area of Natural Language Processing (NLP) that focuses on processing and understanding the content of textbooks to generate coherent and accurate responses. Textbooks contain various types of information, such as text, diagrams, and formulas, which require the model to have abilities in reasoning, analyzing, and synthesizing information. The training of Large Language Models involves extensive corpora, enabling them to effectively process and comprehend intricate linguistic phenomena, thereby generating coherent and insightful responses. This paper utilizes instructional materials as supporting information to assist LLM in responding to questions, then fine-tunes the LLM to enhance its adaptability to specific disciplines while utilizing the RAG technique to minimize interference from irrelevant content. We conducted ablation experiments on a medium-sized LLM that demonstrated the highest accuracy of 81.91% when both finetuning and RAG techniques were employed.