Research on the framework of low-cost wide-domain Question-Answering system based on knowledge graph

Knowledge graph-based question-answering (KBQA) systems suffer from problems such as low-quality datasets and limited categories of candidate entities, which lead to difficulties in system construction and limited applications. To address this problem, a low-cost wide-domain KBQA system construction framework called LCWD-QA is proposed by combining deep learning, knowledge graphs, and large language models. First, an entity span prediction model was designed to recognize potential entity mentions in sentences. Then, the concept of entity popularity is introduced, and an entity-linking algorithm is designed to link entity mentions to specific entities in the knowledge graph. Finally, the Bert style of text classification models was used for intent recognition, and then different methods were used to generate the answer to that question. In addition, large language models were used to enhance the experimental dataset and generate supplementary answers to the knowledge graph to improve the generalization of the system. The experimental results show that the LCWD-QA system presented in this paper exhibits good performance on both entity span prediction and intent recognition subtasks, with an accuracy of 99.5%, which is better than that of the prevalent benchmark models. The system avoids the high dependence of traditional named entity recognition schemes on manually labeled datasets and has high accuracy and interpretability, which has high application and reference value.