Abstract: 【Purpose/significance】To solve the problem of high cost of dataset construction during the build of the automatic question
answering system and the problem of only considering the relevance of the questions and answers in the process of automatic question answering.【Method/process】This paper proposes a dataset construction method of combining annotated question and answer data? base and community question and answer data; constructs a multi layers heterogeneous network model based on keywords of ques? tions-questions-answers-clusters of answers; and proposes an algorithm of automatic question answering based on this model. Obtain library corpus, process and use it as experimental data. Experiments and analysis are carried out with the BERT-Cos, AINN, and BiMPM models as comparison objects.【Result/conclusion】Through experiments, the effects of each model on the library's automatic question answering task are obtained. The model proposed in this paper is superior to other models in all evaluation indicators, and the accuracy rate of the model reaches 87.85%.【Innovation/limitation】The dataset construction method based on multi data source fu? sion and automatic question answering model proposed in this paper have better performance in the question answering task than exist? ing methods; at the same time, suggestions for the length of the user's question are given based on the analysis of the model effect.