Research On Automatic Question Answering Technology Based On Exogenous Knowledge Assistance

With the rapid development of computer technology in the era of big data,a large amount of data is transmitted on the Internet every day.The rich Internet resources enable users to obtain a large amount of information support,which is very convenient in real life.However,when people are faced with such a huge amount of data,how to quickly obtain the required information has become an urgent problem to be solved.The traditional way is to filter data with the information retrieval technology of keyword matching.However,facing the increasing expansion of data volume,the performance improvement is4 not satisfactory.The automatic question answering system models the question on the language model,and digs deep semantic information to find the connection between the question and the answer,and quickly returns the correct answer that matches the question.The automatic question answering system provides a solution for people to efficiently obtain valuable data.The answer selection technology in the automatic question answering system is studied in this subject.The specific task process is as follows: for the multiple candidate answers of a question,the language model is trained by means of supervised learning,so that the model can automatically screen out the correct answer corresponding to the question.The current mainstream solution is to use a Transformer-based Pre-trained model to model the question and answer text,and fine-tune the pre-trained model through the training data of the answer selection task,so that the pre-trained model is suitable for the answer selection task.However,the pre-trained model is limited to the task data in the fine-tuning stage,and cannot obtain additional information to assist the answer selection task.Thus,we try to add exogenous knowledge to the task data to improve the model accuracy.In the process of applying the automatic question answering system to the pre-training model,it was found that the prediction efficiency is low due to the large amount of parameters of the pre-training model.Therefore,a two-stage model is proposed to improve the prediction efficiency of the pre-training model.The main work of this paper is as follows:（1）The Pre-trained model fine-tuning strategy is studied.Aiming at the fact that the single form of task data cannot exert the maximum performance of the Pre-trained model,the exogenous knowledge of the task data is expanded and fused by using information retrieval and knowledge graph,which improves the performance of Pretrained model on answer selection tasks.（2）The method of balancing the performance and efficiency of the Pre-trained model in the answer selection task is studied,and a multi-layer Bi LSTM model based on co-attention is proposed to screen the candidate answers,and the Pre-trained model BERT is used to make the final prediction of the filtered data.The experimental results show that the prediction efficiency of the Pre-trained model is effectively improved when the performance of the Pre-trained model is stable.