Implementation Of FAQ Question And Answer System Based On Deep Learning

The traditional search-based question-and-answer system,by matching the literal similarity,can quickly retrieve the corresponding content,but the accuracy is not high.The question-and-answer system based on deep learning can effectively solve the problem of accuracy.However,the training phase usually relies on a large amount of data resources,and the calculation requires more calculation time.This paper proposes a FAQ question and answer system with text matching as the core.The system consists of text retrieval and text matching.The text retrieval part reduces the amount of data calculation in the system time and text matching stage through rapid screening.The text matching part adopts a two-stage training method of pre-training and fine-tuning.Pre-training is done by the latest BERT model in the field of natural language processing,using a small amount of data and computing resources to achieve higher accuracy.In order to further improve the accuracy of the matching,the following improvements were made during the fine tuning phase.Firstly,based on the BERT model,the convolutional neural network incorporating the Attention mechanism is used to extract the local features and distinguish the importance.On the test data set,better results than the BERT model are obtained.Secondly,based on the above model,the collaborative training algorithm Tri-Training is introduced to improve the semantic ability of the model as a whole through the differential learning of multiple classifiers.Further,the Tri-Training training algorithm brings noise to the classifier while improving the performance of the model.To reduce the influence of noise,a dynamically adjustable probability threshold is added to each classifier.Combined with the above methods,the final model achieved better results in the objective evaluation criteria.Compared to the baseline model BERT,overall performance increased by 1.5% on the QQP data set and by 1.9% on the IQA data set.