The Method Of Question Semantic Matching For Retrieval-based Question Answering

With the development of the Internet,more and more enterprises are beginning to provide services to users through the Internet.As the number of users grows,the demand for online customer service is also growing.It is difficult to provide roundthe-clock service,and there are a lot of repetitive problems that users ask.Combining the FAQ database and using the retrieval-based question answering(QA)technology as an assistant to the customer service staff can free the customer service staff from a lot of repetitive work and have time to solve problems with more value.Therefore,retrieval-based QA technology has great practical value.Based on the characteristics of retrieval-based QA,this thesis studies the question matching technology in retrieval-based QA to improve the efficiency and accuracy of question matching.The main contents of this thesis are as follows:Compression and knowledge distillation of BERT model for text semantic matching.This thesis experiments with some existing semantic matching methods.Results show that the BERT model outperforms all other models in accuracy,but BERT is also much slower and heavier.The high requirement of the BERT model for computing resources limits its use in practical applications.In order to reduce the BERT model's computational requirements,this thesis experiments several methods to simplify the BERT model to obtain a lightweight student model.Then we introduce a BERT model compression method on text semantic matching task.This method combines knowledge distillation and data augmentation and is used to transfer the knowledge of the BERT model to the pruned BERT model with only four layers of Transformer structure.The accuracy of the pruned BERT model on question semantic matching task is close to that of the BERT model,and the inference speed is three times that of the BERT model.Text semantic matching methods based on the BERT model.This thesis analyzes the BERT model on text semantic matching task.Combining the BERT model with several semantic matching paradigms we propose the BERT-interaction model.Experiments show that the performance of the representation-based model with BERT as the text encoder outperforms the other representation-based model and not much worse than BERT model.The BERT-interaction model with BERT as the text encoder and simple interaction on top of that can achieve comparable performance as the BERT model.The model's inference speed when using the cache mechanism to sort candidate questions in a retrieval-based QA scenario is significantly improved compared to the BERT model.A variant of the model can improve the speed of retrieval-based QA and the accuracy of question semantic matching compared to the BERT model.The implementation of the QA platform and the collection of the corpus.This thesis built a QA system that supports administrators flexibly configure various corpus and provides multi-scenario user status management and retrieval-based QA functions.The semantic matching model in the system trained on instance-based selected LCQMC corpus has an MRR of 0.7496 on the system's FAQ test set.