Research On Text Multi-tag Prediction And Question-answering Matching Method Based On Deep Learning

In order to allow users to obtain valuable answers from massive information,intelligent question and answer has always been one of the research hotspots.There are many questions in the Q&A area that need to be studied.Text multi-label prediction and question-answer matching are two of the more critical issues.Prior to the questionand-answer process,multiple tagging techniques were used to first predict the subject tag of the current question and answer,and then to filter the answers that were not related to the subject of the question.This approach can improve Q&A data for the Q&A match phase.The quality improves the accuracy of the returned answer information and improves the experience.This paper uses the deep learning model to carry out text multi-tag prediction and question-answer matching.The main work is as follows:In the text multi-tag prediction,this chapter proposes a multi-label prediction method that integrates the semantic relationship between tags and texts.It mainly integrates the relevant tag libraries and text semantics for the current LSTM to solve multi-label prediction problems.The TBLSTM-TSS multi-label prediction model was constructed by predicting the influence factors of the tag.Without the need to artificially design complex feature engineering,the TBLSTM-TSS model can deepen the semantic relationship between the key information in the relevant tags and texts and the current prediction tags,avoiding the key information when predicting tags for long sequence texts.Semantic weakened or disappeared.Finally,based on the key information is not lost,using the TBLSTM-TSS model itself strong learning ability,improve the accuracy of multi-label prediction.In the case of question-answer matching,this chapter proposes a depth-based question-answer matching method based on attention,which is based on the deep learning model(GRU,CNN),and integrates semantic granular learning of phrase granularity in question sentences and candidate answer sentences.The ATPH-BGRUCNN model.The use of the ATPH-BGRU-CNN model in question-answering matches can increase the weight of the correct answer information in the candidate answer sentence,and avoid the occurrence of interference when the irrelevant information in the candidate answer sentence matches the question and answer.The ATPH-BGRUCNN model needs to extract phrases in sentences.When dealing with the task of phrase segmentation,this paper proposes a phrase sequence model based on BGRU-HS based on traditional methods to improve the accuracy of phrase sequence annotation..In the experimental part,through the comparative experimental analysis with relevant researchers,it is proved that the proposed TBLSTM-TSS multi-label prediction model and ATPH-BGRU-CNN question and answer matching model have higher accuracy and recall than the existing models.