Research And Implementation Of Question Tags Prediction Based On Deep Learning

The tags of question as the most critical factor in the community question answering system,how to predict tags for new or answered questions effectively and efficiently,have become a pressing problem for many community question answering system.In the traditional community question answering system,word segmentation techniques are used to segment the title and description of a question,and the TF-IDF score of each word in the entire question set is counted.Even though this approach computes keyword or frequent tags as tags,it does not make full use of the semantic and structural information of title and description,what's worse,this method is so difficult to find new tags.For the prediction tasks of unanswered and answered questions,this paper utilizes the deep feature extraction capabilities of deep learning to solve the problems of the traditional tag prediction model.Our contributions are summarized below:(1)The existing tag recommendation algorithm of questions ignores the semantics and structural information of questions.This paper makes use of the high-level text feature extraction capabilities of the neural networks to solve the lack of text feature extraction in traditional machine learning algorithms.(2)Recurrent neural networks need to extract feature sequences from the question sequences recursively,so there are performance problems.Convolutional neural networks based text classification model,there will be difficulties in the deep-level network convergence.This paper tries to synthesize the ability for convolutional neural networks in text feature extraction,the ability to speed up the model training of the highway network and the faster running speed of the extreme learning machine,to solve the problems traditional neural networks.Then we submit a deep neural network model that integrates convolutional neural networks,highway networks and extreme learning machine.Experiments on the Zhihu data set verify the validity of the unanswered question tag prediction model.(3)For the tag prediction of an answered question,the existing community question answering systems share the model of unanswered question tag prediction,so it doesn't consider the potential ability of answer to be able to supplement the question.However extracting answer feature requires lager cost of multi-document feature extraction.In addition to the problems of multi-topic clustering and a large amount of computing time needed to solve the problem of multi-document extraction,this paper proposes to use Wilson score algorithm to select the highest confidence answer.The answers obtained based on the highest score in user voting are used to supplement the questions.(4)If the answer with the highest score of the user's vote is used as the auxiliary input for the question tags prediction,this requires us to abstract the answer.The traditional abstraction algorithms are mainly divided into extractive methods and abstractive methods.The traditional s extraction algorithm is mainly divided into abstract and abstract,the focus of extraction is to find important sentences in the text,and the coherence and consistency of the generated abstracts cannot be guaranteed.While,the abstractive methods based on deep learning,such as the attention mechanism,can realize the understanding of texts to a certain extent and can generate highly readable abstracts.Base on the previous work,this paper proposes a question tag prediction model which using multi-head attention model to extract the characteristics of the answer and combines the unanswered question tag prediction model to predict the tags of answered questions.Experiments on the Zhihu data set verify the validity of the answered question tag prediction model.(5)The emphasis researches of the prediction of tags for unanswered and answered question mentioned above,as well as other question tags prediction modules,including question preprocessing technology,web server,data cache and load balancing and model server modules etc.are applied to the practical question tags prediction system.And a Restful service framework based on pre-training model was designed.The practical production verifies the validity of the question tags prediction model base on deep learning proposed in this paper.