Reordering Of Question Answering System Based On Deep Learning And Gradient Lifting Tree Algorithm

In today's life and work,the Internet,with its high availability and convenient accessibility,as well as its huge amount of data sources,has taken human to an era of information explosion.Getting the necessary information accurately and quickly from huge information pool becomes more and more important.Q&A system is created based on this demand.Q&A system is widely used in practical fields such as intelligent customer service,shopping recommendation,knowledge quiz,and seeking medical advice etc.When the user enters his or her own question or a keyword,the system will return a piece of knowledge or a sequence of urls or texts.However,many Q&A systems has a problem in results sorting,leading to unsatisfactory sorting of answers,which further causes mismatch with the needs of users,and reducing the user experience and affecting the efficiency of acquiring knowledge from the Internet eventually.The reason why the sorting efficiency is not high enough may be that the feature is not accurate enough,or the feature is not comprehensive enough or the feature weight setting is not reasonable enough.In order to optimize the sorting of Q&A system,in this paper,the characteristics of Q&A system are diversified and studied combing traditional feature engineering and deep learning semantic similarity calculation,aiming to find the intrinsic relationship between the Question and Answer by combining statistics and semantics.After removing the stop words from the questions and answers,then the features are extracted,the traditional feature engineering is established,the features are normalized,the features to eliminate the magnitude effects between the features are filtered and the redundancy features and noise are reduced.Then the text is mapped to a vector,and the deep learning method is used to calculate the semantic similarity between the question and the answer.Then the author combines the two parts as input,and calculate the gain of the feature through the gradient lifting tree learning method to achieve the final optimized sorting result.In this paper the efforts are made to use a variety of short text semantic calculation methods,including traditional methods like word frequency statistics,editing distance,and language models such as word2 vec and bert,in order to improve the accuracy of similarity calculation using a hybrid multi-feature hybrid strategy.Taking into account the recall time of the streamlined Q&A system,the question was identified by an entity and the answers containing the question entity were weighted.The refined sorting results on the public dataset are verified and final sorting effect is proved to be improved.