Research And Implementation Of College Enrollment Question And Answer Service System Based On Deep Learning

At present,in the consultation work of college admissions,with the increase in the number of consultants and the different business processing procedures of various departments,to a certain extent,it causes heavy workload and management pressure.So,how to provide professional,reliable and real-time help to the students in a timely manner is a problem that needs to be solved urgently in the enrollment scenario.Therefore,this article uses college admissions as an application scenario to build an intelligent enrollment question and answer service system that will help staff answer questions for students.This can not only improve work efficiency and satisfaction,but also improve medical efficiency and satisfaction.Banking and other similar demand scenarios also have a positive reference.The research goal of this paper is to build an intelligent question answering system for the limited field of college admissions.The system consists of question understanding module,question retrieval module and answer generation module.The key research contents involved include text preprocessing,text vectorization,text classification,text similarity calculation,answer confidence calculation.For the system,how to improve the ability of the system question understanding module to belong to the category of query sentences;how to improve the ability of the system question retrieval module to measure the semantic equivalence;how to improve the semantic matching of the query sentence and the candidate answer set in the answer generation module,These are the research focuses of this article.Therefore,the main research work and achievements of this paper are: constructing a CNN_BiGRU problem classification model based on part-of-speech features;constructing a question similarity model based on attention k-max pooling BiLSTM network;constructing a Answer confidence model based on attention_BiGRU interactive matching.(1)Constructing a classification model CNN_BiGRU based on part-of-speech features.The model includes text vectorization technology,CNN,BiGRU.Among them,it is proposed to introduce part-of-speech attributes to assist the vectorized text in view of the shortcomings of the traditional vector method to distinguish the polysemy in the paragraph text.In addition,in view of the fact that CNN and BiGRU can capture higher-order features of different granularities,it is proposed to combine CNN and BiGRU to form CNN_BiGRU.The model extracts information at different granularity levels to obtain higher-order representations with different semantics,thereby enhancing classification capabilities.(2)Constructing a question similarity model based on attention k-max pooling BiLSTM network.The model consists of CNN convolution,attention pooling,k-max pooling,and BiLSTM network.The traditional similarity algorithm learns shallow semantic information,resulting in poor semantic similarity.Therefore,a new similarity model is constructed.Among them,for the CNN convolution,the contribution of the convolution representation cannot be quantified.Therefore,combining the attention pooling and k-max pooling to focus on the convolution representation,and then,using the BiLSTM bidirectional encoding mechanism to focus and capture the representation Potential information in order to generate advanced features.Experimental analysis shows that the model can pay more attention to and obtain the corresponding level of semantic representation,which effectively improves the accuracy of semantic similarity.(3)Constructing an answer confidence model based on attention_BiGRU interactive matching.This model draws on the Siamese Network structural framework and interactive matching mode,and applies the attention mechanism and BiLSTM network.For semantic matching tasks,if parallel matching is used,the semantic granularity of the matching will be relatively shallow.Therefore,interactive matching is used to focus on text interaction information at different matching levels and extract deep interactive semantics.Based on the experimental results,the model performs well on evaluation indicators,proving excellent semantic matching capabilities.