Research And Application Of Key Technology For Automatic Review Of Professional Subjective Questions

The intelligent review plays an important role in assisting education supervision.It helps quantify the quality of human reviews,evaluate the ability of reviewers,and improve the fairness of the examination.By combining the analysis on students'subjective questions and objective questions,we can conduct a fine-grained analysis of the candidate's knowledge points,assist in diagnosing the candidate's professional ability and evaluating the difficulty of the test paper.This paper studied the automatic review of professional subjective questions with reference answers.The main challenges and contributions are as follows:(1)Semantic vector processing of professional vocabulary.Word segmentation and word vector are the basis of text semantic understanding.Because the professional corpus is small,the common Chinese word segjmentation tool and the traditional vector learning method can't handle professional vocabulary well and affect the later semantic analysis.This paper used the general word segmentation tool to get the word segmentation results,and obtained the word vector dataset based on the general corpus.For the professional vocabulary that is not registered in the general corpus,the word segmentation tool was modified,the professional vocabulary was granulated,and the word vector of the smaller unit was used for fusion representation to solve the professional word vector problem.(2)Semantic vector representation based on deep network and interpretability review based on attention mechanism.Aiming at the problem of semantic ambiguity of Chinese vocabulary,this paper used two-way long-short memory cycle neural network to learn the semantic vector of sentences,which can fuse the context information of vocabulary and complete sentence semantics.Because the ways of discussing knowledge points of different candidates' answers were different,this paper introduced the attention mechanism based on the reference answer knowledge point,calculated the semantic relevance of the candidate's answer and the reference answer sentence,reconstructed the candidate's answer sentence vector,and generated the segment vector of the candidate's answer for classification review.The semantic similarity and knowledge point coverage detection were considered,which increased the interpretability of the model.(3)Verifying model with real data sets.For a national-level professional qualification examination,the subjective question candidate dataset with reference answers in the past two years was used to test the model performance from four aspects.Firstly,the influences of different sentence vector and segment vector generation models,and attention mechanism on the performance of the model were compared.Secondly,the reusability of the model was discussed by using targeted trainingstrategies for specific topics and mixed training strategies for different topics.Thirdly,the model was semantically analyzed,and the T-SNE method was used to visually analyze the candidate's answer vector that incorporates the attention mechanism.Finally,comparing the model results with the actual reviewers' results,it shows that the model is comparable to the manual reviewer on specific topics.The above results verify the validity and practicability of the model and show that the model is interpretable.