Research On Automatic Scoring Method Of Text Subjective Homework

With the rapid development of artificial intelligence,automatic scoring technology has been widely studied and applied.Machine review not only reduces the impact of subjective and objective factors such as review environment,time and reviewers' psychological activities,but also reduces the time cost and economic cost of manual review.Text subjective homework is helpful to test students' comprehensive mastery of knowledge and cultivate students' innovation ability,so it occupies a large proportion in the teaching process of colleges and universities.At present,text subjective homework is mainly evaluated manually,so the automatic scoring technology remains to be further studied.Considering the characteristics of text subjective homework assignments and the past scoring experience of reviewers,this thesis takes the assignments of software engineering and curriculum design in our university as the experimental subjects.The specific research contents are as follows:(1)In-depth analysis of the scoring rules of subjective type of text homework is carried out,and the Automatic Scoring Model(ASM)is constructed.The ASM model is composed of automatic scoring module and performance intelligent analysis module.The automatic evaluation module realizes the evaluation from four dimensions: similarity index,topic fit index,workload index and text structure index.The specific evaluation ideas are as follows:Firstly,the text features are extracted by statistical methods to achieve the scores of workload indicators and text structure indicators.Then,the similarity algorithm and Text Rank algorithm are integrated to quantify the text content,and the results of similarity index and topic fit index are calculated.Finally,according to the different types of operations,the organization of applicable evaluation indicators and flexible adjustment of evaluation parameters to achieve a key scoring function.The score intelligent analysis module includes the score statistics module,students' personal ability analysis,human-computer scoring comparison and other links.The chart can make the reviewer more intuitive to understand the completion of students' homework.At the same time,in order to make up for the lack of machine scoring,the function of online browsing and performance modification is added.It is integrated into an intelligent auxiliary scoring system for subjective text type tasks,which integrates job management,automatic scoring and performance statistical analysis.(2)The evaluation principle of topic fit indicators is to propose a variety of calculation methods to improve the flexibility and universality of the model.A balanced compensation-Graph Attention Network(BC-GAT)based on GAT is proposed to solve the problem that the topic classification effect of GAT model is not good for small sample imbalanced data.Its core idea is to make balanced compensation for small-scale samples in the data set and optimize the construction method of GAT model.In this thesis,through the reasonable organization of EDA algorithm and web crawler algorithm,the expansion of small proportion samples in the data set is more reasonable and efficient,and the GAT model is more suitable for small sample imbalanced topic classification.The experiments on public and real data sets show that the positive and negative sample recognition accuracy of BC-GAT method is significantly improved compared with that of GAT model,which can effectively solve the problem of minimum sample and data tilt in practical application.