Application Of Deep Learning In Automatic Grading Of English Tests

With the continuous development of artificial intelligence in the field of education,various methods to improve work efficiency emerge in endlessly,especially the emergence of automatic scoring system,which reduces the workload of teachers.Now it has been widely used in the automatic scoring of objective questions,while the subjective question scoring has always been a research difficulty and hot spot in the field of educational artificial intelligence because of its complexity and difficulty.The workload of English marking is large,and there are two types of subjective questions in English,so it is of great significance to study the automatic scoring of English test.The core problem of automatic scoring algorithm for subjective questions is the study of text similarity,and deep learning is widely used in the field of natural language processing because of its advantages in feature extraction.The text content is represented by feature vector by using deep learning knowledge,and the similarity of text pairs is judged by the distance of vector.In this thesis,when using convolution neural network to measure text similarity,this thesis not only use multi-scale convolution to extract local basic features,but also use dual channels convolution model to extract multi granularity features of text.On this basis,this thesis interact the features extracted by dual channels.When using recurrent neural network model to measure text similarity,this thesis studies the bidirectional GRU model which integrates attention pooling and self-attention.the main work and innovation points of this thesis are as follows:Firstly,multi-scale convolution is used to extract text features,and convolution kernel with different scales is used to obtain rich semantic information of different scales.It solves the problem of semantic loss when a single scale obtains local information.In addition,a dual channel convolution neural network based on word granularity and word granularity is proposed,which makes the deeper information of the article be mined.Two kinds of granularity text representation matrix are used to learn the similarity features of different granularity semantics of the two channels,and the text similarity is judged by the semantic similarity matrix after interaction.Secondly,based on the bidirectional GRU model,self-attention mechanism and attention pooling are integrated into the study.Self-attention mechanism can not only capture long-distance text information,solve the problem of long-distance information loss,but also capture the interaction between different words in the same sentence.Attention pooling can capture the information of interaction between text pairs and obtain the degree of influence between different words through the interaction between texts.When the bidirectional GRU model is used to process the text information of time series,it can obtain the forward and backward semantic information of sentences.Finally,the experimental analysis of the method model proposed in this thesis is carried out,and the effects of convolution neural network and recurrent neural network are compared.There are two kinds of corpora in this thesis,Chinese and English.For these two different corpora,different preprocessing methods are used.The experimental data show that the effect of recurrent neural network in short text data set is better,which proves that the bidirectional network integrated with attention pooling can obtain the interaction between texts and make the text features more comprehensive,and the model effect is better when studying the similarity of text pairs.