Research On Key Techniques Of Automated Essay Scoring

More and more attention has been paid to English education,and writing is an important part of it.Automated essay scoring can solve many problems of traditional manual marking,such as reducing the workload of teachers,speeding up the feedback of students' writing,improving the fairness of examinations,etc.,which attracts many scholars' research.Automated essay scoring system doesn't really understand composition,but indirectly evaluate essays by building features that can reflect words,sentences,chapters and other scales.Therefore,mining deeper features hidden in the data has a great impact on improving the system effect.A good article always has a special high-level logic and topic structure,in which the actual word and sentence selection and their arrangement serve the high-level structure,so the sentence smoothness of an article can be an important indicator of automated essay scoring.There are a lot of vocabulary,grammar and semantic information in the model essay.By calculating the text matching degree between students' answers and the model essay,there is an important deep reference standard of combining structure and semantic information in the essay scoring.The main contents and innovations of this paper are as follows:(1)We propose a deep sentence smoothness algorithm that fuses multiple features.For the problem that the traditional word vector representation is rough and irrelevant words and sentences cannot be found,we introduce external knowledge to add prior information to the training process,and for the knowledge representation learning of the knowledge base,we can learn the knowledge information of words;using representation of the synonyms of words can help better solve the problem of polysemy;by embedding dependencies,we can better model the grammatical information of text.At the same time,aiming at the problem of high complexity and low efficiency of the traditional way of using similar matrix to model the relationship between sentences,we introduce the selfattention mechanism to investigate the relationship between the current type of sentence vector and all other sentence vectors.Experiments show that our algorithm has a good effect.(2)We propose a text semantic matching algorithm based on graph neural network.Compared with traditional neural networks,graph neural networks have proven to be able to learn more complex and hidden features in many fields.We use graph convolutional networks to model the matching relationship between long texts.Using the idea of divide and conquer,we assign the sentences of the article to each node to form a "concept".We use a network to measure the similarity feature of some sentences in documents as the feature vector representation of nodes.After the training of graph convolution networks,the similarity information of the whole document is obtained.At the same time,we use recursive autoencoder to pre-train sentence vectors,thereby considering more sentence structural features.Experiments verify the effectiveness of our proposed algorithm.(3)We integrate the two deep features of sentence smoothness and text matching into the deep automated essay scoring algorithm.In the LSTM layer and the multi-layer perceptron layer,we fuse the sentence smoothness vector and text matching vector respectively,which has a significant improvement compared with the traditional way.