Research On The Method Of Composition Recognition And Correction For Chinese Learning

As more and more people begin to learn Chinese,the demand for Chinese teaching is also increasing.Language teaching plays a vital role in the process of language learning.It can help learners better master the language they are learning.By establishing a good language learning system,it is possible to discover problems in the content written by learners in a timely manner,modify grammatical errors in a personalizing manner,and give guidance and suggestions.It is of great significance for improving the efficiency of language learning and teaching.Due to the rise and vigorous development of deep learning,natural language processing has undergone major changes.Grammatical error correction(GEC),as one of the important research areas of natural language processing,has also been affected by this wave of change,attracting a large number of researchers.As a result,many GEC methods based on deep learning have appeared,and then many intelligent corrections systems have been produced.They have greatly improved the efficiency of language learning and teaching.The intelligent correction platform needs to solve the following problems:(1)Since the handwritten composition cannot be directly used as the input of the intelligent correction model,and the processing method of manually inputting it into the system is time-consuming and laborious,it is necessary to automatically recognize it as text before correction;(2)Studies have shown that second language learners with different first languages make different grammatical errors.Modeling these characteristics can effectively improve the effect of GEC models.However,the current deep GEC methods for Chinese do not consider the personalizing characteristics of the Chinese learner;(3)Although many GEC methods based on deep learning have achieved satisfactory results in general domains,they often can not achieve high performance in specific domains due to domain shift and the limited in-domain data.Few work addresses few-shot GEC domain adaptation.To address the above challenges,the main work of this paper is as follows:(1)We implement a text detection model based on border regression and construct a composition image dataset.We carry out the experiments on the constructed dataset.The detection model trained by the composition image dataset can achieve high precision.On the basis of the detection results,we implement a text line recognition model based on contextual information.We train the recognition model using a public dataset and composition text line images.Finally,we achieve the automatic extraction of composition information.(2)We propose personalizing GEC for Chinese as a Second Language(CSL)learners and correct the mistakes made by CSL learners with different first languages.We realize the personalizing GEC by adapting the GEC model to the different domains of CSL learners through the transfer learning method.We cast GEC as a translation task,and correct the grammatical errors by translating an error sentence into a correct sentence.To verify our method,we construct domain adaptation datasets.The experimental results on the domain adaptation datasets demonstrate the rationality and effectiveness of our proposed method.(3)We propose a method based on meta-learning for few-shot GEC domain adaptation to solve the problem of the GEC model quickly adapt to the domain under the condition of few samples.We regard GEC in different domains as different GEC tasks and use a set of resource-rich source tasks to learn the initialization parameters of the GEC model.These initialization parameters can be quickly adapted to new tasks using only a small amount of data.On this basis,we construct a few-shot GEC domain adaptation dataset and carry out experiments.The experiment results on constructed dataset demonstrate that the proposed method can effectively solve the problem of GEC domain adaptation under low-resource conditions.