A Grammar Correction Assistance System For Non-native Chinese Learner

The Silk Road Economic Belt and 21 st Century Maritime Silk Road have fostered a close relationship between China and its neighboring countries,with more and more foreign friends opting to use Chinese as their second language.However,Chinese has developed over 3000 years unlike other languages in the world,with its complex grammatical structure,confusing morphological similarities and homophones,which poses great challenges to non-native Chinese learners in the world.The task of correcting Chinese grammar errors is a demanding one in the realm of natural language processing,and research into this task can help non-native Chinese learners to master it,which holds both practical and financial importance.(1)In this paper,we design a grammar error correction assistance system for nonnative Chinese learners in response to the lack of grammar software to assist non-native Chinese learners in learning Chinese in the current market.The system supports users to input by voice,then converts the voice input into text for error correction,and finally gives the results back to users.This system can help non-native Chinese learners practice Chinese,find errors in the learning process and assist in correcting them in a timely and effective manner,and greatly reduce the investment in human cost.(2)The grammatical errors generated in the process of using the assisted system mainly consist of two parts: the first part is that when converting speech into text,the presence of a large number of homophones and near-sounding characters in Chinese,the speaker’s speed of speech or non-standard pronunciation,etc,will make the speech recognition technology unable to accurately identify grammatical errors;the second part is that non-native Chinese learners’ grammatical errors are caused by their poor knowledge of The second part is grammatical errors caused by non-native Chinese learners’ poor knowledge.Since the two types of errors are generated in different ways,this paper treats them separately,adds the post-processing process of speech recognition,and adopts the fastcorrect model to correct the grammatical errors in this process.(3)The Chinese grammatical structure is flexible and there are many words with multiple lexical forms,when the same word appears in a sentence with different lexical forms,it represents a different relationship.To address this problem,this paper integrates linguistic knowledge into the model as auxiliary information,improves the word embedding layer without changing the sentence sequence length,proposes three different word embedding methods: full-difference word embedding,word-difference word embedding and lexical-difference word embedding,and combines them with Transformer model,so that the model can focus on the grammatical differences between the source and target utterances when extracting lexical features,and better captures the grammatical differences between sentences.The model can focus on the grammatical differences between source and target utterances when extracting lexical features,and better capture the grammatical features of sentences.(4)Experimental findings demonstrate that the F0.5 value of the model is augmented to varying extents by the three word embedding techniques.Of these,fulldifference word embedding has the most remarkable effect,with an enhancement of2.73% and a 6.27% improvement in the BLEU and F0.5 values respectively.In contrast,the proposed method’s F0.5 value is 0.74 higher than the Transformer model when only the decoder is kept.