Research And Design Of A Kind Of Hierarchical Language Model For English Grammar Error Correction

In recent years,with the rapid development,English grammatical error correction field has produced a large number of outstanding academic achievements,but there is no great breakthrough in the design of language model.For example,n-gram is not very targeted for language error correction,the error correction effect of the newly proposed syntactic n-gram model and some kinds of the tree typed language models is not ideal,and neural language models are too complicated.Therefore,the goal of this thesis is to integrate the advantages of the existing common language model,and to study and design a hierarchical language model that can be applied to the English grammar error correction system under the limited resources.In order to achieve better error detection and correction effect than traditional language models,first of all,I refer to several common language models,and propose a kind of hierarchical language model that can overcome the problem of long-distance inter-word interdependence and can preserve semantic information.The language model uses the dependencies between the words in the sentence,decomposes the sentences into different levels of clauses which are highly correlated internally,and the upper and lower clauses are modified and supplemented.Secondly,I train the hierarchical language model,design the corrective decoding algorithm which uses the approximate words as the candidate words,and uses the probability information of the model to correct the decoding.In this way,a general grammar error correction module is implemented for a variety of grammatical errors correction.Then,I use the hierarchical language model to extract the context information of the sentence to be corrected as the features of the classifier which uses the approximate verb-noun collocations of the collocation of the sentence as an alternative collocation set to generate the matching set.In addition,the language model is used to reorder the error correction results of the matching set.In this way,a correction module for English verb-noun collocation errors is implemented.Finally,the error correction effects of this two modules are evaluated.Experiments show that the proposed hierarchical language model has the advantages of stability and simplicity.And using the hierarchical language model,the description of sentences is more accurate.The hierarchical language model can be used directly to construct the decoder to perform error correction for a variety of English grammatical errors,or to extract the contextual characteristics as the features of machine learning in English grammar error correction,or to evaluate the error correction results and reorder the result.In these three uses,the hierarchical language models can get better performance than traditional linear language model.