Research Of Chinese Text Correction Based On Neural Machine Translation

With the booming Chinese economy,more and more foreigners are learning Chinese,but learning Chinese is not easy for them.Therefore,it is increasingly important to detect and correct grammatical errors in Chinese articles written by CSL?Chinese as a Second Language?learners.The effective Chinese Grammar Error Correction?CGEC?system can provide instant feedback to CSL learners and has important value in the learning process.The mainstream CGEC model is based on neural machine translation?NMT?,but it is not optimized for the characteristics of CGEC.Based on the characteristics of the CGEC task,the thesis uses different parameter initialization methods in the embedding layer,and improves the NMT model by using the multi-layer target-attention calculation method in the decoder.The main work of the thesis is as follows:1.Different initialization methods for embedding layer parameters are proposed.In the NMT-based CGEC model,the embedded layer parameters are typically initialized using pre-trained word vectors.However,the language habits of CSL learners and native speakers are not the same,and the pre-training word vectors are usually trained on large-scale Chinese corpus.Therefore,the thesis proposes different embedding layer parameter initialization methods,the encoding uses random initialization,and the decoding uses the pre-training word vector to initialize.The experimental results show that the initialization method improves the error correction performance of the model by 1.48 F_0.5.2.Based on the NMT model,a multi-level target-attention calculation method is proposed.In the CGEC task,the input and output statements use the same language,and the NMT model deals with translation problems between two different languages.It may not be appropriate to use the NMT model directly as a CGEC model,and the input and output statements in the parallel corpus have similar words and syntax structures.Therefore,the thesis improves the NMT model by using a multi-layer target-attention calculation method in the decoding,so that the model can pay attention to the previous information multiple times and multiple weights when predicting the next word.The experimental results show that the NMT model with the calculation method has improved the error correction performance by 1.7 F_0.5.3.All models in the thesis are integrated by using model ensemble techniques.Based on NMT model,the thesis uses a variety of granular training data and a variety of embedded layer parameter initialization methods,and combined to form a variety of different configurations of NMT error correction model.Model ensemble can make full use of the diversity between individual model,comprehensively consider the error correction information of different angles,and improve the error correction performance by 0.86 F_0.5.