Research On Braille-chinese Translation Based On Neural Machine Translation Model

Braille-Chinese translation is a key part of Braille text processing,it cannot only effectively promote the communication between visually impaired people and discerning people,but have important significance for Braille promotion and other related works.In recent years,with the development of deep learning,neural machine translation has gotten massive attention due to its excellent performance,BrailleChinese translation based on neural machine translation model has also become a worth studying subject.This thesis proposes strategies to improve the quality of translation and achieve the lightweight of model based on current neural machine translation models,and also designs solutions for the insufficient Braille-Chinese parallel corpus and the incomprehensive evaluation standards.The main contributions can be summarized as follows:(1)In order to improve the quality of Braille-Chinese translation,in view of the problem that the current neural machine translation models are difficult to model semantic relations when translating fine-grained text,this thesis proposes the NeighborTransformer for Braille-Chinese Translation(NTBC).In NTBC,a neighbor module is integrated to the multi-head attention of Transformer.The neighbor module can dynamically affect the attention weight,and improve the model’s ability to capture contextual features and better model semantic relations.Experimental results show that NTBC can generate higher quality translation results compared to current neural machine translation models.(2)For the insufficient Braille-Chinese parallel corpus,this thesis builds a BrailleChinese translation dataset.Based on this dataset,the study of using neural machine translation methods to handle Braille-Chinese translation tasks can be pursued.The experiments in this thesis are based on this dataset.(3)For the problem that BLEU,a widely used evaluation standard for machine translation,cannot comprehensively evaluate the results of Braille-Chinese translation,this thesis proposes a Position-Wise Bilingual Evaluation Understudy(PBEU)for Braille-Chinese translation,which can better evaluate the uniqueness and sequence of translation results.PBEU can be used as a supplement of BLEU to comprehensively evaluate translation quality.(4)In order to build a lightweight model that can maintain high-quality translation results,this thesis starts with solving the problem that the limitations of the Transformer layers when reducing the hidden layer dimensions,and proposes the Lightweight Transformer for Braille-Chinese translation(LTBC)based on the lightweight Transformer.LTBC uses a one-way expansion layer to replace the original layer of Transformer.The one-way expansion layer combines local self-attention layer and global self-attention layer to extract different features vertically,where different layers specialize in different features,therefore the model can achieve better efficiency.Besides,using the high-performance NTBC to perform sequence knowledge distillation on the lightweight LTBC can further improve the performance of LTBC.Experimental results show that LTBC can maintain better performance in lightweight scenarios,and has faster inference speed when computing resources are sufficient.