基于深度学习的英语翻译机器人错误文本检测系统研究

Research on Error Text Detection System of English Translation Robot Based on Deep Learning

针对传统英语翻译机器人在错误文本检测中存在语法纠错准确率低的问题,以英语翻译机器人错误检测系统为研究对象,提出构建一个双编码器的语法错误纠正模型。此模型分别采用Transformer编码器和Bi-GRU编码器对句子上下文信息和源句进行特征提取;之后利用解码器端的门控结构对输入的相关特征信息进行整合,由此实现各个特征与注意力机制的适配。实验结果表明,在CoNLL-2014数据集中,本模型的精确率、召回率和F0.5分别取值为85.42%、42.31%和75.33%,均高于传统的Nested-GRU*模型和SMT+NMT+FST-LM混合模型,本模型的F0.5值比前两者模型分别高出了32.97%和14.62%。且本模型的GLEU值取值为85.93%,超出前两种模型39.14%和25.62%。由此可知,本模型可实现英语翻译机器人语法错误准确检测和纠正,语法纠错精确率高达81.08%,可在英语错误文本检测系统中进行应用。

Based the English translation robot error detection system, a grammar error correction model with dual encoder is proposed. Transformer encoder and Bi-GRU encoder is used to extract sentence context information and source sentence features. Later, the decoder gate structure integrates the input related feature information to realize the adaptation of each feature and attention mechanism. Experimental results show that the exact rate, recall and F0.5 values are 85.42% in the CoNLL-2014 data set, respectively, which are higher than the traditional Nested-GRU * model and SMT + NMT + FST-LM hybrid model. The F0.5 values are 32.97% and 14.62% higher, respectively. Moreover, the GLEU value of this model is 85.93%, which exceeds 39.14% and 25.62% in the first two models. Therefore, this model can accurately detect and correct the English translation robot grammar errors, and the grammar error correction accuracy rate is as high as 81.08%, which can be applied in the English error text detection system.