基于DNN与规则学习的机器翻译算法研究

Research on machine translation algorithm based on DNN and rule learning

摘要:

通过以目标信息为指导的卷积体系总结相关源信息,提出了一种系统的处理语言方法。利用在解码过程中使用不同的引导信号,经过特殊设计的卷积+门控体系结构可以查明与预测目标单词相关的源句子部分,并将其与整个源句子的上下文融合在一起形成统一表示形式。研究结果表明,模型将表示形式与目标语言单词一起馈入深度神经网络(DNN),形成更强大的神经网络联合模型(NNJM)。通过两个NIST汉英翻译任务的实验验证,在相同设置下,tagCNN和inCNN在Dep2Str基线上的改善幅度分别为+1.28,+1.75 BLEU,所提出的模型分别优于NIST MT04和MT05的平均值+0.36,+0.83 BLEU,比传统DNN机器翻译平均提高了+1.08 BLEU点。模型为统计机器翻译研究提供了新思路。

Abstract:

Based on the convolution system guided by the target information, this paper summarizes the relevant source information and proposes a systematic processing language method. By using different guide signals in the decoding process, the specially designed convolution + gating architecture can identify the source sentence part related to the predicted target word, and fuse it with the context of the whole source sentence to form a unified representation. The results show that the model feeds the representation and the target language words into DNN to form a stronger neural network joint model (NNJM). The experimental results of two NIST Chinese-English translation tasks show that under the same settings, the improvement of tagCNN and inCNN on the Dep2str baseline is + 1.28 and + 1.75 BLEU, respectively. The proposed model is superior to the average of NIST MT04 and MT05 + 0.36 and + 0.83 BLEU, respectively, which is + 1.08 BLEU higher than the traditional DNN machine translation. The model provides a new way for statistical machine translation research.