Neural Network Machine Translation Method Based on Unsupervised Domain Adaptation

Rui Wang,Wei Wang

doi:10.1155/2020/6657344

Abstract

Relying on large-scale parallel corpora, neural machine translation has achieved great success in certain language pairs. However, the acquisition of high-quality parallel corpus is one of the main difficulties in machine translation research. In order to solve this problem, this paper proposes unsupervised domain adaptive neural network machine translation. This method can be trained using only two unrelated monolingual corpora and obtain a good translation result. This article first measures the matching degree of translation rules by adding relevant subject information to the translation rules and dynamically calculating the similarity between each translation rule and the document to be translated during the decoding process. Secondly, through the joint training of multiple training tasks, the source language can learn useful semantic and structural information from the monolingual corpus of a third language that is not parallel to the current two languages during the process of translation into the target language. Experimental results show that better results can be obtained than traditional statistical machine translation.

Highlights

At present, with the gradual deepening of international exchanges, people’s demand for language translation is increasing day by day [1, 2]
The matching degree of translation rules is measured by adding relevant topic information to the translation rules and dynamically calculating the similarity between each translation rule and the document to be translated during the decoding process
Experimental Setup. e experiment selects 10 million single sentences in English, German, and French from the WMT2007 to WMT2010 corpus. e experiment uses Adam as the optimizer, the deactivation rate is set to 0.1, the dimension of the word is set to 512, the maximum sentence length is 175, and sentences with more than 175 words will be intercepted by the superlong part. e training step is 3.5 × 105 and the rest of the model parameters are set to the default parameters of the Transformer model

Summary

Introduction

With the gradual deepening of international exchanges, people’s demand for language translation is increasing day by day [1, 2]. There are so many kinds of languages in the world, and the Internet has become the most convenient platform for obtaining information, and users have an increasingly urgent demand for online translation [3]. On the basis of this work, Imankulova et al [11] generated a pseudoparallel corpus for training through noise reduction and reverse translation and obtained good experimental results. Morente-Molinera et al [13] selected granular information of words and characters in the encoder and used multiple attentions on the decoding side to make information of different granularities collaboratively help translation. Park et al [15] proposed regularization of subwords, using a unary language model to generate multiple candidate subword sequences, enriching the input of the encoder to enhance the robustness of the translation system. Zhang et al [18] used noise-

Objectives

Results

Conclusion