| Myanmar is one of the key nodes along the"The Belt and Road".It is also an important hub for China to connect with Southeast Asia.There is a huge space for cooperation and communication between China and Myanmar in economic,political and cultural fields.The study of machine translation in the two languages is of great significance.However,the development of machine translation in China and Myanmar depends on the parallel corpus.At present,machine translation achieves extremely high performance in language pairs rich in corpus resources,but not mature enough with low-resources language,so studying how to construct a parallel corpus with low-resources is of great significance for further machine translation work in the future.At present,Chinese and English resources are abundant,and there are many news websites on the Internet,and it is very easy to obtain a large number of Chinese-English,English-Myanmar parallel corpora.Therefore,this paper uses English as an intermediate language to bridge the semantic space of Chinese and Burmese bilinguals to complete the construction of the parallel corpus.（1）Contruction for English-Myanmar Corpuswe use the massive resources on the Internet to collect English-Myanmar translation news websites.Then,we analyze its page structure characteristics,crawl English news and Burmese news,and construct English-Myanmar parallel corpus.At the same time,this paper constructs a comparable English-Myanmar corpus to support the research on extraction methods of English-Myanmar parallel sentences,so as to expand the English-Myanmar parallel corpus,This paper also constructs a monolingual corpus of Myanmar language to support the study of the construction of the public semantic space of the Chinese-English-Myanmar trilingual.Since the corpus of the Chinese-English parallel corpus is relatively sufficient,we use the UN corpus[63].（2）English-Myanmar sentence extraction based on Siamese frameworkThe Siamase framework is composed of two parts:one is the combination of Bi-directional Long Short-Term Memory（Bi-LSTM）and Convolutional Neural Networks（CNN）,the other is the classification layer composed of the full connection layer.Parallel sentence pair extraction is mainly statistical methods to construct parallels which is well-labeled training date.selecting the words mutual translation and sentence length,bilingual mutual translation word,sentence dependent syntax structure and other characteristics extraction classifier,which need a large number of well-labled date to support.This paper attempts to use the Siamese network framework combined with Bi-directional Long Short-Term Memory（Bi-LSTM）that consider the contentual information and long-distance.And Convolutional Neural Networks（CNN）obtain deep semantic representation of sentences.Finally,the English-Myan parallel classifier is constructed by computing semantic similarity.Thus,a large number of English-Burmese parallel corpora can be extracted from the comparable corpus to provide a expand-corpus for further experiments.（3）Chinese-Myanmar sentence extraction based on pivot languageMikolov discovery that concepts with similar semantics in vector spaces of different languages have very similar distributions[23],it is believed that linguistic invariant semantics can be obtained in cross-linguistic spaces.Therefore,parallel sentence pairs get closer in the semantic space and non-parallel pairs get farther away.Correlational Neural Networks（CorrNet）was used to conduct small-scale of Chinese-English,English-Myanmar parallel corpus,which mapped to the same space to achieve the Chinese-Myanmar sentenses pair extraction.（4）Chinese-Myanmar prototype systemCombining the above steps,constructing Chinese-Myanmar classifier prototype system.Finally,the Chinese-Myan parallel corpus classifier is constructed to prototype system.The modules of the system includes a classifier module,a web front page display module,and a web backend calling module.The implementation of this system is based on the open source seq2seq code,a lot of modifications.The system can provide accurate parallel corpus sources for Chinese-Myanmar machine translation and information retrieval. |
