結合神經機器法及中間語法具調校功能之多語言語音翻譯分析

A multilingual speech-to-speech calibration system based on neural machine translation and interlingual translation

張紫文(Tzu-Wen Chang)

指導教授 ： 魯大德

健行科技大學/電資學院/資訊工程系碩士班/碩士(2017年)

摘要

翻譯系統的研發利於旅遊、工作或會議等不同國籍之間的溝通。但現今的翻譯系統多為一對一的翻譯，且須建置多個語言間對應的資料庫，非常龐大且管理複雜度高，現階段的技術大多以結合語音辨識、機器翻譯以及聲音合成來進行，將源語言語音輸入後，經過語音辨識轉換成文字，再將轉換完成的源語言文字，利用機器翻譯進行文本翻譯的動作，最後將翻譯後的目標語言文字經過聲音合成，輸出目標語言的語音而翻譯的方式。
本研究結合神經機器翻譯與中間語法，建置一個一對多的中間語翻譯系統，以減少資料庫的建置，同時探討不同語言的中間語，所翻譯出來的精準度差異性，評價使用bilingual evaluation understudy (BLEU)分數，並使用本實驗所提出二元語法之資料庫比對法進行調教，實驗數據顯示，二次翻譯調校後以英文到日文為最佳，與一次翻譯BLEU分數只相差了1.81，其次是中文到英文，BLEU分數相差了3.53，第三是日文到英文，BLEU分數相差了5.38，中間語為中文的調校源語言為日文的部分提升67.54% ，源語言為英文的部分提升61.45%，而以調校後的中文中間語翻譯為目標語言的部分，目標語言為英文的部分提升22.93%，目標語言為日文的部分提升362.59%。中間語為日文的調校實驗數據顯示，源語言為中文的部分提升70.86% ，源語言為英文的部分提升195.18%，而以調校後的日文中間語翻譯為目標語言的部分，目標語言為英文的部分提升9.81%，目標語言為中文的部分提升11.67%，以實驗三種源語言為例，二次的翻譯較一次翻譯減少1/3的複雜度。

關鍵字

機器翻譯 ； 神經機器翻譯 ； 中間語法翻譯 ； BLEU ； 二元語法之資料庫比對法

並列摘要

The development of the translation system facilitates communication between different nationalities such as travel, work or meetings. One-to-one translation needs to create large numbers of database, is very large and high management complexity. Most of the technology at this time is to combine voice recognition, machine translation and speech synthesis. The source language voice input, after the voice recognition into text, and then used the machine translation to translation the text to target language text. Finally used speech synthesis output target language voice.
In this study, we combined a neural machine translation and interlingual translation to construct a one-to-many multilingual language translation system to reduce a large amounts of database for ont-to-one translation. Bilingual evaluation understudy(BLEU) algorithm used for evaluating the quality of translation. The 2-gram method proposed to calibrate these quality of translations under a threshold after BLEU. Experiment results showed the Chinese or Japanese interlingual languages after BLEU needed for calibration. The Chinese as interlingual language, the performances of source languages from Japanese and English parts after calibrations increased 67.54% and 61.45%, respectively. The performances of target languages English and Japanese parts after calibrations increased 22.93% and 362.59%, respectively. The interlingual languages used Japanese, the source languages from Chinese and English increased 70.86% and 195.18%, respectively. The performances of the target languages English and Chinese increased 9.81% and 11.67%, respectively.

並列關鍵字

machine translation ； neural machine translation ； interlingual machine translation ； BLEU ； database comparison with 2-gram method

參考文獻

[1] 陳鍾誠, “JavaScript (10)–Google 的語音辨識 API 之使用”, http://programmermagazine.github.io/201310/htm/article2.html, 2013

[2] Adam L. Berger, et al., “The Candide system for machine translation”, HLT '94 Proceedings of the workshop on Human Language Technology, pp. 157-162, 1994

[3] Makoto Nagao, “A framework of a mechanical translation between Japanese and English by analogy principle”, Proc. of the international NATO symposium on Artificial and human intelligence, pp. 173 – 180, 1984

[4] “Google Neural Machine Translation”, https://en.wikipedia.org/wiki/Google_Neural_Machine_Translation, 2017/6/26

[5] Norioki Goto, et al., “Domain adaptation of a speech translation system for lectures by utilizing frequently appearing parallel phrases in-domain”, Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA), pp. 1-4, 2016