Towards better translation performance on spoken language

Towards better translation performance on spoken language Chao Bei, Hao Zong Global Tone Communication Technology Co.,Ltd. {beichao,zonghao}@gtcom.com.cn Abstract In this paper, we describe GTCOM’s neural machine transla- tion(NMT) systems for the International Workshop on Spo- ken Language Translation(IWSLT) 2017. We participated in the English-to-Chinese and Chinese-to-English tracks in the small data condition of the bilingual task and the zero-shot condition of the multilingual task. Our systems are based on the encoder-decoder architecture with attention mecha- nism. We build byte pair encoding (BPE) models in paral- lel data and back-translated monolingual training data pro- vided in the small data condition. Other techniques we ex- plored in our system include two deep architectures, layer nomalization, weight normalization and training models with annealing Adam, etc. The official scores of English-to- Chinese, Chinese-to-English are 28.13 and 21.35 on test set 2016 and 28.30 and 22.16 on test set 2017. The official scores on German-to-Dutch, Dutch-to-German, Italian-to- Romanian and Romanian-to-Italian are 19.59, 17.95, 18.62 and 20.39 respectively. 1. Introduction This paper describes the submission of the Global Tone Com- munication Technology Co., Ltd. (GTCOM) for the first participation in IWSLT evaluation. We participated in the zero-shot condition in the multilingual task and the English- to-Chinese and Chinese-to-English tracks in the small data condition of the bilingual task. Our neural machine transla- tion systems are developed as encoder-decoder architecture [1] with attention mechanism [2] and the experiment toolkit we used in the evaluation is Nematus [3]. The intuition of this participation is to verify whether the model architechture and techniques we applied in our generic system 1 with large training data is also effictive in spo- ken language domian with small training data. In bilingual task, since the training data is very small in both Chinese- to-English and English-to-Chinese directions, Chinese word segmentation, tokenization, binary pair encoder(BPE), dif- ferent size of hidden layer, deep transition model and back- translation are involved in our experiments. In multilingual task, we uesd different pre-processing strategy and annealing Adam to enhance the translation performance. This paper is arranged as follows. We firstly describe the 1Our generic translation system covers 10 languages and is available at http://translateport.yeekit.com:4305/index.html task, including the data size and evaluation method. Then we introduce the techniques used in our system. After that, we present the experiments for the two task, including data pre- processing and model architecture. Finally, we analysis the experiment results and draw the conclusions. 2. Task Description The task focuses on bilingual and multilingual text transla- tion in spoken language domain; the provided data is mainly collected form TED talks. We participated in Chinese-to- English and English-to-Chinese directions of the bilingual task, as well as zero-shot translation of the multilingual task. 2.1. Bilingual task For the bilingual task, we focused on Chinese-to-English and English-to-Chinese directions of the small data condition, which only the in-domain training and development data is allowed to use. The detail information about the data is shown in Table 1. In addition, Chinese texts were evaluated at character level. Before evaluation, …

Content truncated to 500 words