Robust Neural Machine Translation with Joint Textual and Phonetic Embedding

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3044–3049 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3044 Robust Neural Machine Translation with Joint Textual and Phonetic Embedding Hairong Liu1 Mingbo Ma1,3 Liang Huang1,3 Hao Xiong2 Zhongjun He2 1Baidu Research, Sunnyvale, CA, USA 2Baidu, Inc., Beijing, China 3Oregon State University, Corvallis, OR, USA {liuhairong, mingboma, lianghuang, xionghao05, hezhongjun}@baidu.com Abstract Neural machine translation (NMT) is notori- ously sensitive to noises, but noises are al- most inevitable in practice. One special kind of noise is the homophone noise, where words are replaced by other words with similar pro- nunciations.1 We propose to improve the ro- bustness of NMT to homophone noises by 1) jointly embedding both textual and phonetic information of source sentences, and 2) aug- menting the training dataset with homophone noises. Interestingly, to achieve better trans- lation quality and more robustness, we found that most (though not all) weights should be put on the phonetic rather than textual infor- mation. Experiments show that our method not only significantly improves the robustness of NMT to homophone noises, but also sur- prisingly improves the translation quality on some clean test sets. 1 Introduction Recently we witnessed tremendous progresses in the field of neural machine translation (NMT) (Sutskever et al., 2014; Cho et al., 2014; Bahdanau et al., 2014; Luong et al., 2015; Gehring et al., 2017), especially the birth of transformer network (Vaswani et al., 2017). Despite tremendous success, NMT models are very sensitive to the noises in input sentences (Be- linkov and Bisk, 2017). The causes of such vul- nerability are multifold, and some of them are: 1) neural networks are inherently sensitive to noises, such as adversarial examples (Goodfellow et al., 2014; Szegedy et al., 2013), 2) every input word can affect every output word generated by the de- coder due to the global effects of attention, and 3) all NMT models have an input embedding layer, which is sensitive to noises in the input sentences. 1In this paper, the word “homophone” is loosely used to represent characters or words with similar pronunciations. In this paper, we focus on homophone noise, where words are replaced by other words with similar pronunciations, which is common in real- world systems. One example is speech transla- tion (Ruiz et al., 2017; Ruiz and Federico, 2015; Ma et al., 2018), where an ASR system may out- put correct or almost correct phoneme sequences, but transcribe some words into their homophones. Another example is pronunciation-based input systems for non-phonetic writing systems such as Pinyin for Chinese or Katakana/Hiragana for Japanese. It is very common for a user to acciden- tally choose a homophone instead of the correct word. Existing NMT systems are very sensitive to homophone noises, and Table 1 illustrates such an example. The transformer model can correctly translate the clean input sentence; however, when one Mandarin character, ‘有’(yˇou), is replaced by one of its homophones, ‘又’(y`ou), the transformer generates a strange and irrelevant …

Content truncated to 500 words