Translating Short Segments with NMT: A Case Study in English-to-Hindi

Translating Short Segments with NMT: A Case Study in English-to-Hindi Shantipriya Parida Ondřej Bojar Charles University, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostranské náměstí 25, 118 00 Prague, Czech Republic {parida,bojar}@ufal.mff.cuni.cz Abstract This paper presents a case study in trans- lating short image captions of the Visual Genome dataset from English into Hindi using out-of-domain data sets of varying size. We experiment with three NMT mod- els: the shallow and deep sequence-to- sequence and the Transformer model as im- plemented in Marian toolkit. Phrase-based Moses serves as the baseline. The results indicate that the Transformer model outperforms others in the large data setting in a number of automatic met- rics and manual evaluation, and it also produces the fewest truncated sentences. Transformer training is however very sen- sitive to the hyperparameters, so it requires more experimenting. The deep sequence- to-sequence model produced more flawless outputs in the small data setting and it was generally more stable, at the cost of more training iterations. 1 Introduction In recent years, neural machine translation (NMT) systems have been gaining more popularity due to their improved accuracy and even more flu- ency compared with “classical” statistical ma- chine translation systems such as phrase-based MT (PBMT), see e.g. the shared tasks of WMT and IWSLT (Bojar et al., 2017; Cettolo et al., 2017). The major advantages of NMT include the consid- eration of the entire sentence, capturing similarity © 2018 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. of words, and the capacity to learn complex rela- tionships between languages. At the same time, it has been observed that NMT is more sensitive to the shortage of or noise in the parallel training data (Koehn and Knowles, 2017). Our goal is to create the Hindi version of Visual Genome (Krishna et al., 2017).1 Hindi, with 260 million speakers, is the fourth most widely spoken language on the planet (after Chinese, Spanish and English). Hindi is a morpho- logically rich language (MRL), with e.g. the gen- der category being reflected in the forms of nouns, verbs and also adjectives (Sreelekha S and Bhat- tacharyya, 2017). The structural and morphologi- cal differences between English and Hindi result in translation difficulties (Tsarfaty et al., 2010). Visual Genome is a dataset of images, captions and relations. As such, it is potentially useful for many NLP and image processing applications. The Hindi version would allow to exploit this dataset e.g. to create Hindi image labellers or other practi- cal tools. The textual part of Visual Genome consists pri- marily of short sentences or noun phrases that were manually attached to rectangular regions in an in- put image. In the current version, Visual Genome contains 108K distinct images with 5.4 million such labelled regions in total. On average, an im- age is thus associated with 50 text segments. Text segments can repeat across images and indeed, when de-duplicated, the set of unique strings re- duces to …

Content truncated to 500 words