Improving Low Resource Machine Translation using Morphological Glosses (Non-archival Extended Abstract)

Improving Low Resource Machine Translation using Morphological Glosses Steven Shearing ssheari1@jhu.edu Christo Kirov ckirov1@jhu.edu Huda Khayrallah huda@jhu.edu David Yarowsky yarowsky@jhu.edu Johns Hopkins University Abstract Low-resource machine translation is a challenging problem, especially when the source lan- guage is morphologically complex. We describe a simple procedure for constructing glosses, or mappings between complex, inflected source-language words and equivalent multi-word English expressions. We demonstrate the utility of glosses, especially compared to entries in bilingual dictionaries, across several data-augmentation strategies designed to mitigate a lack of training data. In our experiments, we achieve improvements of up to 1 BLEU point in a Russian-English translation task and 2.4 BLEU points in a Spanish-English translation task over a strong baseline translation system. 1 Introduction Low-resource machine translation, where only a small amount of parallel data is available be- tween source and target languages, poses a significant challenge. Machine translation systems, especially those based on neural network models, tend to be data-hungry. Highly-inflected source languages further complicate the situation, presenting a significant sparsity problem in low-resource settings. Most possible inflected wordforms are likely to appear only once in the data or not at all. In an effort to improve performance when limited parallel data is available for learning how to translate from a highly inflected source language into English, we experiment with two simple data augmentation strategies—appending and substitution. To alleviate sparsity, we experiment with appending entries from multilingual dictionaries directly to the bitext. We also leverage linguistic knowledge about the morphological grammar of the highly-inflected source language to generate multi-word English glosses. We show that these glosses, which better mimic in-situ translations, are more effective than dictionary entries when appended to the training data. We also employ glosses for a second strategy, directly substituting them in place of complex inflected forms in the source language. The overarching idea is to create a new version of the source, source′, that is more similar to the target language. In theory, this should improve performance by solving some portion of the translation problem before a final translation model is trained. We show that gloss substitution has a positive effect on BLEU scores compared to baseline systems. We present experimental results translating from Russian and Spanish into English. While Russian and Spanish are not low-resource languages, we simulate extremely low-resource sce- narios by relying only on representative language packs from DARPA’s LORELEI (LOw RE- source Languages for Emergent Incidents) program as our source of parallel training data. Proceedings of AMTA 2018, vol. 1: MT Research Track Boston, March 17 - 21, 2018 | Page 132 These packs typically contain less than 50,000 bilingual sentence pairs in total, orders of mag- nitude below the amount used to train most state-of-the-art MT systems. We run our exper- iments using traditional phrase-based statistical machine translation models (PBMT). While neural machine translation offers state-of-the-art performance when training data is plentiful, PBMT remains competitive or superior in the low resource conditions we focus on (Koehn and Knowles, 2017). 2 Multilingual Dictionaries Versus Glosses We …

Content truncated to 500 words