Are Ellipses important for Machine Translation?

Abstract

The present paper describes an experiment to evaluate the impact of different types of ellipses described in linguistics on the Neural Machine Translation (NMT) process using English-Hindi as source and target language pair. The empirical results show that the state-of-the-art NMT system achieves much poorly on sentences containing ellipses in comparison to the ones that do not. But, more importantly, we note that some ellipses types are harder for NMT in comparison to others, which advocates for a disparate treatment of different ellipses phenomena for Natural Language Processing (NLP) tasks. We also report significant gains in the translation quality when the text supplied to the NMT system is preprocessed with resolved ellipses through the reconstruction of the elided material-- a simple syntactic procedure that can be easily extended for MT of other language pairs.