Data Augmentation by Adjunct Deletion for Neural Machine Translation
Hakimeh Fadaei, Heshaam Faili · 2018
Neural machine translation is shown to be very successful for translation between many language pairs. The experiments have shown that the quality of NMT largely depends on the amount of training data and parallel corpora with large sizes are unavailable for many language pairs. To be able to use this approach for low resource language pairs, generating parallel data for these languages is of great importance. In this paper we introduce an approach for augmenting the size of the parallel corpus. In our work we generate new parallel sentences by deleting the optional constituents from the source and the target sentences in the base parallel corpus. To detect these optional constituents, we consult the Tree Adjoining Grammar (TAG) derivation tree of the source sentence. Adjoined words in these trees are good candidates for deletion. To evaluate the proposed method, we conducted our experiments on English-Persian translation task. The results show that augmenting the parallel corpus with adjunct deletion improves the quality of translation 3.2 and 0.8 BLEU points when using the training sets of 1.6M and 3.2M words respectively.