Translation rescoring through recurrent neural network language models

Abstract

This work is framed into the Statistical Machine Translation field, more specifically into the language modeling challenge. In this area, have classically predominated the n-gram approach, but, in the latest years, different approaches have arisen in order to tackle this problem. One of this approaches is the use of artificial recurrent neural networks, which are supposed to outperform the n-gram language models. The aim of this work is to test empirically these new language models. For doing that, the translation rescoring of three tasks of different complexity has been performed: in first place, the translation problem has been solved by means of the classic n-gram language models. Next, the different translation hypotheses have been rescored through the language models based on neural networks and the results have been compared. This comparison shows that the translations produced by the neural network language models have a better quality in all the experiments: the perplexity of the language models has been lowered and the BLEU score of the translations outputted by the system has yielded higher values with the neural network language model than with the classical n-gram language model.