Machine Translation System Combination with Enhanced Alignments Using Word Embeddings

Abstract

Machine Translation (MT) is a challenging problem and various techniques proposed for MT have their own strengths and weaknesses. Combining various MT systems has shown promising results. Confusion network decoding is one such approach. In this work, we propose using word embeddings for aligning words from different hypotheses during confusion network generation. Our experiments, on English-Hindi language pair, have shown statistically significant improvement in BLEU scores, when compared to the baseline system combination. Four data-driven MT systems are combined, namely, a phrase based MT, hierarchical-phrase based MT, bi-directional recurrent neural network MT and transformer based MT. All of these have been trained on IIT Bombay English-Hindi parallel corpus.