Controlled beam search for neural machine translation using subword units leveraging phrase-based statistical machine translation outputs

The decoding phase is a crucial component in machine translation systems, alongside the creation of the model. Beam search is the most commonly used algorithm for decoding in these systems. Statistical machine translation systems, such as phrase-based models, are very efficient and produce translations phrase by phrase. These models perform decoding locally. Recently, neural machine translation systems have become increasingly popular due to their ability to produce fluent translations. However, neural machine translation systems can sacrifice accuracy for fluency during decoding. The beam search method used by neural machine translation can lead to errors if an incorrect word is selected. This research aims to improve the quality of neural machine translation by controlling the beam search during decoding with the output sentences of phrase-based statistical machine translation systems that use a subword approach. The proposed method guides the decoding of NMT by incorporating the results of SMT into the beam search process through a constrained decoding mechanism that utilizes phrase-based suggestions. Our approach significantly improves the quality of neural machine translation as demonstrated in experiments with two-way translations (German-English and English-German). Experimental results show that our method improves BLEU and METEOR scores on average by up to 0.6 BLEU and 0.5 METEOR for different approaches and by 0.4 BLEU and 0.3 METEOR over the baseline NMT system.