Improving Arabic neural machine translation via n-best list re-ranking

Abstract

Even though the rise of the neural machine translation (NMT) paradigm has brought a great deal of improvement to the field of machine translation (MT), the current translation results are still not perfect. One of the main reasons for this imperfection is the decoding task complexity. Indeed, the problem of finding the one best translation from the space of all possible translations was and still is a challenging problem. One of the most successful ways to address it is via n-best list re-ranking which attempts to reorder the n-best decoder translations according to some defined features. In this paper, we propose a set of new re-ranking features that can be extracted directly from the parallel corpus without needing any external tools. The feature set that we propose takes into account lexical, syntactic, and even semantic aspects of the n-best list translations. We also present a method for feature weights optimization that uses a quantum-behaved particle swarm optimization algorithm. Our system has been evaluated on multiple English-to-Arabic and Arabic-to-English MT test sets, and the obtained re-ranking results yield noticeable improvements over the baseline NMT systems.