Hybridization of Rule-based and Neural machine translation systems:
Rule-Based Machine Translation (RBMT) systems rely on formal linguistic explicit knowledge to identify lexical units and morphological and syntactic structures in the source language and to convert them into the appropriate target language constructions. Their strength is the correctness of their translations due to the high level of control they enable over the generated structures, while their weakness is the rigidity of such structures – which often leads to a lack of fluency in the resulting translation - and the poor lexical selection at transfer time.
On the other hand, Neural Machine Translation (NMT) systems rely on data to learn those structures and how to translate them. Their strength is the use of idiomatic expressions and their ability to resemble human translations, while their weakness is their black-box nature, which prevents control over badly translated structures.
This project aims at combining the strengths of both systems while avoiding their weaknesses by exploring ways of exploiting the large amount of explicit linguistic knowledge in RBMT systems in Neural MT systems.