The last twelve months of the MMT project were rather intense and turbulent. We started the year still believing that phrase-based MT (PBMT) could be competitive with Neural MT (NMT) in the realm of technical translation and, in particular, when a single system has to handle translation requests from many different domains. Thus, during Q1 we invested significant effort in finalising our PBMT implementation and released the first MMT plug-in for a commercial CAT tool (Trados Studio). In parallel, we continued investigating an original NMT adaptive solution developed at FBK that seemed to fit pretty well the use cases of the project and, more importantly, to overcome the limitations of NMT under the multi-domain setting. Finally, the imminent finalisation of the MMT PBMT software gave also impetus to a large-scale collection of training data for 7 language pairs (14 translation directions) and to the development of a large-scale multi-domain evaluation benchmark.
During Q2 we presented the MMT PBMT solution to several large companies, which later started to run tests with the MMT open source code, either independently or with our support. During Q2, in accordance with the work plan, FBK spun off a newco, MMT Srl, which was soon joined by Translated in the role of investor.
By the end of Q2, however, empirical evidence convinced us that NMT would outplace PBMT much quicker than we thought. From Q3, we started to completely redesign the MMT architecture and to rewrite and replace most of its core components. In just a few months of very intensive work, the MMT team was able to release the first real-time adaptive NMT open source software and to complete the implementation of a plugin for the MateCat tool. Hence, 14 translation engines for seven language pairs were trained and put into production with the plugin. The new software was then widely disseminated and showcased at several public venues and companies.
In Q4, evaluation activities of the new NMT architecture took off. Fast development of new and better baseline neural systems also called for an effective and simple human evaluation protocol to rapidly monitor progress. This period was very intense but rewarding as we e achieved significant enhancements of the NMT baselines. Until the very last days of 2017, our work has focused on improving the 14 baseline NMT systems as well as the MMT open source code, which thanks to feedback from several users has been relentlessly improved by fixing bugs and improving its functions and performance.
