D4.3 Tools and Models for Multimodal Multilingual and Discourse-Aware Machine Translation

In this deliverable, we report on our final releases of machine translation models and tools that were developed through our efforts in WP4.

We introduce OPUS-MT, a WebSocket-based translation server, and our release of the MeMAD subtitle translation pipeline, both including pre-trained models suitable for general-purpose translation. Next, we introduce and discuss the \texttt{subalign} toolbox, and its key utilities that implement heuristics to convert between plain sentences and SRT-formatted subtitle segments with time codes. In connection with this, we introduce fine-tuned models for subtitle translation with the capability of token alignment for improved synchronisation. Afterwards, we introduce our releases of the MeMAD image caption translation and end-to-end speech translation systems. These systems are based on our work on multimodal machine translation discussed previously in D4.1, evaluated as part of our submissions to the WMT 2018 multimodal translation and IWSLT 2019 speech translation shared tasks, respectively. Furthermore, we also introduce our release of the MeMAD document-level translation models, which were developed through our experiments on discourse-aware machine translation, and evaluated as part of our submission to the WMT 2019 document-level translation shared task. Finally, we also describe our release of a dataset tailored for benchmarking document-level machine translation performance.

All of our software releases are open source with permissive licences of use, and our pre-trained models have been made freely available for download following the guidelines for open access. Our scripts and documentation have been organised into individual repositories located in the common MeMAD Github space, linking to the relevant pre-trained models~(where applicable) hosted in the MeMAD community space on Zenodo. We include additional explanations and usage instructions in this deliverable as necessary.