Partners: University of Avignon, FHG, PRIBERAM, DW, IMCS
SELMA builds a continuous deep learning multilingual media platform using extreme analytics. Large amounts of multilingual text and speech data are available in the internet, but the potential to fully take advantage of this data has remained largely untapped. Recent advances in deep learning and transfer learning have opened the door to new possibilities – in particular integrating knowledge from these large unannotated datasets into plugable models for tackling machine learning tasks. The aim of the Stream Learning for Multilingual Knowledge Transfer (SELMA) is to address three tasks: ingest large amounts of data and continuously train machine learning models for several natural language tasks; monitor these data streams using such models to improve multilingual Media Monitoring (use case 1); and improve the task of multilingual News Content Production (use case 2), thereby closing the loop between content monitoring and production. SELMA has eight goals: 1. Enable processing of massive video and text data streams in a distributed and scalable fashion 2. Develop new methods for training unsupervised deep learning language models in 30 languages 3. Enable knowledge transfer across tasks and languages, supporting low-resourced languages 4. Develop novel data analytics methods and visualizations to facilitate the media monitoring decision-making process 5. Develop an open-source platform to optimize multilingual content production in 30 languages 6. Fine-tune deep learning models from user feedback, reducing recurring errors 7. Ensure a sustainable exploitation of the SELMA platform 8. Encourage active user involvement in the platform. Achieving these aims requires advancing the state of the art in multiple technologies (transfer learning, language modelling, speech recognition, machine translation, summarization, speech synthesis, named entity linking, learning from user feedback), while building upon previous project results and existing services.