NEUREM3 Interim Research Report

• http://www.fit.vutbr.cz/research/groups/speech/publi/2022/NEUREM3_research_report_2019-21.pdf PDF

Neural Representations in Multi-modal and Multi-lingual Modeling (NEUREM3) is a project funded by the Czech Science Foundation (GAČR) program Research, Experimental Development and Innovation for the Support of Basic Research Grant Projects EXPRO 2019 from January 2019 till December 2023. This report covers its first three years, 2019--2021. The scientific work was articulated around 5 broad areas: Foundations, Interpretability and task-dependence, Tight integration, Robustness, and Relation of neural representations to multi-lingual concepts. Its description is clustered according to 5 tasks defined in the project proposal and several technical topics. In several of these, we have reached beyond state-of-the-art results. Foreign cooperation in the project was intensive ranging from hiring foreign specialists, through student interns in respected foreign laboratories, to synergies with EU and US projects. The team of the project is consolidated and has a good balance of top-class PI and co-PI, researchers/post-docs, and both Czech and international PhD students. The project is competitive on the international level, the assessment is done via international technology evaluations (challenges), bibliographic metrics, and organization of top international scientific events. NEUREM3 team efficiently cooperates and is at the core of building both Czech and international speech/NLP/MT communities. So far, the project led to a total of 96 publications of which 12 were in peer-reviewed journals, 49 at top conferences, and 35 at local workshops, challenge and evaluation workshops, etc. We are also regularly releasing research data and code in open repositories.