The existing deep learning models for natural language processing (NLP), while sometimes impressive, are often unreliable and even misleading: they do not generalize well to new domains, they do not exploit contextual information, they are poorly calibrated, and their memory is not traceable. These limitations stem from their monolithic architectures, good for perception, but unsuitable for tasks that require higher-level cognition.
With the DECOLLAGE project (DEep COgnition Learning for LAnguage GEneration), André Martins intends to present solutions to some of the fundamental problems of NLP, by using an innovative interdisciplinary methodology that brings together tools from artificial intelligence, sparse modeling, neuroscience, and cognitive sciences.
First, uncertainty and quality estimates will be used as a guiding principle for controlled generation, combining this control mechanism with the efficient encoding of contextual information and the integration of multiple modalities. Second, sparse and structured models of long-term memory will be developed, together with attention to descriptive representations. Third, new mathematical models for sparse communication will be devised (bridging the gap between discrete and continuous representations), supporting end-to-end differentiability and enabling a shared workspace where multiple modules and agents can communicate. These innovations will be applied to highly challenging generation tasks, including machine translation, dialogue, and open-ended generation.
The researcher expects that this project will make it possible to “advance one step further in overcoming the limitations of current NLP technologies, making it possible for humans and machines to communicate effectively in natural language and to work collaboratively to solve increasingly harder problems”.
To reinforce the interdisciplinarity of the project and to maximize its technological impact, collaborations are planned with neuroscientists at the Champalimaud Foundation and with Unbabel, a Portuguese scale-up company in the crowdsourcing translation industry.
André Martins is a researcher at the Instituto de Telecomunicações, Vice-President of AI Research at Unbabel, Associate Professor at Instituto Superior Técnico, University of Lisbon, and alumnus of CMU Portugal. Having already won an ERC Starting Grant in 2017, in the amount of 1.4 million euros, the researcher has now the opportunity to consolidate this research under the hosting of Instituto de Telecomunicações.
The European Research Council (ERC) supports and funds pioneering projects and researchers at progressive stages of their careers. The ERC Consolidator Grant is intended for researchers in the consolidation phase, with a value of nearly 2 million euros for five years.
André Martins is one of the three researchers that received this grant (ERCCoG 2023) in Portugal.
Listen to the testimonial of André Martins in the podcast "How cool is you science?" (in Spotify):