Iprova team advances the state of the art in Natural Language Processing
Our invention team creates great inventions thanks to our unique data-driven invention software. Members of the team developing that software are also contributing to the state of the art in Machine Learning and Natural Language Processing (NLP).
NLP systems often understand text by looking at words in context. A recent paper co-authored by one of our developers, Matteo Pagliardini and researchers at EPFL introduces a new method to improve the quality of text representation by changing the way a word’s context is analysed.
The new algorithms are made publicly available to allow anyone to benefit from this advance. Methods such as this will help Iprova more precisely identify the inventive insights on which our inventions are built.
The paper is here: https://arxiv.org/abs/1904.05033
Additionally, Iprova developer, Yassine Benyahia, recently presented his paper at the International Conference on Machine Learning. The paper, in collaboration with EPFL (École Polytechnique Fédérale de Lausanne) and Swisscom, presents a novel solution to a problem which they called multi-model forgetting. The problem arises when several machine learning models are trained sequentially on the same task while sharing parts of their parameters. They showed that this phenomenon can hinder the performance of several approaches to the automatic design of ML systems.
Solving this problem reduces the randomness of the automatic design of ML systems. Ultimately this can help improve the quality of ML solutions by reducing the cost required to develop high quality systems.
The International Conference on Machine Learning is held in Long Beach, California in June and is one of the primary academic conferences on ML and AI research in the world.
The paper is here: https://arxiv.org/abs/1902.08232