Myriadisation et éthique pour le traitement automatique des langues
Karën Fort · HAL (Le Centre pour la Communication Scientifique Directe) · 2022
In the past ten years, Natural Language Processing (NLP) has undergone two revolutions : the extreme shortening of the distance between research outputs and the end user and the advent of deep learning. As a result, data needs have exploded alongside ethical issues. This "habilitation à diriger des recherches" presents the work I have carried out in the field of the production of manual annotations for NLP by crowdsourcing, in particular using games with a purpose, and in that of ethics for the TAL. I redefine crowdsourcing and citizen science in general and I present in detail the games with a purpose, their strengths and their limits. I focus more particularly on ZombiLingo, which was used to collect dependency syntax annotations for French, and RigorMortis, a game aiming at collecting multiword expressions (MWE). In the last part, I focus on ethics for NLP, a sub-field that was only truly recognized from 2016 and of which I was a forerunner. I return to its history, its recent evolution and present my work, carried out in a more deontological than consequentialist perspective, allowing to have a systemic vision of NLP and the ethical problems it poses.