Project: Hyperbolic Geometry for compositional generalization in NLP Keywords: NLP, compositional generalization, hyperbolic geometry We invite applications for an 18-month post-doc position at the Laboratoire d'informatique de Grenoble (LIG, Université Grenoble Alpes). Scientific context: Despite their success in many NLP tasks, seq2seq models fail at learning simple generalization rules, i.e. interpreting novel combinations of known lexical items, as illustrated by results on challenge datasets such as SCAN [1] or COGS [2]. The HyperboTAL project is a multi-disciplinary collaboration aiming at applying the tools of hyperbolic geometry to improve the systematic compositional capabilities of Natural Language Processing models. The project will build on recent proposals, as [3] and [4], that use hyperpolic geometry to better encode linguistic information (e.g. Poincaré embeddings). Main tasks: - Literature review on hyperbolic geometry applications in NLP and deep learning approaches to systematic compositionality. - Design compositionality models that bring together the tools of hyperbolic geometry and those of deep learning. - Implementation and evaluation of proposed models on existing challenge datasets (COGS). The scientific orientations of the post-doc may vary according to the candidates' background and interests. Requirements: - PhD in computer science. - Background and/or strong interest in Mathematics and Natural Language Processing. - Programming skills: proficiency in python and experience with a deep learning library. - Proficiency in either French or English. Work environment: - Location: Laboratoire d'Informatique de Grenoble (LIG), on Université Grenoble Alpes campus. - Cosupervision by Émilie Devijver (PI), Gérard Besson, Maximin Coavoux, Éric Gaussier. - Gross salary: minimum 2400 euros per month (depends on experience). - Starting date: September 2022 (flexible). To apply, please send a cv and a cover letter to contacts below as soon as possible (applications are open until the position is filled). Contacts (for applications or any questions regarding the position/project): emilie.devijver@univ-grenoble-alpes.fr g.besson@univ-grenoble-alpes.fr maximin.coavoux@univ-grenoble-alpes.fr eric.gaussier@imag.fr [1] Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. B. M. Lake, M. Baroni. ICML 2018. [2] COGS: A Compositional Generalization Challenge Based on Semantic Interpretation. N. Kim, T. Linzen. EMNLP 2020 [3] Poincaré Embeddings for Learning Hierarchical Representations. M. Nickel, D. Kiela. Neurips 2017. [4] Hyperbolic graph neural networks. Q Liu, M Nickel, D Kiela. NeurIPS 2019.