Skip to Main content Skip to Navigation

Apprentissage de Représentation dans les Réseaux de Documents : Application à la Littérature Scientifique

Abstract : The work presented in this thesis, made in collaboration with the company Digital Scientific Research Technology, aims to develop representation learning models for networks in order to address the resolution of different tasks of information retrieval, in particular, on data extracted from the scientific literature. We present GVNR, a network embedding algorithm whose algorithmic time complexity is lower than other mechanism between documents. Finally, we present IDNE, a document network embedding model based on a new mechanism, the topic-attention. We experimentally study the performances of these 4 models on transductive and inductive tasks of classification of nodes and of link prediction with 9 datasets. We show that these models achieve state-of-the-art performances in most datasets on all tasks.In addition, we present our work on expert finding. We introduce a new evaluation methodology and we provide 4 new annotated datasets. We experimentally show the relevance of our evaluation protocol and highlight the remaining steps for the design of an expert finding model based on document network embedding techniques.
Complete list of metadata
Contributor : Abes Star :  Contact
Submitted on : Wednesday, November 24, 2021 - 12:08:24 PM
Last modification on : Thursday, November 25, 2021 - 3:31:19 AM


Version validated by the jury (STAR)


  • HAL Id : tel-03446041, version 1



Robin Brochier. Apprentissage de Représentation dans les Réseaux de Documents : Application à la Littérature Scientifique. Théorie et langage formel [cs.FL]. Université de Lyon, 2020. Français. ⟨NNT : 2020LYSE2087⟩. ⟨tel-03446041⟩



Record views


Files downloads