HAL will be down for maintenance from Friday, June 10 at 4pm through Monday, June 13 at 9am. More information
Skip to Main content Skip to Navigation
Theses

Représentation des mots et des connaissances : construction, utilisation et évaluation des plongements euclidiens et hyperboliques

François Torregrossa 1
1 LinkMedia - Creating and exploiting explicit links between multimedia fragments
Inria Rennes – Bretagne Atlantique , IRISA-D6 - MEDIA ET INTERACTIONS
Abstract : Texts, graphs and hierarchies are heterogeneous form of data. In this thesis, we deal with these three categories by leveraging geometric spaces suiting each of them. Starting with the study of word embedding for representing texts, we try to evaluate their quality and provide properties highlighting effectiveness. Enriching those word embeddings with expert knowledge entails a better processing of an industrial task. Building from this fact, we develop new ways to accumulate structured knowledges (e.g. hierarchies) from free online sources (Linked Open Data), or, from information contained in flat contents such as texts (taxonomy induction) or other databases. These hierarchies are then represented using hyperbolic spaces, and we propose a new hyperbolic embedding method. The ultimate goal being to restrain information loss from the structure to the continuous space, and then to combine this new knowledge with traditional word embeddings.
Document type :
Theses
Complete list of metadata

https://tel.archives-ouvertes.fr/tel-03592848
Contributor : Abes Star :  Contact
Submitted on : Tuesday, March 1, 2022 - 3:36:40 PM
Last modification on : Friday, April 8, 2022 - 4:08:03 PM

File

TORREGROSSA_Francois.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03592848, version 1

Citation

François Torregrossa. Représentation des mots et des connaissances : construction, utilisation et évaluation des plongements euclidiens et hyperboliques. Automatique. Université Rennes 1, 2021. Français. ⟨NNT : 2021REN1S078⟩. ⟨tel-03592848⟩

Share

Metrics

Record views

33

Files downloads

15