Séminaires

Prochain séminaire :

Jeudi 28 Février - Matthieu RIOU
Jeudi 07 Mars - Antoine Caubriere
Jeudi 14 Mars - Lucile Sassatelli (invité de Rachid ELAZOUZI)
Jeudi 21 Mars - Mathias Quillot
Jeudi 11 Avril - Vincent Labatut
Jeudi 18 Avril - Carlos Cicak
Jeudi 13 Juin - Jiajia LIU (invité de Abderrahim BENSLIMANE)

 

Study of the European Parliament votes through the multiple partitioning of signed multiplex networks (jeudi 24 janvier 2019 - Nejat Arinik)

Summary: For more than a decade, graphs have been used to model the voting behavior taking place in parliaments. However, the methods described in the literature suffer from several limitations. The two main ones are that 1) they rely on some temporal integration of the raw data, which causes some information loss; and/or 2) they identify groups of antagonistic voters, but not the context associated to their occurrence. In this article, we propose a novel method taking advantage of multiplex signed graphs to solve both these issues. It consists in first partitioning separately each layer, before grouping these partitions by similarity. We show the interest of our approach by applying it to a European Parliament dataset. By comparison to existing approaches, our method has the following advantages. First, it undergoes much less of the information loss appearing when integrating the raw voting data to extract the voting similarity networks. Second, in addition to antagonistic groups of voters, it allows identifying sets of legislative propositions causing the same polarization among these groups. Third, it does not require to filter out (quasi)-unanimous propositions, or to discard week links appearing in the model for interpretation or computational purposes. Fourth, it explicitly represents abstention in each roll-call vote layer, which allows detecting relevant groups of abstentionists.

Présentation

 

Speech recognition with quaternion neural networks (jeudi 17 janvier 2019 - Titouan Parcollet)

Summary: Neural network architectures are at the core of powerful automatic speech recognition systems (ASR). However, while recent researches focus on novel model architectures, the acoustic input features representation remain almost unchanged. Traditional ASR systems rely on multidimensional acoustic features such as the Mel filter bank energies alongside with the first, and second order derivatives to characterize time-frames that compose the signal sequence. Considering that these components describe three different views of the same element, neural networks have to learn both the internal relations that exist within these features, and external or global dependencies that exist between the time-frames. Quaternion-valued neural networks (QNN), recently received an important interest from researchers to process and learn such relations in multidimensional spaces. Indeed, quaternion numbers and QNNs have shown their efficiency to process multidimensional inputs as entities, to encode internal dependencies, and to solve many tasks with up to four times less learning parameters than real-valued models. The presentation will first introduce the basics of QNNs alongside with motivations to use quaternion-valued models over traditional real-valued ones. Then, a sum up of all the conduced experiments with state-of-the-art QNNs architectures (code is provided) and competitive results will be detailed.

Présentation

 

Génération de texte à partir d’AMRs (jeudi 29 novembre 2018 - Guy Lapalme)

Summary: Après avoir présenté les "Abstract Meaning Representation" (AMR) et leurs utilisations, nous décrirons un système symbolique qui fournit une traduction "littérale" en anglais. Nous comparerons ensuite ces résultats avec ceux obtenus par des méthodes basées sur l'apprentissage automatique. Nous terminerons par des suggestions pour les développeurs d'AMRs inspirées par cette vue "générative".

Bio: Guy Lapalme est professeur associé au département d'informatique et de recherche opérationnelle. Depuis plus de 30 ans, il travaille sur la correction d'orthographe, l'édition de dictionnaire et la génération automatique de texte. En 1997, il a créé le RALI un des plus grands laboratoires sur le traitement de la langue au Canada. En plus de la recherche universitaire, le RALI a effectué des recherches en collaboration avec des industriels dans les domaines du résumé automatique, de l'extraction d’information, de la traduction automatique et assistée par ordinateur et du recrutement en ligne.

Présentation

 



Laboratoire Informatique d'Avignon

Université d'Avignon et des Pays de Vaucluse
339 chemin des Meinajaries, Agroparc BP 91228, 84911 Avignon cedex 9
+33 (0)4 90 84 35 00