Laboratoire Parole et Langage
UMR 7309 CNRS | Aix-Marseille Université
5 avenue Pasteur ? 13100 Aix-en-Provence (France)
Tél : +33 (0)4 13 55 36 20+33 (0)4 13 55 36 20 ? Fax : +33 (0)4 13 55 37 88
Site du LPL : www.lpl-aix.fr
LPL Newsletter : lpl-newsletter-subscribe@lpl-aix.fr
*****
Vendredi 15 mai 2015
9h-12h15 LPL, salle de conférences B011, 5 avenue Pasteur, Aix-en-Provence
Séminaire thématique, cycle TAL
Apprentissage automatique et sciences du langage
Philippe Muller (IRIT, Toulouse)
Pascal Denis (INRIA, Lille)
Vincent Labatut (LIA, Avignon)
Benoît Favre (LIF, Marseille)
Stéphane Rauzy (LPL, Aix-en-Provence)
Résumé L'apprentissage automatique (AA) (ou Machine Learning ML) a d'abord constitué une ligne de fracture entre le traitement automatique des langues (TAL) et les sciences du langage en raison d'une part du grain relative grossier des premiers travaux de ML et de petite taille des corpus linguistiques. L'évolution de ces deux domaines renouvelle cependant l'intérêt de travaux croisés. D'une part, les techniques d'AA ont évolué très rapidement au point de proposer des méthodes d'apprentissage de structures complexes (voir en en particulier l'exposé de Pascal Denis sur les résolutions anaphores des coréférences en discours). par ailleurs, les outils d'apprentissage automatique sont devenus beaucoup plus accessibles aux non-spécialistes qu'ils ne l'étaient via (i) des package pour les outils statistiques (e.g R), (ii) des API pour des langages de scripts (Orange, Scikit,...), (iii) des outils stand-alone (Weka,...).
D'autre part, la taille des corpus utilisés en linguistique a considérablement évolué, tant pour l'écrit (où utiliser le web comme corpus est devenu pratique courante) qu'à l'oral où la taille des corpus de parole augmente de manière régulière. Le contexte est donc favorable à l'émergence de méthodologies permettant par exemple d'ajouter des contraintes à l'approche typiquement data-driven de l'apprentissage automatique (en particulier présentation de Philippe Muller). Cette journée est aussi l'occasion d'écouter des collègues d'AMU et de l'Université d'Avignon travaillant sur ces domaines.
Programme
9:00 - 9:15 : Introduction et présentation de la journée
9:15 - 9:55 : Apprentissage de structures linguistiques sous contraintes (Philippe Muller, IRIT, Toulouse)
9:55 - 10:35 : Latent trees for joint anaphoricity and coreference resolution (Pascal Denis, INRIA, Lille)
10:35 - 10:45 : Pause Café
10:45 - 11:15 : Natural language and complex networks (Vincent Labatut, LIA, Avignon)
11:15 - 11:45 : Detecting semantic change with word embeddings (Benoît Favre, LIF, Marseille)
11:45 - 12-15 : Parsing the CID, a corpus of French spontaneous speech in interaction (Stéphane Rauzy, LPL, Aix)
12:15 - 12:45 : Apéritif
Plus d?infos sous http://lpl-aix.fr/event/1614
*****
Vendredi 22 mai 2015
11h-12h LPL, salle de conférences B011, 5 avenue Pasteur, Aix-en-Provence
Christopher T. Kello
University of California, Merced
Complexity Matching in Vocal Interactions
Résumé Spoken conversations require coordination, but it is challenging to formalize and quantify coordination in vocalizations. Much progress has been made by expressing coordination in related terms of synchronization, entrainment, alignment, and convergence. But coordination in speech appears to go beyond these terms of behavioral matching, because interlocutors make distinct contributions to vocal interactions. In this talk, two studies will be presented that employed measures of complexity matching to examine coordination in the acoustics of vocal interactions that are inherently irregular and variable across a wide range of timescales. One study examined speech signals in adult conversations between pairs of individuals, and the other study examined infant vocalizations in conjunction with those of caregivers and other adults in the infant vocalization environment.
In both studies, the temporal dynamics of vocalizations were boiled down to time series of onsets in acoustic energy, and a single power law was found in all conditions that quantified the nested clustering of onsets across a range of timescales. The parameters of this power law matched for vocalizations of individuals engaged in interaction, indicating a different kind of coordination compared with behavioral matching. This measure of complexity matching was sensitive to different types of conversations and different types of infant vocalizations, and changed with infant age. The talk will end with a discussion of the implications of these results for theories of language development and interaction.
Plus d?infos sous http://lpl-aix.fr/event/1617
*****
Vendredi 5 juin 2015
9h30-12h15 LPL, salle de conférences B011, 5 avenue Pasteur, Aix-en-Provence
Séminaire thématique, cycle Cognition
Language across domains and modalities
Friedemann Pulvermuller*, Philippe Blache**, Kristof Strijkers**
*Brain Language Laboratory, Freie Universtät Berlin (Germany)
** LPL, CNRS & AMU, Aix-en-Provence (France)
Résumé The goal of this seminar is to demonstrate that we can substantially advance in our understanding of the mechanics, architecture and function of language by adopting an interdisciplinary research strategy. Traditionally, language has been approached as a rather encapsulated skill (segregated from other areas of cognition and human behavior). Furthermore, within the domain of language each modality could count on its own separate research tradition, with relatively little cross talk between the production and comprehension of language. While such segregated approach to the topic of language has certainly produced significant progress, by now it is evident that a modular view on language processing comes short to explain the full complexity and diversity of this uniquely human ability. Instead, research in the last decade suggests that there is in fact much more overlap between the different language modalities as originally assumed. Likewise, across domains, evidence has been brought forward which highlights that both the representations and dynamics underpinning language can rely on processes general to other domains of cognition and even recruit systems that were traditionally believed to be fully separated of the linguistic system. If so, it is clear that in order to advance upon our knowledge of how language comes about in the human mind, we need to address empirical questions which take into account the domain-interactive and cross-modal nature of linguistic processing. In this seminar three different talks will be presented which aimed at studying language across domains and modalities. More concretely, in the first talk Friedemann Pulvermuller will demonstrate how the integration of basic neurobiological principles with linguistics predicts that words in the brain rely on overlapping networks between the perception (sensori) and action (motor) of language. In a second talk, Philippe Blache will detail a language architecture, relying on the integration of computer science with linguistics, which relies on the convergence of different and mutli-modal sources of information to construct meaning and syntax. And finally, Kristof Strijkers will discuss the role of top-down processes such as attention during language processing and specify how such domain-general regulatory properties of the brain affect basic psycholinguistic phenomena.
Programme
9h30: Welcome
9h45-10h45: Friedemann Pulvermuller (Brain Language Laboratory, FU Berlin) :
Brain Circuits For Language Understanding: Theory ? Model ? Experiment
10h45-11h: Coffee break
11h-11h30: Philippe Blache (LPL, Aix-en-Provence) :
Integrating domains and modalities: a construction-based architecture for language processing
11h30-12h: Kristof Strijkers (LPL, Aix-en-Provence) :
The Role of Intention and Goal-Directed Behavior in Language Processing.
12h-12h15: Concluding remarks
Plus d?infos sous http://lpl-aix.fr/event/1618 | 
