Research Interests
Projects and Collaborations
Scientific Dissemination and Responsabilities
Program Committees
Ph.D. Students and Interns Mentored
Challenges

Associate Professor
Visiting Scientist at Microsoft
Research Assistant
Ph.D. in Computer Science

Research Interests

  • Artificial Intelligence
  • Machine Learning
  • Neural Networks
  • Speech and Language Processing
  • Spoken Language Understanding
  • Semantic Web
  • Image Processing

Projects and Collaborations

ANR Projects

Responsable of funded Projects
  • 2017-2020 : Orkis - (CIFRE, start on March 2017­ - project leader, 12K€/year) Complex-valued deep neural networks for speech and language processing. (Ph.D. thesis of Titouan Parcollet).
  • 2016-2020: MAGIS - (GDR CNRS 2017-2020 – responsable for the LIA, 20K€) Méthodes et Applications pour la Géomatique et l'Information Spatiale. This fund is obtained from a collaboration with CNRS UMR ESPACE team and the LIA (international papers). Participants: CNRS UMR ESPACE, ...

Participant of ANR Projects
  • 2014-2018 : GaFes Project - Gallerie des Festivals, Web content analysis, the Web as cultural media, video Re-purposing. Participants: Centre Norbert Elias, Syllabs, GECE, EURECOM.
  • 2013-2015: ContNomina Project - Exploitation of context for proper names recognition in the diachronic audio documents. Participants: LORIA (University of Lorraine).
  • 2012-2015: Projet DECODA - Speech Analytics in recorded call-center conversations. Participants: LIF (University of Aix-Marseille), RATP (industrial, Paris), SONEAR (industrial, Avignon).
  • 2011-2014: Projet SuMACC - Semi-Supervised Cooperative learning Multimedia concepts for the Categorization and Detection of Novel Concepts. Participants: EURECOM (University of Sophia Antipolis, Nice), WIKIO (industrial, Paris), Syllabs (industrial, Paris).
Academic Collaborations
Industrial Collaborations

Scientific Dissemination and Responsabilities

    Scientific responsibilities
    • Since March 2017: Scientific leader at the Structure Fédénative de Recherche (SFR) Agorantic of the UAPV (France) with Graham Ranger full professeur at UAPV
    • Axe 5 of the SFR Agorantic called "Structuration and Exploitation of Corpuses (SEC)" : The aim of this axe is to merge and structure common resources from different domains, to better process this broad spectrum of knowledge across the heterogeneous domains of researchers of the SFR Agorantic.

    • 2012: EACL, European Chapter of the Association for Computational Linguistics.
    Presentations and seminars
    • 2016 : Université d'Avignon et des Pays de Vaucluses (UAPV) (France)
    • During the inter-disciplinary workshop in the context of the Structure Fédénative de Recherche (SFR) Agorantic of the University of Avignon, I have presented a broad spectrum of methodologies of neural networks.

    • 2015: Laboratoire d’Informatique Fondamentale de Marseille (LIF) (France)
    • I have presented during this seminar the studies that I have done during my Ph.D. related to the thematic representations of noisy documents in abstract spaces.

    • 2015: Centre de recherche Inria Nancy - Grand Est (France)
    • I have given a seminar related to my works on i-vectors representation of noisy documents from ASR and from Reuters corpora.

    • 2014: Apple, Infinite Loop, Cupertino (US)
    • Presentation of a robust representation of highly imperfect transcriptions from an automatic speech recognition (ASR) system. I have detailed a new modelisation of these documents in a reduced space, named total variability space. The documents are dialogues between customers and agents from the Paris transportation customer care service (RATP) (DECODA project).

    • 2013: Microsoft Research, Cambridge (UK)
    • During the selection of projects to be presented during the annual meeting Tech'Fest 2014 of Microsoft, I have presented the works done during my visit, for the PowerPoint slides modelisation in order to provide relevant contents to Microsoft customers. Thus, I have used original methods to allow Microsoft customers to enrich their PowerPoint presentations with relevant content. This project was selected to be showed at annual Microsoft meeting TechFest'14 (40 projects were submitted and 22 were selected).

    Popularization
    Professional societies

Program Committees

Ph.D. Students and Interns Mentored

    Graduated Ph.D Students
    • Killian Janod (2013-2017).
      Title: Possibility theory applied in a speech processing context.
      CIFRE thesis with Orkis company. Now in the research department of ISMART. 8 common publications (1 international journal, 3 international conferences and 4 national conferences)
    • Mohamed Bouaziz (2013-2017).
      Title: Delinearization of synchronous video streams.
      CIFRE thesis with EDD company started in november 2013. Now in the research department of Airbus Defence and Space. 6 common publications (3 international conferences, 2 national conferences and 1 national workshop)
    Ph.D. Thesis
    • Titouan Parcollet (2017-...).
      Title: Complex-valued neural networks for speech and language processing.
      CIFRE thesis with Orkis company started in march 2017.
    Interns Mentored
    • Maxime Bisotto - B.Sc (3 months started in febuary 2019). Sujet Agorantic : Développer une plateforme web permettant aux chercheurs de l'université d'Avignon de partager leurs corpus (Ensemble de données de recherche, généralement sous forme textuelle) avec leurs collègues et leurs étudiants. Le stagiaire devra participer à la conception – architecture logicielle et modèle relationnelle de la base de données -, développer avec un outil de gestion des sources et procéder à des tests – unitaires et fonctionnels. Co-direction avec Graham Ranger et Emili Volpi (Agorantic).
    • Rida El Allami - B.Sc. (3 months started in febuary 2019). Sujet Agorantic : Développer une plateforme web permettant aux chercheurs de l'université d'Avignon de partager leurs corpus (Ensemble de données de recherche, généralement sous forme textuelle) avec leurs collègues et leurs étudiants. Le stagiaire devra participer à la conception – architecture logicielle et modèle relationnelle de la base de données -, développer avec un outil de gestion des sources et procéder à des tests – unitaires et fonctionnels. Co-direction avec Graham Ranger et Emili Volpi (Agorantic).
    • Nour Elhouda Ayari - Graduate from Sup’com, Tunisia (6 mounths started in febuary 2018). Title: Mobility prediction in Wireless Networks based on Recurrent Neural Networks. Direction with Majed Haddad and Rachid Elazouzi.
    • Titouan Parcollet - Research engineer (6 mounths started in september 2016). Title: Quaternion Neural Networks for Natural Language Processing. Direction with Linarès Georges.
    • Afef Arfoui - Graduate from INPT, Morocco (6 mounths started in febuary 2018). Title: User Experience Assessment Control: measures and metrics. Direction with Majed Haddad and Rachid Elazouzi.
    • Titouan Parcollet - Graduate (6 mounths started in febuary 2016). Title: Quaternion Neural Networks for Natural Language Processing. Direction with Linarès Georges.
    • Amina Brahem - Engineering school (6 mounths started in march 2014). Title: Electroencephalogram signals processing based on Speech Processing methods. Direction with Linarès Georges.
    • Etienne Papegnies - Graduate (3 mounths started in june 2014). Title: Electroencephalogram signals processing based on Speech Processing methods. Direction with Linarès Georges.
    • Mathias Quillot - Undergraduate (2 mounths started in june 2014). Title: Demo for the ANR ContNomina project. Direction with Dufour Richard and Linarès Georges.

Challenges

Papers are available here

  • 2016 : DSTC5'16 - Representation of dialog states and updates them at each moment on a given on-going conversation.
  • 2014: IWSLT'14 - English-to-Slovene and English-to-Polish Translation Tasks.
  • 2013: MediaEval'13 - "Crowdsourcing", "MusiClef tracks" and "Spoken Web Search" tasks.
  • 2013: Deft'13 - The organizers address this edition with a new application domain on a theme that has been studied in an evaluation campaign in the past (Computer Cooking Contest): cooking recipes. We focuse on two analysis functions in DEFT2013, document classification (task 1 to 3) and information extraction (task 4), in a speciality domain.
  • 2013: RepLabs'13 - An evaluation campaign for Online Reputation Management Systems.
  • 2012: Inex'12 - Tweet Contexualization Track.
  • 2011: MediaEval'11 - Discover events and detect media items that are related to either a specific social event or an event-class of interest.


Associate Professor (2015-...)

Since september 2015, I am an Associate Professor within the Laboratoire Informatique d'Avignon (LIA), France. Overall, my current and future research projects reflect my strong interest in a broader research agenda that seeks to study the role that both document content and representation play in different NLP tasks. Therefore, my research interests lie on novel issues related to words embeddings applied in different tasks such as possibility theory for acoustic models and confidence measure (K. Janod Ph.D. thesis). I also participate to the Structure Federative de Recherche (SFR) Agorantic and to some Agence Nationale de la Recherche (ANR) projects, and I am involved in different evaluation compaigns as well.


Visiting Scientist at Microsoft (2013-2014)

I spent time visiting the Microsoft Research (Cambridge, UK) Laboratory to work under the supervision of Professor Youssef Hamadi for 3 months (nov. 2013 - feb. 2014) on the PowerPoint Assistant project. The PowerPoint Assistant project's aim is to provide relevant contents to Microsoft customers. Thus, I proposed original methods to allow Microsoft customers to enrich their PowerPoint presentations with relevant content. This project was selected to be showed at annual Microsoft meeting TechFest'14 (40 projects were submitted and 22 were selected). During this meeting, Microsoft Research scientists from across the globe gather to display and discuss their latest projects, which encompass a broad spectrum of computer-science investigations. You can found out a demo of the PowerPointAssistant in the section Demo.


Research Assistant (2014-2015)

I have been an Teaching and Research Assisitant at the Laboratoire Inforatique d'Avignon (LIA) within the University of Avignon from september 2014 to august 2015. My works were related to first extend the paradigm of expansion and fusion in the total variability space of relevant information contained in documents such as dialogues from DECODA projects or Reuters articles (this work has been pblished in the international journal IEEE/ACM Transactions on Audio, Speech, and Language Processing).
Still in the context of highly imperfect dialogues from the DECODA corpus, I have proposed to compare differents normalization and denoising approach in the highly imperfect transcriptions context (DECODA) (study presented during the annula international conference ISCA InterSpeech).
The stat-of-the-art methods for most of the Natural Language Processing tasks, use Machine Learning approaches such as Deep Neural Networks (DNN). I have participated to a common study with the Language and Speech Technology (LST) team of the University of Maine (le Mans, France) led by Professor Yannick Estève. We have proposed to extend the c-vector representation of a dialogue from the DECODA prject, with some other features such as semantic features. This study has been presented during the international conference IWSDS and has been edited on the international book Natural Language Dialog Systems and Intelligent Assistants.
I have finally proposed to study different thematic representations (Latent Dircihlet Allocation (LDA), Supervised Latent Dircihlet Allocation (sLDA), ...) for the theme identification task in the context of dialogues from the DECODA project with the collaboration of Professor Youssef Hamadi from Microsoft Research Cambridge (UK). This study has been presented during the international conference CICLing.

These machine learning methods are often costly in terms of time processing. I have then proposed to compare the quality (in terms of perplexity, KL divergence) as well as the time processing of of thematic models from an LDA learned from summarized documents and from the whole documents from Wikipédia. I have demonstrated that the quality is at least equivalent even better when the LDA model is learned with summarized documents. Obviously, the time porcessing is massively reduced (this work has been published during the international conference CICling).

The LIA strongly participate to the Structure Fédérative de Recherche (SFR) Agorantic. I have collaborated with Didier Josselin from the UMR ESAPCE team of the University of Avignon, to locate automatically a short text message (tweet) for the "JeSuisCharlie" event in january 2015. These works have showed that a thematic representation (Auhtor-Topic Model) with different hidden variables allows us to take into account not only the tweet content itself (words), but also the loacate (country). The results observed have pointed out that this thematic representation allows us to obtain an accuracy of roughly 95%. This work has been presented in both SPATIAL STATISTICS (computer science) and ISPRS GW (geography).


Ph.D. Thesis

Title: Robust Representation of noisy documents in homogeneous spaces

Defended at November the 25th 2014 at University of Avignon, France.

The composition of the doctoral thesis jury:

Frédéric BÉCHET (Professor)
University of Aix-Marseille, Marseille, France
President
Jérôme R. BELLEGARDA (Apple Distinguished Scientist)
Apple Inc., Cupertino, USA
Reviewer
Laurent BESACIER (Professor)
University of Joseph Fourier, Grenoble, France
Reviewer
François YVON (Professor)
University of Paris-Sud, Paris, France
Reviewer
Youssef HAMADI (Senior Scientist)
Microsoft Research, Cambridge, UK
Examiner
Benjamin PIWOWARSKI (CNRS Scientist)
LIP6, Paris, France
Examiner
Georges LINARES (Professor)
University of Avignon, Avignon, France
Adviser
Richard DUFOUR (Assistant Professor)
University of Avignon, Avignon, France
Adviser

Abstract: In the Information Retrieval field, a document is usually considered as a "bag-of-words". This model does not take into account the temporal structure of the document and is sensitive to noises which can alter its lexical form. These noises can be produced by different sources: uncontrolled form of documents in micro-blogging platforms, automatic transcription of speech documents which are error-prone, lexical and grammatical variabilities in Web forums... The work presented in this thesis addresses issues related to document representations from noisy sources.

My thesis consists of three parts in which different representations of content are available. The first one compares a classical representation based on a term-frequency representation to a higher level representation based on a topic space. The abstraction of the document content allows us to limit the alteration of the noisy document by representing its content with a set of high-level features. The major problem with this topic-based representation is that its parameters are chosen empirically.

The second part presents a novel representation based on multiple topic spaces that allow us to solve three main problems: the closeness of the Titles discussed in the document, the tricky choice of the "right" values of the topic space parameters and the robustness of the topic-based representation. Based on the idea that a single representation of the contents cannot capture all the relevant information, we propose to increase the number of views on a single document. This multiplication of views generates "artificial" observations that contain fragments of useful information. However, it has the disadvantage of being very large, redundant and of containing additional variability associated with the diversity of views. I have proposed a method based on factor analysis to compact different views to obtain a new robust representation of low dimension which contains only the informative part of the document while the noisy variabilities are compensated.

Nonetheless, during the learning process of topic spaces, the document is considered as a "bag- of-words" while many studies have showed that the word position in a document is useful. A representation which takes into account the temporal structure of the document based on hyper- complex numbers is proposed in the third part. This representation is based on the hyper-complex numbers of dimension four named quaternions.

Keywords: Robust representation, noisy document, latent Dirichlet allocation, multi-views, factor analysis, quaternion.