Séminaire R-D - Vers des systèmes de génération du langage naturel cohérents, fluides et adaptés au contexte

Séminaire R-D - Vers des systèmes de génération du langage naturel cohérents, fluides et adaptés au contexte
24/04/18 11h00
CRIM (405, avenue Ogilvy, bureau 101, Montréal)

Vers des systèmes de génération du langage naturel cohérents, fluides et adaptés au contexte : études de cas sur la prédiction de la détermination, la détection des déclencheurs de présupposition et le transfert de style

Conférencier
Jad Kabbara, boursier au sein de l'équipe Parole et Texte du CRIM et étudiant au doctorat à l'Université McGill (School of Computer Science). 
 
Présentation en anglais

Towards coherent, fluent and context-appropriate Natural Language Generation systems: Case studies on definiteness prediction, detecting presupposition triggers and style transfer
 
Speaker
Jad Kabbara, scholarship student within the Speech and Text team at CRIM, and Ph.D. candidate at the School of Computer Science at McGill University. 
 
Abstract
Nowadays, it is becoming increasingly common to rely on intelligent systems such as smartphones to carry out day-to-day tasks. There is thus a need to improve these virtual personal assistants such that they interact with users in a more fluent way using natural language. From translation to dialogue systems and other tasks, there is a need to design natural language generation (NLG) systems that produce language that is humanlike, context-appropriate, fluent and diversified. To this end, we present in this talk three research efforts in that direction.
 
The first part of the talk focuses on definiteness prediction, the task of determining whether a noun phrase should be definite or indefinite. In English, one instantiation of this task is to predict whether to use a definite article (the), indefinite article (a/an), or no article at all. We will examine the potential of recurrent neural networks for handling pragmatic inferences involving complex contextual cues for this task. We will present an LSTM network with an attention mechanism and empirically show that our model outperforms a previous state-of-the-art system and other baselines.
 
In the second part of the talk, we will introduce the task of predicting adverbial presupposition triggers such as “also” and “again”. Solving such a task requires detecting recurring or similar events in the discourse context, and has applications in NLG tasks such as summarization and dialogue systems. We will present a deep learning model with a novel attention mechanism tailored to this task and demonstrate that our model statistically outperforms a number of baselines, including an LSTM-based language model.
 
Linguistic style is a crucial aspect of making NLG systems natural and humanlike. To that end, we will close the talk by presenting first steps that have been taken in the direction of textual transfer of linguistic style and highlight ongoing research on the transfer (controlling) of linguistic attributes such as sentiment (positive, negative), tone (personal or not), length/brevity, etc. We will present preliminary results using conditional language modelling and highlight avenues for future research.
 
Biography
Jad Kabbara is a Ph.D. candidate at the School of Computer Science at McGill University where he works under the supervision of Prof. Jackie CK Cheung. He’s a member of the Reasoning and Learning Lab (RL-Lab) and the Montreal Institute for Learning Algorithms (MILA). Before his Ph.D., Jad completed his Master’s degree in Electrical and Computer Engineering in 2014 also at McGill University and his Bachelor’s degree in Computer and Communications Engineering at the American University of Beirut (Lebanon) in 2011. His research interests lie at the intersection of artificial intelligence and natural language processing. Specifically, his work falls under computational pragmatics and natural language generation. 

Les séminaires scientifiques du CRIM, gratuits et ouverts à tous, sont donnés par des experts de renommée internationale, des collaborateurs universitaires, le personnel de R-D et les boursiers de 2e et 3e cycles du CRIM. Au programme, des présentations conviviales sur les dernières avancées scientifiques et technologiques.

Eventbrite - Journée Techno du CRIM : 30 ans d'innovation collaborative en TI

Conférence gratuite. Inscription requise.

Mardi, 24 avril 2018, de 11 h à midi.  Au CRIM, 405, avenue Ogilvy, bureau 101, Montréal. 

  • #Bientôt ???? Le CRIM présentera le projet Application des technologies vocales aux langues autochtones le 19 novembre… https://t.co/FnO5IKw79H
  • Tom Landry RT @Tom_Landry_: Ce fût un réel plaisir que de partager ma vision de l'avenir de la géomatique. Et que dire des mots-clés sélectionné par l…