Post Doc position on NLP (Natural Language Processing) for Human-Robot Interaction @ICAR (Interactions, Cognitions Team), Ecole Normale Supérieure de Lyon
4 Juillet 2023
Catégorie : Post-doctorant
Duration : 12 months (potentially extensible to 14 months) - Expected starting date September 2023, can be adapted
Team : ICAR (Interactions, Cognition Team), Ecole Normale Supérieure de Lyon
Project : PepperMint funded by ASLAN Labex (https://peppermint.projet.liris.cnrs.fr/)
Partners: LIRIS (SyCoSMA, SAARA Teams), ICAR (InSitu Team), University of Oulu-Finland (GenZ),
Supervision: Dr Heike Baldauf-Quilliatre, Dr Frédéric Armetta, Dr Mathieu Lefort
PepperMint project (https://peppermint.projet.liris.cnrs.fr/) is funded by the ASLAN Labex (https://aslan.universite-lyon.fr/). It proposes an exploratory study of embodied turn-taking practices in task-related Human-Robot Interaction (HRI) to improve the social abilities of robots and make HRI more natural to humans. The project initiates a cooperation between researchers in AI (Artificial Intelligence) (LIRIS) and CA (Conversation Analysis) (ICAR and GenZ Oulu - Finland). It investigates if and how CA findings on natural occurring interaction can be used to develop innovative and effective AI models for HRI.
The use case is a Pepper robot that is expected to inform and orient users in a library. In order to perform its task, the robot currently makes use of afinite-state machine for its dialog policy, which one is designed to manage the semantics of the conversation (and provide answers to questions such as where to find a biology book for example). From the robot perspective, the course of the conversation moves between states, thanks to transitions triggered by intentions task-oriented only, caught from the sentences spoken by the user. Of course, such a systematic interaction does not make the interaction natural, because it excludes the emergent character of interaction, i.e. every conversational step can project different types of responses, with regard to the way humans interpret the action.
The interaction thus follows specific rules that help humans to make sens of it. As an example, an offer (“how can I help you ?”) has to be followed by an acceptation (“yes, I would like …”) or a rejection (“well, hum …”). Many of such sequential patterns are imbricated through the interaction, and are at least as important as the semantic content to make the interaction natural.
The ongoing project has already collected some 30 hours of recording sessions from which such patterns are currently analyzed and tagged. In order to propose a better interactive schema, we aim to recognize such patterns thanks to natural language processing (NLP), but also to make use of the natural flow of the conversation to define the “right” interpretation and define the “best” natural answer to the user. Specifically, the analysis of the spoken text will require Natural Language Understanding (NLU) based on Few-Shot Learning. The dialog policy will then make use of this complementary channel (this can be done thanks to a simple adaptation of thefinite-state machine but a more generic and innovative policy framework dedicated to our approach could be developed).
The goal of this position is then to use the annotated dataset for machine learning methods to propose a new AI model for HRI (Human Robot Interaction). The learning will be based on the language modality, and will consider as far as possible other collected modalities that could advantageously help to detect patterns (posture, etc.).
For this subject, the Post Doc will have:
- To collaborate with a Post-Doc in the field of Conversation Analysis, in order to clean and prepare the annotated data produced by the CA researchers
- To develop an improvementof the HRI application (turn-taking patterns detection and integration in the dialog policy)
- To contribute to the scientific communication activities of the PepperMint Project.
We are looking for a candidate having the following skills and background:
- Expertise/Experience/Background in AI and Machine Learning (specifically NLP methods)
- Fluent or good level in written English
- Open-mindedness, teamwork, autonomy and capacity to interact with other disciplines like social sciences. Interest in interdisciplinary research
- Interest in Social Robotics (Human Robot Interaction) would be a plus
Apply to the corresponding offer "Postdoctorant en NLP (H/F) pour interaction humain-robot" on the CNRS website: https://emploi.cnrs.fr/Offres/CDD/UMR5191-MAXPEN-021/Default.aspx