Accueil > Colloques et journées d’études > IWSDS 2019 Special Session — Dialogue systems and lifelong learning

IWSDS 2019 Special Session — Dialogue systems and lifelong learning

Deadline : 0000

Bibliographie

IWSDS 2019 Special Session — Dialogue systems and lifelong learning
April 24-26, 2019
Siracusa, Sicily, Italy

* DSLL description

The topic of dialogue systems and chatbots has been gaining renewed
interest in the recent years, particularly thanks to the recent
development of deep neural networks. Nevertheless, most of the proposed
approaches require a very large amounts of data, which is difficult to
obtain when talking about dialogue. Proposing Methods that fill the data
gap will allow data-driven dialogue systems trained in a specific domain
and task to improve over time, and even learning in a cumulative way new
domains or tasks is of great interest to fill the data gap. This
direction of research is called lifelong learning or continuous
learning. From another angle, another research paradigm that allows for
continuous learning is to design systems are able to learn a new task or
domain through interaction as a student with a teacher could do.

The main obective of this special session is to gather researchers
interested in dialogue systems that interact with the users in order to
learn about new domains or acquire new knowledge.

We invite submissions on all aspects of dialogue systems, lifelong and
interactive learning.

Topics include but are not limited to :

- Dialogue systems that improve over time
- Intelligent systems that use interaction to gather new information
- Specific techniques that can enable learning through interaction,
such as online reinforcement learning, imitation learning, etc.
- Corpora for interactive learning with dialogue
- Demonstration of systems that learn through interaction
- Evaluation methodologies

* Session Committe

Eneko Agirre, University of the Basque Country, Spain
Mark Cieliebak, Zurich University of Applied Sciences, Switzerland
Olivier Galibert, LNE, France
Sahar Ghannay, LIMSI, Univ. Paris Sud, France
Arantxa Otegi, University of the Basque Country, Spain
Anselmo Peñas, Universidad Nacional de Educación a Distancia, Spain
Camille Pradel, Synapse Développement
Sophie Rosset, LIMSI, CNRS, France
Anne Vilnat, LIMSI, Univ. Paris Sud, France

* Important dates

Submission paper : January 15
Author notification : January 25
Camera ready : February 15

For paper submission process, please check the IWSDS 2019 website
(https://iwsds2019.unikore.it/) and the submission paper website
(https://easychair.org/conferences/?conf=iwsds2019)

* Background

Artificial intelligence has made significant advances on solving
prediction and dialogue tasks. But most of the approaches are based on
off-line and supervised learning, where algorithms take as input
annotated data and build a model. Further work is necessary to build
autonomous agents which are capable of learning from the environment and
the interactions, without explicit supervision for each new task. The
goal of Lifelong Learning (LL, also known as Learning to Learn) is to
research methods for continuous learning of various tasks over time and
learning commonalities among them (Chen and Liu, 2018). Current LL
systems exploit similarities between the learned models for past tasks
using task meta-features (Eaton and Ruvolo, 2013) and corresponding
methods to learn representations of tasks, using for instance neural
networks and ensembles of learners. Still, LL assumes that manual
annotations exist for each item to be learned, while autonomous agents
rarely have access to such supervision. In a realistic scenario the
agent receives feedback only after completing a complex task comprising
of several decisions, and needs to guess which of the decisions were
correct or incorrect.

Current interactions between humans and computers are limited to
constrained dialogues, where dialogue systems (aka ChatBots or
Conversational Agents​) are trained on a number of annotated sample
dialogues of a narrow domain. The development cost is considerable, both
in building the representation of the knowledge for the target domain
and in the dialogue management proper, where one of the most important
shortcomings is the variability of human language and the large amount
of background knowledge that needs to be shared for effective
dialogue. In addition, most of the learned knowledge needs to be learned
nearly from scratch for each new dialogue task, including both the
domain knowledge (learned using knowledge induction or knowledge bases)
and the dialogue management module (adapted to the new
domain). Interestingly, humans use dialogue to improve their own
knowledge of a domain. That is, people interact with other people in
order to confirm, retract or refine their understanding. This topic of
learning through dialogue is an emerging one with more and more attempts
to propose framework and tasks to evaluate such system. Most recent work
in this area concern learning through conversation where the supervision
part is given by the user feedback (Weston, 2016), the way the learning
system can ask questions in an online reinforcement learning framework
(Li et al., 2017) and also on how to learn and infer new knowledge
during a dialogue (Mazumder et al., 2018 ; Letard et al., 2016).

This special session will focus on methods and evaluation methodologies
for learning through dialogue. All aspects involved in dialogue (natural
language understanding, dialogue management, natural language
generation, knowledge management) are of interest.

This special session will provide a focal point for the growing research
community on interactive learning with and by dialogue.

** References

Z. Chen, and B. Liu. Lifelong Machine Learning (2nd Edition). Synthesis
Lectures on Artificial Intelligence and Machine Learning. Morgan and
Claypool Publishers. August 2018, 207p

E. Eaton and P. L. Ruvolo. 2013. ELLA : An efficient lifelong learning
algorithm. In ICML 2013.

Sahisnu Mazumder, Nianzu Ma, Bing Liu. Towards a Continuous Knowledge
Learning Engine for Chatbots. arXiv:1802.06024 [cs.CL], 16 Feb. 2018.

Jiwei Li, Alexander H. Miller, Sumit Chopra, Marc’Aurelio Ranzato, Jason
Weston, Learning through Dialogue Interactions by Asking Questions, ICLR
2017.

Jason E. Weston. Dialog-based language learning. NIPS 2016.

Vincent Letard, Sophie Rosset, Gabriel Illouz. Incremental Learning From
Scratch Using Analogical Reasoning. ICTAI 2016.

Retour à l'annuaire