Program
Program:
Thursday, March 21, 2024
09:15 - 09:30 Opening Remarks
09:30 - 10:30 Invited Talk - Ondrej Dusek
10:30 - 11:00 Coffee Break
11:00 - 12:00 Invited Talk - Dimitra Gkatzia
12:00 - 13:30 Lunch Break
13:30 - 14:00 Shared Task Discussion and Findings
14:00 -15:15 Contributed Talks:
14:00-14:15 Improving Dialog Safety using Socially Aware Contrastive Learning - Souvik Das and Rohini K. Srihari
14:15-14:30 Reliable LLM-based User Simulator for Task-Oriented Dialogue Systems - Ivan Sekulic, Silvia Terragni, Victor Guimarães, Nghia Khau, Bruna Guedes, Modestas Filipavicius, Andre Ferreira Manso and Roland Mathis
14:30-14:45 Evaluating Modular Dialogue System for Form Filling Using Large Language Models - Sherzod Hakimov, Yan Weiser and David Schlangen
14:45-15:00 KAUCUS: Knowledgeable User Simulators for Training Large Language Models - Kaustubh Dhole
15:00-15:15 Advancing Open-Domain Conversational Agents: Designing an Engaging System for Natural Multi-Turn Dialogue - Islam A. Hassan and Yvette Graham
15:15 - 15:30 Closing Remarks
Invited Talk Speaker 1: Ondrej Dusek
Title: Looking for LLMs’ Limits in Dialogue & Data-to-Text
Abstract:
This talk is centered around large language models (LLMs), which are operationally based in dialogue and represent the state of the art in various language processing tasks, including dialogue. The LLMs’ main advancement is their ability to provide meaningful results in virtually any domain based only on simple instructions or examples on the input, with no domain-specific finetuning. However, they still retain some of the problems of the previous generation of language models, in particular their opacity and lack of controllability. Through experiments in dialogue evaluation, task-oriented dialogue and data-to-text generation, we search for the LLMs’ limits, mainly focusing on consistency of their outputs with respect to context. We show that current LLMs can be useful for these tasks and point out potential areas of improvement.
Invited Talk Speaker 2: Dimitra Gkatzia
Title: Dealing with Data & Evaluation Challenges in Low-Resource NLG and Dialogue
Abstract:
Many domains and tasks in natural language generation (NLG) and dialogue are inherently ‘low-resource’, where data and linguistic tools (including evaluation tools) are scarce if not absent. This poses a particular challenge to researchers and system developers in the era of machine-learning-driven NLG, but also calls for more efficient data collection and evaluation procedures. In this talk, I will initially present the challenges researchers & developers often encounter when dealing with low-resource settings in NLG/dialogue and discuss current approaches to low-resource NLG/dialogue. I will then introduce our challenge proposal for the NLG community: low-resource language corpus development (LowRe), a framework designed to collect a single dataset with dual tasks to maximize the efficiency of data collection efforts and respect participants' time. Finally, I will present a case study with Scottish Gaelic using the LowRe framework.