GEMINI – Graph-based knowlEdge for Modelling Intelligent Natural Interaction
This special session focusses on knowledge and knowledge modelling for dialogue systems, in particular on the opportunities and challenges for enhancing and stabilizing dialogue capabilities of chatbots, robots, and virtual agents with the use of LLMs. The workshop aims to explore new methods and techniques for knowledge modelling, especially knowledge graphs and graph databases, and their effective use in various dialogue-related tasks, to discuss technologies to combine knowledge graphs and LLMs for robot interactive applications, to share experience and results on experiments and underlying principles, and to find ground-breaking results to advance LLM-based dialogue modelling with reliable and uptodate knowledge.
Motivation and background
Knowledge has always been one of the fundamental aspects of dialogue modelling. Representing and extracting knowledge from conversational interactions is a challenge needed to be addressed in order to scale up dialogue systems to various topics and domains, to extend conversational capabilities of practical systems with relevant responses in case the database falls short of answers, to adapt to the user’s styles, interests, features, and wishes, and to accommodate and build rapport for long-term interactions, in a natural manner.
A forerunner of the Special Session is a series of IJCAI workshops “Knowledge and Reasoning in Practical Dialogue Systems” which were organised at IJCAI and focussed on various challenges arising when implementing (conversational) dialogue systems for different types of users, applications, and scenarios: https://www.ida.liu.se/~arnjo82/ijcai-ws-01/
In more than 10 years, the landscape for dialogue modelling has changed drastically. From rule-based systems with hand-crafted knowledge to statistical modelling with large language models, paving way to chatty generative-AI systems that happily produce responses for any topic, any style, and any format. However, despite the success of LLMs in dialogue modelling, it has become clear that LLM-based applications need to have reliable and uptodate knowledge sources and be able to deploy new methods and techniques that enable appropriate reasoning and grounding of the utterances in the real-world facts.
In a recent final seminar for the EU-Japan project e-VITA, a special workshop on LLMs and KG was organised to discuss especially the natural language-based context modeling and reasoning as well as new LLM and KG technologies for spoken dialogue systems, including applications (https://www.e-vita.coach/homepage/finalconference/).
This special session will continue on these lines. Focus is on the technologies, opportunities and challenges that LLMs provide for dialogue systems and the solutions that have been developed to address the notorious aspects of “hallucinations” in LLMs, and combine the knowledge needed for meaningful interaction. In particular, the special session tries to explore solutions to balance smooth conversational capabilities of LLMs with the control of and symbolic nature of knowledge bases for reliable and enjoyable interactions between humans and AI-agent.
Themes
The session aims to support discussion and fruitful exchange of ideas on the following general themes.
- The promising techniques to address the above-mentioned reliability issues include RAG (Retrieval Augmented Generation) and its deployment to generate dialogue responses by retrieving information from various types of knowledge sources such as document collections, knowledge graphs, set of images, and dynamic data streams. Papers using vector-based knowledge graphs, knowledge graph databases like neo4j, or other knowledge bases in order to explore the technical feasibility and practical solutions of the approach are welcome.
- There is also a need to include knowledge about conversation, domain, environment, users, etc. in the dialogue processing and reasoning. In this respect, the special session welcomes contributions that develop efficient solutions for how relevant knowledge can be stored, represented, and deployed in interactive applications.
- Issues on data collection will also be addressed in terms of sample dialogues for model training as well as using and re-using existing knowledge sources as reliable information sources for the system to conduct conversations with the user. The knowledge includes structured knowledge bases (databases, knowledge graphs) and document collections, and we also welcome explorations on the use of environmental information through sensors.
- Finally, important aspects concern ethical issues, including not only user information and privacy considerations, but diversity and sustainability of the models for training, adaptation, and modularisation. The papers may also address novel evaluation methods and frameworks for the assessment of the accuracy and appropriateness of the models in practical applications and tasks relevant for users and society in which the models can be used.
We seek papers on the successful and fruitful experiments and results that address these topics and bring in a fresh view of integrating the two trends in language and dialogue modelling: smoothness of conversation and trustworthiness of knowledge graphs. We encourage the authors to explore the research area for ground-breaking research on new methods and techniques in order to enhance conversations with reliable and uptodate knowledge, and to share their experience on LLMs and knowledge graphs for various dialogue modelling tasks thus paving way to new standards in storing and representing knowledge for its efficient use in dialogue management.
Paper format and submission
The special session papers will be handled by the special session organisers with the help of expert reviewers. The submissions will undergo the same review process as regular papers and if accepted, they will appear in the SIGDIAL proceedings.
For the required paper format and submission details, please consult the SIGDial main website: https://2024.sigdial.org/call-for-papers/
The deadline for submissions is May 17.
Organisers
- Kristiina Jokinen
AIRC, National Institute of Advanced Industrial Science and Technology, Japan
E-mail: kristiina.jokinen@aist.go.jp - María Inés Torres
University of the Basque Country UPV/EHU, Spain
E-mail: manes.torres@ehu.eus - Michael McTear
Ulster University, UK
E-mail: mf.mctear@ulster.ac.uk - Gérard Chollet
SAMOVAR, Institut Polytechnique de Paris, IMT-TSP, France
E-mail: gerard.chollet@telecom-sudparis.eu - Jérôme Boudy
Polytechnique de Paris, IMT-TSP, France
E-mail: jerome.boudy@telecom-sudparis.eu - Hugues Sansen
Institut Polytechnique de Paris, IMT-TSP, France
E-mail: hugues.sansen@telecom-sudparis.eu