The paper “MODULE SELECTION: A NEW TASK FOR DIALOG SYSTEMS” [1] presents the idea of combining different dialogue engines (DE) into a single one by distributing the user inputs to the best matching DE. They propose the task idea of “module selection” to identify the dialogue engine that should handle the user input.
I think it is an interesting idea, but it is also not new. Please see the related work section of the paper for more information on existing systems that go in this direction. The difference to most previous work (aside from the “Black Box Agent Integration (BBAI) framework” [2]) is the focus on the high-level and treating the DEs as sub-modules in the form of black boxes.
The Modular Dialog System treats the underlying conversational agents as black boxes and works with
[1]
any dialog system without further adaption.
Using an MDS could increase the reusability and improve the scalability of Conversational User Interfaces. It also allows combining and experimenting with different dialogue control solutions like combining rule-based approaches, “traditional” machine learning approaches, and Large Language Models (LLMs) into one CUI.
The idea of an MDS is really broad and would include systems and use cases from A/B testing in a user study to commercial systems that could separate different customer support tasks into independently trained and managed DEs.
Module Selection Approaches
I would categorize three types of how the module selection could be realized based on the paper and add LLMs as well.
- Rule-based
- Train a classifier
- Use a LLM like ChatGPT or GPT-4
I currently recommend using an LLM based on some experiments I did. You just have to specify the task and the capabilities of the different dialogue engines and that’s it. I would suggest adding in the prompt that an explanation should be included for the module selection so that the prompt could be improved if wrong selections are encountered.
[1] https://www.dfki.de/web/forschung/projekte-publikationen/publikation/13137