Its first version (Blabinha 1.0) was implemented as a game in a social robot. Built using imperative programming in Java, its artificial intelligence was limited to text-to-speech and speech-to-text services available on the Robios robot (https://www.humanrobotics.ai/).
Its second version (Blabinha 2.0) was implemented as an instance of a goal-oriented chat-like system. In this case, with interaction via chat (text), it uses OpenAI models and prompt engineering as the core of its reasoning. Since it is a goal-oriented system, it has some objectives to achieve during the conversation: introduction and contextualization of the dialogue; persuasion and engagement; scope restriction of interaction; use of appropriate language (for conversing with children); formulation of multiple-choice questions; and topic analysis. To learn more about Blabinha 2.0, you can access the conversation logs, the evaluation form used to evaluate it, and the results of this evaluation, all available in this directory.
An article about the Goal-Oriented Chat-Like System instantiated with Blabinha has been accepted at the Encontro Nacional de Inteligência Artificial (2024).
@inproceedings{teodoro2024,
author = {Guilherme Sebastião Teodoro Junior, Sarajane Marques Peres, Marcelo Fantinato, Anarosa Alves, Franco Brandão},
title = {A Goal-Oriented Chat-Like System for Evaluation of Large Language Models},
booktitle = {Anais do XX Encontro Nacional de Inteligência Artificial e Computacional},
year = {2024},
publisher = {SBC},
address = {Porto Alegre, RS, Brasil},
note = {accept_to}
}
For more information, contact [email protected] or [email protected], or access https://sites.usp.br/keml/.