Planejamento não-determinístico para o gerenciamento do agente de diálogo plantão coronavírus

• Ano de defesa: 2024
• Autor(a) principal: Pinho, Bruno da Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Coronavírus; Planejamento automatizado; Algoritmo não determinístico; CNPQ: CIÊNCIAS EXATAS E DA TERRA: CIÊNCIA DA COMPUTAÇÃO
• Link de acesso: http://repositorio.ufc.br/handle/riufc/77500

Natural conversation is a fundamental aspect of intelligent behavior, and for this reason, dialogue agents have been an important research topic in Artificial Intelligence (AI) for decades. Chatbots are implementations of these systems and have been widely adopted, particularly during the recent COVID-19 pandemic. As an example, we have Plantão Coronavírus dialogue agent, which was used by the Government of the State of Ceará to assist in patient care during critical moments of the pandemic. In general, the tasks performed by a dialogue agent are: (i) understanding the elements of the user input, (ii) based on these elements, selecting an action to be taken, and (iii) generating a response to the user based on this action. Although there are well-established methods for natural language understanding, based on statistical reasoning, such as machine learning algorithms, dialogue management poses challenges for this kind of reasoning, once they do not guarantee the reliability of responses or provide explanations about the reasoning process. These requirements, however, are very important for critical systems, such as those related to healthcare. Therefore, the use of well-established symbolic reasoning methods in AI has been explored in recent years for dialogue management tasks, such as Automated Planning, which is an approach to reasoning about actions. In Planning, the agent’s actions can be deterministic or have uncertain effects, i.e., non-deterministic actions. This work presents a dialogue manager for Plantão Coronavírus based on the approach of Automated Planning with Non-Deterministic Actions, which models the different possibilities of user responses. The main contributions of this work are: (i) the acquisition of the non-deterministic planning domain, based on dialogues between patients and the bot; (ii) the system architecture that combines machine learning for natural language understanding tasks and automated planning for dialogue management tasks; (iii) the implementation of the proposed architecture using Rasa as a conversational system and a state-of-the-art non-deterministic planner for automatic generation of possible dialogue paths; and (iv) the evaluation conducted by simulating whether the solutions obtained by the manager could effectively conduct the dialogues that occurred in the system.