| Abstrakt: |
The use of chatbots, or conversational assistants, is increasingly present in several areas of activity. In the governmental sphere, the use of these solutions provides the improvement of the provision of public services, as well as the reduction of their costs. However, the diversity of subjects under the responsibility of the same public body presents itself as a challenge to the development of a chatbot. This is because the more subjects to be dealt with, the more likely the chatbot is ambiguous in their responses. To solve this problem, this work proposes the use of text classification techniques, to enable a tree-based architecture of conversational assistants. In this architecture, a general knowledge chatbot can identify the subject of the user’s message and forward it to specific chatbots. The feasibility of the proposal was evaluated from the performance of experiments, which compared three implementation strategies for text classification: multiclass classifier, binaries classifiers and one-class classifier. The results of the experiments showed a better performance, in relation to the F1-score and accuracy metrics, of the multiclass and binary classifiers. However, it was concluded that the use of one-class classifiers greater benefits to the hierarchical chatbots, because it allows independence between the models created, which results in greater scalability and simplicity of the proposed architecture. [ABSTRACT FROM AUTHOR] |
