Estudo comparativo entre modelos transformers aplicados ao desenvolvimento de chatbots

Abstract

Chatbots are software that use natural language to communicate with their users. Transformers models can be used to create them, a type of neural network that has been showing promising results in areas of study such as natural language processing (NLP). The variety of existing models poses challenges in selecting the best options for the niche market in which the chatbot will operate. This work presents a comparative study between the transformer models DIET, LaBSE, BERTimbau, DistilBERT, RoBERTa, GPT, GPT-2 and XLNET applied to the development of a chatbot submitted to data captured in public services carried out by the Court of Auditors of the State of Rio Grande do Norte (TCE/RN). The study considered metrics such as accuracy and precision, in addition to training and response processing times. DIET was considered the best model for managing to balance good results in quality and time metrics together, standing out in the time factor for responding up to 2.5 times faster than the others.