Development of a Sentiment Analysis System for Chatbot Responses Enhanced with Fuzzy Logic

Natural Language Processing is a branch of artificial intelligence that focuses on Human-Machine Interaction and currently enables the execution of complex tasks such as text generation or sentiment analysis thanks to Transformers technology and the application of the concept of attention in language models. In this work, the pre-trained language models Mistral-7b and BERT have been used for sentiment analysis task and fuzzy logic membership functions have been employed to optimize the final classification considering a human approach. For evaluation, an input text has been used to generate a response, which undergoes an analysis of the conveyed emotion. The precision to assert opposite emotions as “Very Happy” is up to 72%, while asserting “Very Sad” is up to 92%. On the other side, the “Neutral” documents just the 32% really corresponds to a real “Neutral” emotion. These results establish that Fuzzy Logic techniques emphasize the clear options while spare other ones because of the difficult of identify any kind of emotion just like a human feeling.