Desarrollo de un Asistente Inteligente  para la atención al cliente y solución  técnica en empresas del sector del agua  mediante modelos preentrenados

Fernández Salcedo, Antonio

Desarrollo de un Asistente Inteligente para la atención al cliente y solución técnica en empresas del sector del agua mediante modelos preentrenados

Abstract

The project addresses the development and evaluation of an intelligent assistant based on pre-trained natural language models. Its goal is to create a system capable of providing accurate answers and improving user support for Canal de Isabel II through artificial intelligence and NLP techniques. The methodology follows a cascading approach, involving progressive data collection and continuous analysis of labeled question-and-answer datasets from Canal de Isabel II. The BERT model is selected, fine-tuned, and trained using specific techniques. Additionally, a confidence-based filtering mechanism is implemented to enhance the quality of generated responses. For making inferences, the Mixtral model is chosen due to its ability to follow specific instructions and guide Canal de Isabel II customers or technicians in problem-solving and task execution. The interface design utilizes the Hugging Face framework. Performance evaluation of the models employs various lexical metrics. Furthermore, the capacity of different LLM models to provide responses to specific queries from Canal de Isabel II users is analyzed and compared. In conclusion, the project lays the groundwork for an effective intelligent assistant in user support for Canal de Isabel II. Besides, future work with additional adjustments and improvements is proposed to reach its full potential in terms of accuracy and quality of responses.

Citation

Fernández Salcedo, A. (2024). Desarrollo de un Asistente Inteligente para la atención al cliente y solución técnica en empresas del sector del agua mediante modelos preentrenados. Universitat Oberta de Catalunya (UOC). https://hdl.handle.net/10609/150520

This citation was generated automatically.

Full item page