Reconhecimento e análise de entidades nomeadas em textos não-estruturados usando LLMs e redes complexas

Abstract

The availability of textual data on the internet has been growing exponentially. While access to information was a major challenge in the past, today, the vast amount of available content requires increasingly sophisticated methods to automate information extraction. Whether through books, articles, emails, or social media posts, the information available in these media can contribute to generating innovative ideas, better understanding the target audience in advertising campaigns, and advancing science through literature reviews. In this way, textual data has proven to be valuable in an increasingly connected world. Named entity recognition (NER) allows identifying words in a text that refer to a common theme. Structuring the named entities in a relation graph provides a framework capable of representing the context of the dataset, enabling the identification of patterns. With the recent advances in generative models, such as Large Language Models (LLM), this study evaluated the performance of this type of model in named entity recognition, discussing the advantages and disadvantages through a quantitative analysis of the task performed, comparing ChatGPT with models like BERT and BiLSTM. Additionally, a qualitative analysis seeks to evaluate the capabilities of the BERT and ChatGPT models in generating relation graphs between named entities, as well as proposing exploration techniques for this type of complex network. Thus, the aim of this study was to present a methodology for generating representations in complex networks of textual data. The presented methodology allows two approaches: the use of LLM models, such as ChatGPT, through prompt engineering, requiring a minimal number of examples, and the training or fine-tuning of state-of-the-art models, such as BERT.