Uma análise exploratória e experimental de métodos de resumo automático de texto na saúde

Abstract

Context: Fraud and corruption are among the main crimes that affect public institutions, with the health sector being particularly vulnerable due to its structural complexity, the coexistence of public and private providers, the large number of actors involved, the globalized nature of supply chains, the high financial costs and the asymmetry of information between interested parties. These factors weaken health systems, resulting in wasted resources, reduced resilience in medical emergencies and limited access to essential services. Objective: The general objective of this work is to develop and evaluate Automatic Text Summary Methods, aimed at supporting audits in the health sector. Methods: Initially, a Systematic Literature Mapping was carried out with the aim of investigating the most recent studies on text summarization in the health domain, through the analysis of the methods developed, the challenges faced, the benefits achieved and the areas of application. Then, two methods for extractive text summarization were proposed, and, in addition, we carried out experimental evaluations for the tasks of abstractive summarization, text classification and topic modeling. Results: The Mapping findings highlight the continued predominance of traditional summarization methods, such as those based on Computational Linguistics (73.91%), we also observe that 73.91% of the methods use hybrid approaches, as well as the need for more scalable approaches. Furthermore, the majority of studies approach the problem from an extractive perspective (82%), which highlights a research gap regarding abstractive methods. The wide range of applications highlights the adaptability of text summarization techniques in different medical and biomedical fields. Two unsupervised methods based on optimization and memetic algorithms were proposed, developed and experimentally evaluated for the task of automatic generic extractive multi-text document summarization, seeking to mitigate the limitations identified in the literature, such as evaluation limitations, generalization capacity, computational complexity, dependence on experts, summaries based on query and using only one document. The results of the statistical analyzes indicate that the proposed approaches are better than 16 methods from the bioinspired algorithm literature. Furthermore, experimental evaluations demonstrated the feasibility and reliability of using abstractive methods for text summarization, text classification and topic modeling to support audit activity, contributing to the reduction of information overload. Conclusion: In summary, the results obtained from statistical analyzes indicate the feasibility of using summarization, textual classification and topic modeling techniques as tools to support decision-making, contributing both to mitigating informational overload and improving the efficiency of analytical activities performed by auditors.