Assessing NER tools for dialogue data anonymization

• Autor(a) principal: Pereira, Miguel Alexandre da Silva Sarmento Falco
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10071/30399
• topic: Data anonymization; Entities extraction; Processamento de linguagem natural - -- NLP Natural language processing; Artificial; Named entity recognition; Sensitive data; Anonimização de dados; Extração de entidades; Reconhecimento de entidades mencionadas; Dados sensíveis

As the number of organizations processing sensitive data grows, so does the need for businesses to protect and ensure the privacy of their customers. However, the prevailing methods for protecting sensitive data often involve manual or semi-automatic procedures, which can be resource-intensive and error-prone. This dissertation addresses data anonymization by focusing on Named Entity Recognition (NER) models. Particularly, we investigate and compare various NER models for the Portuguese language to automatically and effectively anonymize unstructured data. The models SpaCy, STRING, WikiNEuRal and RoBERTta are used in the machine learning approach with the goal of identifying classes such as Person, Location, and Organization. On the other hand, the rule-based approach seeks to identify classifications such as NIF, Email, Car Plate and even Postal Code. Additionally, it was created a Flask API tool capable of processing unstructured data and anonymizing it, more specifically, given a string that simulates a message, automatically anonymize the message content that might be considered as sensitive. This tool combines many techniques for identifying and extracting mentioned entities for the Portuguese language, based on rule models and machine learning. The combination of both rule-based and machine learning models in the same tool was crucial to enable the ability to encompass more sensitive classes for anonymization. The results calculated for the extraction of entities from the tool built in this work encompasses the results for the three classes calculated with the SpaCy model, with the addition of the results calculated for the rule-models created.