Linguistically grounded NLP for gender-based violence identification in clinical records

Abstract

AI systems have evolved drastically in the last decade. Much of this progress is visible in their ability to perform a wide range of tasks using language as the primary medium of interaction with users. Interestingly, however, the dominant frameworks used to train these systems are often disconnected from linguistic theories of language and learning. When they fail, the limitations are framed as a matter of training data availability. This thesis argues that more training data is not the only solution and that integrating linguistically grounded resources into the components of core natural language processing (NLP) pipelines can significantly improve performance on semantic-oriented tasks – and tests this hypothesis in Brazilian Portuguese (BP), a language for which NLP performance remains limited despite substantial available data. The application context is the identification of gender-based violence (GBV) in clinical notes from primary care. According to the World Health Organization, one in three women worldwide experience some form of violence in their lifetime, and these experiences are often reflected in health records. This domain is a demanding test bed: clinical notes are noisy and domain-specific, involve sensitive personal information that must be anonymized before any analysis can proceed, and require representations beyond surface-level text classification. Two case studies are presented. The first evaluates named entity recognition (NER) in the anonymization of clinical notes, demonstrating that combining NER with rule-based and lexical resources substantially outperforms a NER-only approach: different categories of personal information call for different kinds of recognition. The second integrates a FrameNet-based semantic parser as a new component into the core NLP pipeline and evaluates it through semantic pattern querying and supervised classification applied to GBV identification. Semantic pattern queries identified 23.8% more cases of violence than those present in official notification systems, with 92.2% of matched cases having no formal report. The semantic classifier built on frame-based features also vastly outperformed both a large sentence embedding baseline and a classifier relying on demographic and structured health record data. The latter broadly flags demographic profiles associated with reported cases rather than specific signals of violence, effectively encoding the structural disparities it is meant to help address – a limitation the frame-based classifier avoids by grounding predictions in the actual textual content of clinical notes. Such systems are not only more accurate and interpretable through human-readable labels – they are also more resource-efficient, with practical implications for environmental sustainability by reducing the energy demands of large-scale AI models. Ultimately, this work demonstrates that bridging computational methods with linguistically motivated representations can make AI systems more accurate, more interpretable, and more ethically aligned with sensitive real-world applications.