ETD

• algorithm
• article
• femicide
• framing
• perpetrator justification
• tone
• victim blaming

This thesis proposes an automated and interpretable framework for the analysis of media framing
in Italian news articles on femicide, combining linguistically grounded rule-based indicators with
supervised learning for validation.
The framework operationalizes three key framing dimensions recurrent in journalistic guide
lines and prior literature: victim blaming, perpetrator justification, and article tone. These
dimensions are modeled through syntactic and semantic indicators derived from lightweight lin
guistic preprocessing and interpretable lexical patterns. To avoid dependence on manual ground
truth, labels are induced via a conservative multi-engine consensus strategy inspired by weak
supervision principles.
The reliability of the proposed indicators is assessed through inter-source agreement and
learnability analyses, comparing supervised models trained on automatically induced labels with
models trained on a manually annotated subset. Both linear TF–IDF classifiers and transformer
based models are evaluated, while post-hoc explainability methods are used to inspect linguistic
signals supporting model decisions. Results show that the proposed framework captures stable
and meaningful framing patterns at corpus level, supporting scalable and reproducible analysis
of femicide framing in Italian journalism.