Ethical Framing Analyzer
A natural language processing tool for detecting and classifying the ethical frameworks implicitly used in public communications, policy documents, and institutional justifications.
Research Question
What ethical frameworks are implicitly used in policy justifications?
Data Source
Government press releases, policy documents
Limitations
Binary frame classification, limited to Western ethical traditions
Demo not available
Note: This tool is currently in development and not yet available for use. The documentation below describes the planned functionality. Expected availability: Q3 2026.
Overview
The Ethical Framing Analyzer (EFA) is a research tool that identifies the implicit ethical frameworks operating within public communications. Policy justifications are never ethically neutral — they invoke principles of fairness, utility, rights, care, or sanctity, often without making these commitments explicit. The EFA surfaces these hidden ethical architectures, enabling researchers and practitioners to examine the moral logic underlying public discourse.
The tool is grounded in moral foundations theory and classical ethical taxonomy. It classifies text segments according to their alignment with major ethical traditions: consequentialist reasoning (focused on outcomes and utility), deontological reasoning (focused on duties and rights), virtue ethics (focused on character and excellence), care ethics (focused on relationships and responsibility), and justice-as-fairness frameworks (focused on equity and procedural legitimacy).
Technical Architecture
The EFA is built on a fine-tuned transformer model trained on a curated dataset of ethically annotated policy texts. Input documents are segmented into argument units, each of which is classified by its dominant ethical frame. The system provides both per-segment classifications and document-level ethical profiles showing the distribution of frames across an entire text.
The tool is served through a FastAPI backend, designed for integration with other DuMonde lab tools or for standalone use via API calls. Response payloads include confidence scores, frame distributions, and highlighted text passages that triggered each classification.
Limitations and Future Work
The current model operates with binary frame classification — each segment is assigned a single dominant frame — which does not capture the ethical complexity of texts that blend multiple traditions. Furthermore, the training corpus draws primarily from Western ethical traditions, limiting the tool's applicability in non-Western governance contexts. Planned improvements include multi-label classification, expanded ethical taxonomies incorporating Ubuntu, Confucian, and Islamic ethical traditions, and a user-facing annotation interface for community-driven model refinement.