API and Demo
Run the demo
Enter an suspicious sentence in the text box.
- Some documents you may want to use for test: https://www.are.na/melodrama-m/un-naturallanguage
Click Submit.
The predicted category and the confidence score will be displayed.
Links for downloading Models and Dataset
Technical Description
This system is built on a pretrained BERT model, fine-tuned specifically for a multi-label classification task on a custom-labeled dataset.
It classifies sentences such as pro-growth, ecologically dominating and neutral.
The model focuses on identifying ecology-, management-, economics- related etc. phrases and hidden ideologies based on actions (e.g., development, extraction, expansion), perspectives (e.g., capitalism, domination, greenwash), and priorities (e.g., economic growth).
The system includes four primary phases: data collection and processing, labeling, training and interpretive output generation.
Data and labeling are essential to this project.
Public documents from the World Bank were collected, cleaned, and segmented into sentences using the NLTK tokenizer
in Python. Labeling was conducted on the Refinery platform using a hybrid, iterative human–machine workflow.
Approximately 50% of the dataset was manually labelled to establish consistent criteria grounded in degrowth values.
Sentences were marked as:
- Pro-growth: language that reflects a growth-oriented mindset, prioritizes economic expansion, or treats ecological and social limits as solvable via technological or market-based interventions.
- Ecologically dominating: language that frames nature through anthropocentric or managerial logics — treating ecosystems as resources to be controlled, extracted, or optimized
These annotations informed rule-based heuristics that assisted in labeling the remaining data. The process was recursive: machine-generated labels were reviewed, refined, and selectively overwritten. In total, approximately 5000 sentences were labeled. The artist re-entered the loop multiple times to incorporate edge cases and update definitions— treating labeling as a critical intervention, not a neutral task.