An art software identifies potential ecological threats in documents.
(Un)Natural Language
(Un)Natural Language is an art software and an interactive online archive that examines and documents how words make worlds, identifying the potential ecological threats in government project documents. The project weaves together ecology, linguistics, and computational technology to offer a new analytical lens on language use in urbanization. Through a custom-labeled dataset and a fine-tuned language model, the system classifies individual sentences in project appraisal documents, returning visualized analyses that reveal patterns of pro-growth and ecologically dominating discourse. Merging net art, environmental activism, and natural language processing, (Un)Natural Language offers a novel framework for interpreting public documents, through a degrowth lens, uncovering hidden narratives of extractivism and economic expansion.
Credit
Artist: Mou Peijing
Developer(Data & ML): Mou Peijing
Developer(Website): Chenyue xdd Dai
Techniques
This system includes three primary phases: data processing, labeling, and training. Public documents from The World Bank were collected, and segmented into sentences. Labeling was conducted by both manual annotation and heuristic labeling. The labeled dataset was exported to fine-tune a customized BERT model, modified with additional dropout and dense layers for multi-label text classification. Once trained, the model assigns category labels to new input sentences, and returns a confidence score for each labeled sentence based on its latent embeddings for visualization.
Lexicon