wwf-seaweed-public-sentiment

This repository contains code and model artifacts for automated classification of global news articles related to seaweed aquaculture. It includes scripts to ingest headlines and full-text articles from PDFs and Word documents into a SQLite database, and a set of transformer-based classifiers built with Hugging Face Transformers. A microsoft/deberta-v3-base model is fine-tuned for three-class headline sentiment analysis (negative, neutral, positive), while two Longformer-based models perform two-stage article body classification: (1) relevance filtering (relevant vs. irrelevant to seaweed aquaculture) and (2) subject classification (seaweed vs. other forms of aquaculture). The body classifiers are deployed as GPU-backed pipelines with text standardization, deduplication, and a simple regular-expression heuristic to correct misclassified seaweed-focused articles. Training follows a supervised setup with held-out validation, accuracy and weighted F1 monitoring, and checkpoint selection, with results logged for downstream selection and analysis. Together, these components enable large-scale, reproducible filtering and labeling of media coverage on seaweed aquaculture for research and reporting.