NLP Data Annotation Pipeline

Manual labelling was slow and inconsistent, and there was no way to measure label quality before data reached training.

• Built ingestion and normalisation scripts to deduplicate and segment raw text.
• Used a baseline classifier to pre-label items, leaving humans to confirm or correct.
• Added blind gold-standard items to every batch to monitor annotator accuracy continuously.
• Automated export with schema validation so malformed data never reached training.

• Throughput roughly doubled with pre-labelling in place.
• Gold-standard accuracy held above 95% across batches.
• Downstream model F1 improved by 6 points on the same architecture.