Semi-automatic sentiment on international news

Course project IKT456 — reduce labelling cost while keeping a deployable NLP workflow.

BERT
Active learning
Label Studio
Python
Map UI

Problem

News volume is high; fully manual sentiment coding does not scale. Stakeholders still need trustworthy labels and a way to explore geography and tone together.

Pipeline

Ingest and preprocess articles; surface uncertain examples for human review.
Active learning: around ~100 carefully chosen labelled examples reached strong classifier behaviour (per project report — exact numbers in thesis/course delivery).
Model: BERT-based classifier integrated with Label Studio for human-in-the-loop iteration.
Presentation: interactive map so non-technical readers can explore sentiment spatially.

Illustration placeholder Screenshot of Label Studio session, confusion matrix, or the map UI when you want to visualise the story for visitors.

Takeaway

Shows the full arc from raw text to human-in-the-loop ML to a stakeholder-facing surface — typical of applied NLP in product and analytics teams.

← Back to project overview