Nina van Diermen - BERTopic to accelerate Ukrainian aid by the Red Cross

BERTopic was implemented to help the Red Cross process Ukrainian refugee social media messages, reducing analysis time from 20 hours to 2 minutes

Key requirements for the system:

• Quick initialization for emergency response
• Dynamic adaptation to changing situations
• Ability to handle unstructured text data

Technical implementation includes:

• Transformation of messages into numerical embeddings
• Dimensionality reduction using UMAP
• Clustering with HDBSCAN
• Topic label generation based on word frequency

System benefits:

• No predefined labels required
• Handles outliers effectively
• Provides topic hierarchies
• Can generate topic summaries using OpenAI
• Allows tracking topic evolution over time

Evaluation methods:

• Topic coherence measurement
• Density-based cluster validation
• Soft reformulation accuracy
• Human interpretability of results

Customization options:

• Replaceable embedding models
• Adjustable clustering methods
• Configurable vectorizer settings
• Optional stop word removal
• Custom summary generation

Practical applications:

• Social media monitoring
• Self-help FAQ generation
• Aid resource allocation
• Trend analysis
• Document categorization