Business Context
TechSupport, a software company handling customer inquiries, receives around 100,000 support tickets daily. Currently, support agents manually categorize tickets, leading to delays in response times and inefficient resource allocation. The goal is to implement an NLP model to automatically classify incoming tickets into predefined issue categories to streamline the triage process and improve response efficiency.
Data Characteristics
- Volume: 2 million labeled support tickets (last 12 months)
- Text length: 10-300 words (median: 50 words)
- Language: English (100%)
- Label distribution:
- Technical Issue: 40%
- Billing: 30%
- Account Management: 20%
- Feedback: 5%
- Other: 5%
Success Criteria
- Achieve ≥90% precision on the Technical Issue category (critical for customer satisfaction).
- Maintain ≥85% macro-F1 score across all categories.
- Inference latency < 200ms per ticket (for real-time processing).
- Handle class imbalance effectively during training.
Constraints
- The model must comply with data privacy regulations (GDPR).
- Must be deployable on a cloud service with limited computational resources.
