Probabilistic Model for Contextual Retrieval

Ji-Rong Wen (Microsoft Research Asia), Ni Lao (Tsinghua University), Wei-Ying Ma (Microsoft Research Asia).

Key Contribution

First comprehensive probabilistic retrieval model addressing three critical challenges in contextual retrieval: incompatible context, noisy context, and incomplete queries.

Applications

• Context-aware mobile search — Exploits user location, time, device state for queries like “bus” to suggest relevant routes
• PC troubleshooting — Uses system state information as context when customer descriptions are vague
• Personalized search — Query log mining for user behavior patterns

Key Insights

• Query logs are the key to building effective contextual retrieval models
• Customer problem descriptions alone are insufficient — system state provides critical diagnostic context
• Probabilistic framework handles real-world noise and incompleteness gracefully

Connection to Vault

Early Ni Lao work at Microsoft Research before CMU. Contextual retrieval concepts later evolved into the Path Ranking Algorithm for knowledge base reasoning. The PC troubleshooting application connects to Automated Fault Diagnosis.