DuIVRS-2: LLM-based IVR System for POI Data at Baidu Maps
At Baidu Maps, researchers have introduced DuIVRS-2, an end-to-end framework utilizing a large language model (LLM) for extensive Point of Interest (POI) attribute gathering. This innovative system tackles the issues of error buildup and the significant maintenance demands associated with conventional modular Interactive Voice Response (IVR) systems. It employs a finite state machine (FSM)-based data augmentation approach to create balanced training datasets, incorporates a selective generation method featuring a Chain-of-Thought (CoT) mechanism to enhance output stability and reduce hallucinations, and implements a collaborative iterative learning framework with dual evaluation for ongoing policy improvement.
Key facts
- DuIVRS-2 is an LLM-based end-to-end framework for POI attribute acquisition.
- It is designed for large-scale deployment at Baidu Maps.
- Traditional modular IVR systems suffer from error accumulation and high maintenance overhead.
- The methodology uses FSM-guided data augmentation to synthesize balanced training data.
- A selective generation scheme with CoT mechanism ensures output stability and eliminates hallucinations.
- A cooperative iterative learning framework with dual evaluation enables continuous policy refinement.
- The system targets the long-tail distribution of real-world interactions.
- The work is published on arXiv with ID 2605.17900.
Entities
Institutions
- Baidu Maps