China has over 1.116 billion mobile Internet users, whose demands continue to evolve rapidly. However, existing Chinese spoken language understanding (SLU) benchmarks are outdated and cover limited domains, failing to reflect real-world mobile assistant scenarios. To this end, we introduce CMPA-SLU, a Chinese SLU dataset for mobile personal assistants collected from real-world applications and annotated with rigorous quality control. CMPA-SLU contains 50 domains, 247 intents, and 74 slot types, expanding the diversity of existing Chinese SLU benchmarks. To better simulate practical deployment scenarios, we further introduce a Partial Category setting, where candidate labels are dynamically filtered using a Bidirectional Guided Embedding (BGE)-based retrieval strategy. We further adopt a label mapping strategy for joint intent and slot prediction within a single generation process. We evaluate multiple large language models (LLMs) under both Full Category and Partial Category settings. Experimental results show that the Partial Category setting improves performance by reducing the candidate label space.

Multilingual spoken language understanding (SLU) involves intent detection (ID) and slot filling (SF) across multiple languages. The inherent linguistic diversity presents significant challenges in achieving performance comparable to traditional SLU. Recent studies have attempted to improve multilingual SLU performance by sharing multilingual encoders. However, these approaches have not directly established information flow between languages. To address this, we first demonstrate the feasibility of such information transfer and pinpoint the key challenges: prediction error mitigation and multilingual slot alignment. We then propose the INformation Transfer network (INT) to tackle these challenges. The gate unit in INT controls the information flow between languages, reducing the adverse impact of prediction errors on both ID and SF. Additionally, we reformulate SF as a span prediction problem and introduce a slot-matching attention mechanism to achieve slot alignment across languages. Experimental results on the MASSIVE and MASSIVE-UG datasets show that our model outperforms all baselines in overall accuracy across all languages, and demonstrates robust performance when different languages are used as the source.