Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Preference-Aligned Data Synthesis for Fine-Grained Intent Detection

AbstractFine-grained intent detection becomes brittle as label spaces expand with many semantically adjacent intents. Although large language models (LLMs) can synthesize labeled utterances to reduce annotation cost, the resulting pseudo data are often ambiguous across intents and repetitive within intents. Existing synthesis pipelines mainly rely on prompt engineering and filtering, without updating generator parameters. We propose a preference-aligned synthesis framework that automatically mines supervision to optimize the generator. Each intent is represented by multi-centroid prototypes, and block-conditioned generation supports diversity-aware candidate construction. From these candidates, we build preferred blocks that are both prototype-consistent and diverse, and rejected blocks composed of confusable negatives from neighboring intents. These preference pairs are used to optimize the generator with Direct Preference Optimization (DPO), together with iterative refresh to mine harder confusions over time. Experiments show consistent gains for downstream intent classification in zero-shot and few-shot settings, while remaining competitive in fully supervised settings. We further observe improved generator reliability and strong transfer across multiple backbone LLM families.

More from EMNLP 2026

All EMNLP 2026 papers →