Few-Shot Learning Methods and Applications in Meituan NLP

Meituan’s NLP services involve many tasks that require large amounts of labeled data, leading to high annotation costs. Few‑shot learning aims to achieve good performance with limited data.

We review several few‑shot techniques: data augmentation (sample and embedding augmentation, EDA, back‑translation, Mixup, adversarial training), semi‑supervised learning (consistency regularization, entropy minimization, methods such as Temporal Ensembling, Mean Teacher, VAT, MixMatch, MixText, UDA), ensemble + self‑training, and domain adaptation (MAML, Meta‑Baseline).

Experimental results on Meituan business benchmarks and public Chinese datasets show that embedding‑based augmentation improves accuracy by ~1 pp, semi‑supervised methods add 1.5‑2 pp, and ensemble/self‑training yields similar gains. Active learning reduces the amount of data needed, achieving the performance of 1000 random samples with only 500 actively selected examples.

Practical deployments include medical‑beauty topic classification, travel‑guide detection, medical‑beauty effect labeling, brand labeling, fake‑review detection, POI sentiment analysis, and internal text classification, all achieving high accuracy with a few thousand labeled instances.

Future work focuses on improving existing models, exploring more domain‑transfer techniques, extending experiments to MRC and NER, building a unified few‑shot learning platform, and training a general model for various text tasks.