Boosting 5G Complaint Intent Detection with Large-Model-Enhanced Few-Shot Learning

In the 5G era, network operators face a growing challenge: user‑complaint tickets are highly imbalanced, with a few frequent categories and many rare ones, making intent classification difficult for small NLP models that rely on abundant labeled data.

The authors propose a novel large‑model‑enhanced few‑shot learning framework. A large language model (LLM) iteratively creates high‑quality synthetic training examples for under‑represented complaint types, while a compact student model evaluates the generated data and feeds quality feedback back to the LLM. This co‑training loop continuously refines the synthetic samples, enabling the student model to achieve superior performance on scarce categories.

Technical Foundations

The paper reviews three mainstream few‑shot techniques—meta‑learning, data augmentation, and transfer learning—and explains why they alone are insufficient for telecom‑specific complaint data, which exhibits a long‑tail distribution and domain‑specific terminology.

Building on recent advances in LLMs, the authors integrate four key mechanisms:

Knowledge distillation: the student model learns from the teacher’s soft outputs, compressing the LLM’s knowledge.

Synthetic data generation: the LLM produces diverse, high‑fidelity complaint examples using prompt engineering (role‑play, chain‑of‑thought, few‑shot exemplars).

Parameterized knowledge transfer: selective LLM parameters are injected into the student to improve initialization for rare classes.

Multi‑party knowledge sharing: federated learning aggregates insights from multiple data owners, mitigating data scarcity and distribution shift.

Algorithm Workflow (Six Steps)

Initial training of the student model on available real complaint tickets.

Triggering LLM‑assisted augmentation when the student’s accuracy on rare categories falls below a preset threshold.

LLM generates targeted synthetic samples for the identified rare categories.

Student model undergoes a second training round using the synthetic data, keeping hyper‑parameters consistent with the first round.

The student evaluates the synthetic data, quantifying influence I(x_i) and similarity S(x_i), and returns a quality score Q(x_i) to the LLM.

The LLM refines its generation strategy based on the feedback, and steps 3‑6 repeat until performance criteria are met.

The quality scoring combines influence (impact on model confidence and variance) and similarity (cosine similarity to historical data), as illustrated in the following equations:

Experimental Setup

A real‑world dataset of 4,500 annotated complaint tickets covering eight categories was split 90/10 into training and test sets. The student model is ERNIE 3.0‑medium; LLMs used for augmentation are Qwen 2.5 14B and GLM‑4‑9B. Table 1 (below) compares model sizes and key properties.

Results and Analysis

Experiment 1: All eight categories were augmented with 200 synthetic samples each (total 1,600). The student model’s F1 score for the three rare categories rose from 0.31–0.65 to 0.71–0.72, a relative improvement of up to 40.66%. Overall accuracy across all categories increased by 9%.

Experiment 2: Only the three rare categories were augmented with varying synthetic sample counts (50‑1,000). Performance peaked at 500 samples for Qwen 2.5 14B (F1 = 0.76, 0.85, 0.90) and at 300‑400 samples for GLM‑4‑9B (F1 ≈ 0.62‑0.84). Excessive synthetic data caused slight degradation, likely due to noise or over‑fitting.

These findings confirm that LLM‑driven synthetic data can substantially alleviate few‑shot scarcity while preserving or enhancing overall model robustness.

Conclusion and Future Directions

The proposed LLM‑enhanced few‑shot framework effectively bridges the data gap in telecom complaint handling, delivering a 21% lift for rare categories and a 9% overall gain. Future work will explore finer‑grained sample generation strategies, model interpretability for synthetic data, and efficiency optimizations such as distillation and pruning to meet real‑time deployment constraints.

References

OUYANG Y et al., “Next decade of telecommunications artificial intelligence,” CAAI AI Research, 2022.

ZENG W et al., “Research on intelligent complaint handling scheme of 5G users based on machine learning,” 2021.

JIANG Y et al., “Exploring intelligent processing of communication complaints based on the BERT algorithm,” 2024.

AsiaInfo & Tsinghua University, “AIGC (GPT‑4) empowering telecom sector,” 2023.

FINN C et al., “Model‑agnostic meta‑learning for fast adaptation of deep networks,” arXiv, 2017.

ZHANG H X et al., “Prompt‑based meta‑learning for few‑shot text classification,” ACL, 2022.

EDWARDS A et al., “Guiding generative language models for data augmentation in few‑shot text classification,” arXiv, 2021.

BAYER M et al., “A survey on data augmentation for text classification,” ACM Computing Surveys, 2023.

GAO J et al., “A hybrid model for few‑shot text classification using transfer and meta‑learning,” arXiv, 2025.

BRAGG J et al., “FLEX: unifying evaluation for few‑shot NLP,” arXiv, 2021.

AGARWAL R et al., “On‑policy distillation of language models: learning from self‑generated mistakes,” arXiv, 2024.

LIANG C et al., “Less is more: task‑aware layer‑wise distillation for language model compression,” ICML, 2023.

ZOU T et al., “Fusegen: PLM fusion for data‑generation based zero‑shot learning,” arXiv, 2024.

YAO Z W et al., “ZeroQuant: efficient and affordable post‑training quantization for large‑scale transformers,” NeurIPS, 2022.

CHENG Y et al., “GFL: federated learning on non‑IID data via privacy‑preserving synthetic data,” PerCom, 2023.

ZHANG Z et al., “FedPETuning: when federated learning meets the parameter‑efficient tuning methods of pre‑trained language models,” ACL Findings, 2023.