Applying Large Language Models to Advertising Copy Generation

In modern advertising systems, copy is essential. With rapid advances in NLP, advertisers increasingly use AI-powered copy generation tools, creating new opportunities for large language models (LLMs) in this domain. This article explores the business background, challenges, and potential improvements brought by LLMs.

The advertising copy workflow faces three main challenges: (1) a wide variety of copy types across different ad formats, (2) the need for rich, creative content that aligns with product information, and (3) emerging demands from short‑video and other new media formats. A unified LLM can reduce maintenance costs and enhance creativity.

Model selection involved evaluating both objective metrics (format compliance, length control, business‑specific differentiation) and subjective metrics (fluency, elegance, product relevance). Objective metrics are measured automatically, while subjective metrics were initially assessed by human judges and later by GPT‑based evaluation with carefully designed prompts to avoid score flattening.

Candidate models included Chinese‑adapted LLaMA, Baichuan, ChatGLM, and QWen series. Zero‑shot prompting achieved reasonable fluency, but fine‑tuning on a curated advertising copy dataset was required for stricter length and format constraints.

Experiments showed that larger models generally improve objective scores, but subjective gains plateau beyond a certain size. Considering resource constraints, the QWen 1.5‑7B model was chosen as the production backbone.

Data preparation emphasized high‑quality, diverse text: (1) cleaning raw copy with rule‑based filters, (2) augmenting with spoken‑style video script data via ASR/OCR pipelines, and (3) generating additional data with powerful LLMs (e.g., GPT‑4) while mixing in generic instruction‑following data to preserve generalization.

Training employed full‑parameter fine‑tuning after testing LoRA, P‑Tuning, and prefix tuning, as full fine‑tuning yielded the best performance for this task.

Deployed LLMs substantially improved copy quality, increased user acceptance rates, and enabled new capabilities such as multi‑style product summaries, richer video script generation, and more diverse external‑placement titles. Business metrics showed higher click‑through and conversion rates after the upgrade.

Future directions include scaling models for personalized copy, exploring smaller, high‑quality data regimes to reduce costs, and leveraging the world knowledge embedded in LLMs for deeper product understanding and user intent modeling.