Applying Large Language Models to Search Advertising Satisfaction: From DNN to ERNIE and Prompt Learning

The presentation introduces how Baidu's Fengchao platform integrates large‑scale models into the search advertising satisfaction workflow, structured into four parts: an overview of search ad satisfaction, the evolution from DNN to ERNIE, prompt learning applications, and AIGC-driven imagination.

Search advertising satisfaction differs from general search relevance by requiring personalized, business‑oriented matching between user queries and advertisers' landing pages; the platform must evaluate both relevance and the quality of the advertiser's service, addressing challenges such as noisy, fragmented landing‑page content and long‑text modeling.

Moving from traditional DNN‑based CTR models to pre‑trained language models (ERNIE) involves extracting massive user log features, converting discrete IDs into dense embeddings via a sparse table, and leveraging distributed training pipelines; the shift enables end‑to‑end learning but introduces hardware, latency, and long‑text processing challenges, prompting solutions like GPU acceleration, model distillation, pruning, and specialized tokenization.

Prompt learning is employed to achieve industry isolation by assigning a fixed soft‑prompt token as an industry identifier, forcing its presence during pre‑training and fine‑tuning; this approach supports incremental learning, maintains performance across evolving industry standards, and facilitates both single‑tower and dual‑tower relevance models.

The AIGC section explores generative model uses such as automated ad‑material creation, debugging/explanation tools for advertisers, and LLM‑based reward models that enhance system‑level feedback loops, illustrating how generative AI can drive a virtuous cycle of higher‑quality ad content and better ecosystem outcomes.

The Q&A addresses practical concerns: implementation of industry‑isolated pre‑training, choice between single‑tower and dual‑tower relevance models, prioritization of token lengths, comparative effectiveness of core‑word versus multi‑level tokenization (favoring the latter), and how soft‑prompt techniques enable dual‑tower models to reuse single‑tower pre‑training.