Scholarlink

We present a practical framework for improving contextual relevance in display advertising through large language model (LLM) assisted labeling and dual-encoder embeddings. Our fine-tuned annotator automatically labels large volumes of ad–context pairs, eliminating costly manual labeling while maintaining high quality. These labels train a Two-Tower BERT model that learns semantically rich embeddings for ads and contexts, enabling scalable and accurate matching. Evaluation with majority-vote LLM ground truth shows substantial gains in AUC and nDCG over generic baselines. The approach demonstrates how LLM-based weak supervision can be effectively combined with deep retrieval architectures to enhance ad relevance in production-scale systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Display Ads Contextual Relevance Modeling with LLM Labels

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chao Gan,
  • Fan Yang,
  • Fangping Huang,
  • Weijie Yuan,
  • Nahid Anwar,
  • Musen Wen,
  • Konstantin Shmakov,
  • Hong Yao,
  • Kuang-chih Lee

摘要

We present a practical framework for improving contextual relevance in display advertising through large language model (LLM) assisted labeling and dual-encoder embeddings. Our fine-tuned annotator automatically labels large volumes of ad–context pairs, eliminating costly manual labeling while maintaining high quality. These labels train a Two-Tower BERT model that learns semantically rich embeddings for ads and contexts, enabling scalable and accurate matching. Evaluation with majority-vote LLM ground truth shows substantial gains in AUC and nDCG over generic baselines. The approach demonstrates how LLM-based weak supervision can be effectively combined with deep retrieval architectures to enhance ad relevance in production-scale systems.