Zero-Shot Image Moderation in Google Ads with LLM-Assisted Textual Descriptions and Cross-modal Co-embeddings
Enming Luo, Wei Qiao, Katie Warren, Jingxiang Li, Eric Xiao, Krishna Viswanathan, Yuan Wang, Yintao Liu, Jimin Li, Ariel Fuxman
TL;DR
This work tackles the scalability challenge of moderating vast and evolving Google Ads imagery by introducing a zero-shot, cross-modal moderation framework. It combines LLM-assisted generation of textual policy descriptions with domain-expert validation and cross-modal co-embeddings to detect policy violations without large labeled datasets. The pipeline comprises three components: generating textual descriptions, validating them, and enforcing policies via a shared embedding space with automated decisions, LLM-assisted review, and human escalation when needed. The approach demonstrates strong precision and coverage in tobacco-image experiments and offers practical benefits, including minimal training data, fast iteration, and reuse across multiple policies, enabling scalable and adaptable ads moderation.
Abstract
We present a scalable and agile approach for ads image content moderation at Google, addressing the challenges of moderating massive volumes of ads with diverse content and evolving policies. The proposed method utilizes human-curated textual descriptions and cross-modal text-image co-embeddings to enable zero-shot classification of policy violating ads images, bypassing the need for extensive supervised training data and human labeling. By leveraging large language models (LLMs) and user expertise, the system generates and refines a comprehensive set of textual descriptions representing policy guidelines. During inference, co-embedding similarity between incoming images and the textual descriptions serves as a reliable signal for policy violation detection, enabling efficient and adaptable ads content moderation. Evaluation results demonstrate the efficacy of this framework in significantly boosting the detection of policy violating content.