Zero-Shot Image Moderation in Google Ads with LLM-Assisted Textual Descriptions and Cross-modal Co-embeddings

TL;DR

This paper introduces a scalable zero-shot image moderation system for Google Ads that uses LLM-generated textual descriptions and cross-modal co-embeddings to detect policy violations without extensive labeled data.

Contribution

It presents a novel approach combining large language models and cross-modal embeddings for zero-shot ad image moderation, reducing reliance on labeled datasets.

Findings

Significantly improved detection of policy-violating images.

Effective use of LLMs for generating policy-representative textual descriptions.

Robust zero-shot classification performance in real-world ad 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.