How Amazon Buying makes use of Amazon Rekognition Content material Moderation to assessment dangerous pictures in product critiques

Clients are more and more turning to product critiques to make knowledgeable choices of their procuring journey, whether or not they’re buying on a regular basis gadgets like a kitchen towel or making main purchases like shopping for a automobile. These critiques have reworked into an important supply of data, enabling consumers to entry the opinions and experiences of different clients. Consequently, product critiques have grow to be an important facet of any retailer, providing useful suggestions and insights to assist inform buy choices.

Amazon has one of many largest shops with tons of of tens of millions of things obtainable. In 2022, 125 million clients contributed almost 1.5 billion critiques and rankings to Amazon shops, making on-line critiques at Amazon a stable supply of suggestions for patrons. On the scale of product critiques submitted each month, it’s important to confirm that these critiques align with Amazon Community Guidelines concerning acceptable language, phrases, movies, and pictures. This apply is in place to ensure clients obtain correct data concerning the product, and to forestall critiques from together with inappropriate language, offensive imagery, or any kind of hate speech directed in the direction of people or communities. By imposing these tips, Amazon can keep a secure and inclusive atmosphere for all clients.

Content material moderation automation permits Amazon to scale the method whereas holding excessive accuracy. It’s a posh downside area with distinctive challenges and requiring totally different strategies for textual content, pictures, and movies. Photographs are a related part of product critiques, typically offering a extra fast affect on clients than textual content. With Amazon Rekognition Content Moderation, Amazon is ready to routinely detect dangerous pictures in product critiques with larger accuracy, lowering reliance on human reviewers to reasonable such content material. Rekognition Content material Moderation has helped to enhance the well-being of human moderators and obtain vital value financial savings.

Amazon Shopping with Rekognition

Moderation with self-hosted ML fashions

The Amazon Buying workforce designed and carried out a moderation system that makes use of machine studying (ML) together with human-in-the-loop (HITL) assessment to make sure product critiques are concerning the buyer expertise with the product and don’t comprise inappropriate or dangerous content material as per the group tips. The picture moderation subsystem, as illustrated within the following diagram, utilized a number of self-hosted and self-trained laptop imaginative and prescient fashions to detect pictures that violate Amazon tips. The choice handler determines the moderation motion and offers causes for its choice primarily based on the ML fashions’ output, thereby deciding whether or not the picture required an extra assessment by a human moderator or may very well be routinely authorised or rejected.

Overall architecture

With these self-hosted ML fashions, the workforce began by automating choices on 40% of the photographs acquired as a part of the critiques and repeatedly labored on enhancing the answer by means of the years whereas dealing with a number of challenges:

  • Ongoing efforts to enhance automation fee – The workforce desired to enhance the accuracy of ML algorithms, aiming to extend the automation fee. This requires steady investments in knowledge labeling, knowledge science, and MLOps for fashions coaching and deployment.
  • System complexity – The structure complexity requires investments in MLOps to make sure the ML inference course of scales effectively to fulfill the rising content material submission visitors.

Change self-hosted ML fashions with the Rekognition Content material Moderation API

Amazon Rekognition is a managed synthetic intelligence (AI) service that gives pre-trained fashions by means of an API interface for image and video moderation. It has been extensively adopted by industries resembling ecommerce, social media, gaming, on-line courting apps, and others to reasonable user-generated content material (UGC). This features a vary of content material varieties, resembling product critiques, consumer profiles, and social media put up moderation.

Rekognition Content material Moderation automates and streamlines picture and video moderation workflows with out requiring ML expertise. Amazon Rekognition clients can course of tens of millions of pictures and movies, effectively detecting inappropriate or undesirable content material, with absolutely managed APIs and customizable moderation guidelines to maintain customers secure and the enterprise compliant.

The workforce efficiently migrated a subset of self-managed ML fashions within the picture moderation system for nudity and never secure for work (NSFW) content material detection to the Amazon Rekognition Detect Moderation API, profiting from the extremely correct and complete pre-trained moderation fashions. With the excessive accuracy of Amazon Rekognition, the workforce has been capable of automate extra choices, save prices, and simplify their system structure.

Deployment diagram

Improved accuracy and expanded moderation classes

The implementation of the Amazon Rekognition image moderation API has resulted in larger accuracy for detection of inappropriate content material. This means that an extra approximate of 1 million pictures per yr can be routinely moderated with out the necessity for any human assessment.

Operational excellence

The Amazon Buying workforce was capable of simplify the system structure, lowering the operational effort required to handle and keep the system. This strategy has saved them months of DevOps effort per yr, which implies they will now allocate their time to growing progressive options as an alternative of spending it on operational duties.

Value discount

The excessive accuracy from Rekognition Content material Moderation has enabled the workforce to ship fewer pictures for human assessment, together with probably inappropriate content material. This has lowered the fee related to human moderation and allowed moderators to focus their efforts on extra high-value enterprise duties. Mixed with the DevOps effectivity good points, the Amazon Buying workforce achieved vital value financial savings.


Migrating from self-hosted ML fashions to the Amazon Rekognition Moderation API for product assessment moderation can present many advantages for companies, together with vital value financial savings. By automating the moderation course of, on-line shops can shortly and precisely reasonable giant volumes of product critiques, enhancing the shopper expertise by making certain that inappropriate or spam content material is shortly eliminated. Moreover, through the use of a managed service just like the Amazon Rekognition Moderation API, firms can cut back the time and sources wanted to develop and keep their very own fashions, which could be particularly helpful for companies with restricted technical sources. The API’s flexibility additionally permits on-line shops to customise their moderation guidelines and thresholds to suit their particular wants.

Study extra about content moderation on AWS and our content moderation ML use cases. Take step one in the direction of streamlining your content moderation operations with AWS.

Concerning the Authors

Lana ZhangShipra Kanoria is a Principal Product Supervisor at AWS. She is obsessed with serving to clients remedy their most complicated issues with the facility of machine studying and synthetic intelligence. Earlier than becoming a member of AWS, Shipra spent over 4 years at Amazon Alexa, the place she launched many productivity-related options on the Alexa voice assistant.

Lana ZhangLuca Agostino Rubino is a Principal Software program Engineer within the Amazon Buying workforce. He works on Neighborhood options like Buyer Critiques and Q&As, focusing by means of the years on Content material Moderation and on scaling and automation of Machine Studying options.

Lana ZhangLana Zhang is a Senior Options Architect at AWS WWSO AI Providers workforce, specializing in AI and ML for Content material Moderation, Laptop Imaginative and prescient, Pure Language Processing and Generative AI. Along with her experience, she is devoted to selling AWS AI/ML options and helping clients in reworking their enterprise options throughout numerous industries, together with social media, gaming, e-commerce, media, promoting & advertising and marketing.

Why Extra Is Extra (in Synthetic Intelligence) | by Manuel Brenner | Aug, 2023

Why Is Characteristic Scaling Vital in Machine Studying? Discussing 6 Characteristic Scaling Strategies | by Rukshan Pramoditha | Aug, 2023