Unlocking multilingual accessibility: Worldreader’s generative AI journey to expand reach using Amazon Bedrock

In today’s digital age, access to engaging educational content remains a critical challenge globally. Worldreader, a pioneering nonprofit organization, is tackling this challenge head-on with a clear mission: inspiring families to read together with young children, nurturing a lifelong love of reading, and supporting school readiness. Through a strategic collaboration with SoftwareOne—an Amazon Web Services (AWS) Partner—Worldreader has revolutionized their digital reading platform BookSmart on AWS. BookSmart offers families a digital library of thousands of captivating books and learning activities. The platform harnesses AWS generative AI technology to create book-aligned activities (BAAs), making reading not only accessible, but also interactive and engaging.

Since 2010, Worldreader has supported over 22 million readers across more than 100 countries through their mobile reading technology, earning them the US Library of Congress Literacy Award’s International Prize. Their innovative approach combines locally sourced children’s books in multiple languages with motivational challenges and rewards, helping children reach their goal of reading 25 or more books annually with comprehension.

The technical foundation of this transformation rests on  Amazon Bedrock and the Anthropic Claude 3 Opus, which SoftwareOne leveraged to design and implement an automated system for generating BAAs across multiple languages. This innovative solution addresses three critical needs: increasing reader engagement through interactive content, eliminating resource-intensive manual content creation, and fostering meaningful connections between young readers and their families through guided reading discussions.

As we delve deeper into this technical journey, we’ll explore how combining cloud computing capabilities with generative AI on AWS has enabled Worldreader to scale their impact while maintaining the quality and educational value of their content.

The challenge: Making reading engaging and interactive

Worldreader’s approach to literacy centers on book-aligned activities (BAAs)—both pre- and post-reading—has transformed passive reading into interactive learning experiences. These BAAs are strategically categorized into three areas:

• Read: Enhances reading comprehension and vocabulary skills
• Play: Delivers engaging games and fun activities
• Grow: Develops emotional learning through self-awareness, relationship skills, and responsible decision-making

However, creating fresh and meaningful BAAs presents significant challenges for Worldreader’s lean team.

“Producing meaningful activities aligned with stories is a labor-intensive effort. It’s especially difficult to continuously come up with fresh content that hasn’t been used before,” explained Sonny Lacey, director of product at Worldreader.

The organization identified two critical needs:

1. Automating BAA creation to reduce development time from days to minutes
2. Developing automated assessments and reading comprehension evaluations

“We already offer BAAs, so we were interested in building book-aligned tests to further engage readers” noted Lacey.

To address these challenges, the Worldreader team decided to explore using generative AI technology as a solution to scale their impact while maintaining content quality.

The Solution: Architecting a generative AI-powered content generation system

The solution the Worldreader team developed leverages Amazon Bedrock and the Anthropic Claude 3 Opus large language model (LLM) within a serverless architecture that combines multiple AWS services to create an efficient, scalable content generation pipeline.

Key elements of technical architecture include, but are not limited to:

• Amazon Bedrock with the Anthropic Claude 3 Opus for LLM capabilities
• AWS Lambda for event-driven text extraction and prompt preparation
• Amazon Simple Storage Service (Amazon S3) for secure digital book storage
• AWS Step Functions for workflow orchestration and automation

The solution’s core strength lies in its intelligent prompt engineering. The team created models containing prompts around book texts, with examples and explanations focused on appropriate book-aligned activities (BAAs), building specific rules to consider socioeconomic status and maintain content relevance across diverse communities.

These rules helped fine-tune how activities consider variations in children’s socioeconomic status and which adult they are exploring reading activities with. This maintains appropriate activities for children from a wide range of backgrounds, whether they are reading with a parent or with another responsible adult. The LLM can also analyze books in other languages like Swahili, enabling Worldreader to generate activities for a near-native speaker experience.

The solution allows Worldreader to take a digital book and complete a language translation (which previously took days) in less than one minute. This enables the creation of appropriate book-aligned activities in its application to provide a personalized experience for a wide range of readers.

A key technical achievement was the solution’s multilingual capability. The Anthropic Claude 3 Opus model demonstrated remarkable proficiency in processing non-English content, with the ability to read Swahili books without translation and deliver BAAs in language close to a native speaker’s experience.

The serverless architecture of the solutions enables cost-effective scalability. AWS Lambda functions process book text extraction on-demand, while Step Functions orchestrate the entire workflow from text extraction to BAA generation. The system can efficiently scale to support BAAs for thousands of books as needed.

Figure 1: Serverless workflow architecture for automated book processing and content generation. Major components are an Amazon S3 bucket, AWS Lambda function, Step Functions, Amazon Translate, and Amazon Bedrock LLM.

The initial architecture design—shown in the preceding diagram—featured two distinct Amazon S3 buckets for storing books: one for English books and another for Swahili books. When a new book was uploaded, an Amazon S3 Event would trigger a notification via Amazon Simple Notification Service (SNS). This notification would then invoke an AWS Lambda function, initiating a processing workflow in AWS Step Functions.

A custom EPUB extraction Lambda function would extract all text from the books. The Step Functions workflow would then apply decision logic. If the book originated from the English bucket, it would be directed to the prompting Lambda function. If the book came from the Swahili bucket, it would first undergo translation with Amazon Translate.

The prompting Lambda function would generate a detailed prompt and send it to the large language model (LLM) in Amazon Bedrock. The output from the LLM would be post-processed and, if necessary, translated back to Swahili.

Figure 2: Native language processing streamlines multilingual content generation.

During implementation, the capabilities of Claude 3 Opus allowed for skipping the translation step entirely, reducing both cost and complexity of the solution. Additionally, working natively in Swahili eliminated compounding translation errors from back-and-forth translations.

Figure 3: Streamlined S3-Lambda event triggering.

Figure 4: AWS Step functions orchestration and workflow tasks.

Prompt Engineering: Designing AI prompts for child-centric learning activities

The system message provided to the model maintains that the generated activities are helpful, empathetic, and suitable for children. The activities are categorized into five skills: Vocabulary, Comprehension, Self Awareness, Social Awareness, and Physical. Each skill has two activities: one to be done before reading the book (Pre) and one to be done after reading the book (Post). Each activity requires assignment of skill category, placement (Pre/Post), activity text, and title. The title is given last to summarize the activity after generation. LLMs generate tokens in sequence based on the tokens already generated.

The prompt emphasizes creating activities accessible to children from all backgrounds, including those from low-income families. It avoids assuming the availability of specific materials or facilities and encourages alternatives when specific items are needed. The activities are designed to be safe for children to perform independently, without requiring adult supervision or potentially dangerous materials.

Moreover, the prompt highlights the need to avoid spoilers in the “pre” activities, as the readers will not be familiar with the plot or characters of the book at that stage. Instead, it suggests using generic terms like “main character” or “protagonist” to refer to the characters. The “post” activities, on the other hand, can refer to the characters by name and dive deeper into the book’s content. Specifically, few-shot prompting was used to provide existing book-aligned activities (BAAs) so the model reproduces their style and better understands what Worldreader needs. Chain-of-thought prompting was used to first detect the language of the book to maintain BAAs in the same language, as the model occasionally encountered challenges with this task. Specifically, the following output template was provided:

Detected language: [English/Swahili]

[ { “skill”: “vocabulary”, “placement”: “pre/post”, “activity”: “…”, “title”: “…” }, … ]

Figure 5: Prompt engineering framework for generating age-appropriate, inclusive book activities.

Example outputs

The examples below illustrate the results of the prompt. The “category” is added in post processing based on the “skill” field. Post processing validates that LLM output can be parsed and contains ten activities. This simple sanity check detects the majority of the errors in LLM generation.

Figure 6: Pre-Reading: Heroes to Zero #4: More Clean for More Green.

Figure 7: Post-Reading: Rini and Jojo Work It Out: The Super-Fast Apple Sorting Machine.

The outcome: Enhancing reader engagement and social awareness

Working with AWS and SoftwareOne, Worldreader created book-aligned activities (BAAs) that will increase reader engagement.

“With SoftwareOne and AWS, we created digital content that went beyond our expectations in terms of creativity,” said Sabrina Abreu González, senior manager of digital publishing at Worldreader. “This is important for readers because they’re not seeing the same book-aligned activities over and over.”

By automating content and BAA creation, Worldreader can focus on strategic goals and reduce reliance on external consultants.

 “With automation, we’ve removed the manual effort of relying on humans to study past BAAs to make sure we’re not reproducing something we’ve already done,” Lacey said. The team can prioritize improving literacy and reading comprehension, while helping children become more socially aware.

“Generative AI allows us to create content that helps young readers become more self-aware and learn to emphasize and interact with others. As a result, we can ultimately help readers build a better world,” stated González.

Conclusion

Worldreader’s successful implementation of AWS generative AI technology showcases how innovative cloud solutions can amplify social impact while solving complex operational challenges. By leveraging Amazon Bedrock and a serverless architecture, the organization has transformed its content creation process, enabling the rapid development of engaging, culturally relevant educational materials that reach millions of young readers globally.

The ability to automatically generate high-quality, multilingual book-aligned activities demonstrates the practical application of generative AI in solving real-world challenges. This transformation not only enhances Worldreader’s operational efficiency, but also enriches the learning experience for children across diverse communities.

As organizations worldwide struggle with similar challenges of scale and resource optimization, Worldreader’s journey provides a compelling blueprint for harnessing AWS generative AI to drive meaningful change. This success story illustrates how technology can be a powerful enabler in making education more accessible, engaging, and impactful for future generations.

If you’re ready to start building your own foundation model innovation with Amazon Bedrock, check out this webpage to learn more about setting up Amazon Bedrock. Dive deeper into the possibilities of Amazon Bedrock and elevate your generative AI journey today.