Breaking language barriers with IDinsight’s generative AI tool – “Ask A Question”

At IDinsight, we consistently see how difficult it can be for people to access reliable information when they need it. For example, citizens want to know of social welfare benefits they are eligible for, new and expectant mothers need access to high-quality health advice, and farmers require guidance with best practices to increase their crop yield. 

To support communities, partner organizations we work with often spend a substantial amount of time curating a knowledge base – i.e. a database of information to be shared – often through frontline workers. However, it can be expensive to build a digital interface through which community members might access this information directly, and when these platforms do exist, it can be challenging to get people to use them, especially if there are language barriers. 

Recent advances in Large Language Models (LLMs) present an opportunity to overcome these barriers. Over the last two years, IDinsight has been refining a question-answering tool, ‘Ask A Question’ that can help nonprofits deliver vital information to their communities by answering questions in their home languages.

How it started

In 2021-2022, we worked with Reach Digital Health on the first version of Ask A Question for the MomConnect platform in South Africa. Before the automated service, a small helpdesk team would field thousands of questions per day. They were overworked and overwhelmed by the volume. To address this challenge, we built an automated question-answering solution that matched mother’s health queries to relevant content in the platform’s information database enabling instant resolution and a 60% reduction in the workload for the helpdesk staff.

With Ask A Question now embedded in MomConnect’s core WhatsApp chat flow, thousands of mothers interact with it to make the best health decisions for themselves and their babies.

A broader vision

Despite Ask A Question’s initial success, we knew there was room to improve. More users could get their questions answered instantly—perhaps 90% rather than 60% of people—if we could increase its question-answering performance. More users would find the platform accessible if we could support other widely spoken languages in South Africa, like Zulu and Xhosa. More organizations could use Ask A Question to reach the communities they serve if we could make it easier and cheaper to run. 

When OpenAI launched GPT-3.5, our vision for Ask A Question expanded even further. With the power of LLMs, we could enable a more natural interaction for users through conversational dialogues and allow them to clarify and refine their questions.

So, we went back to the drawing board and reevaluated our strategy. The result? A new roadmap – one that overhauled the original service design and prioritized inclusivity, safety, and LLM compatibility. This shift was a commitment to breaking down linguistic barriers, ensuring that information was accessible to all, regardless of the language spoken.

What we’re focused on

Recently, we were awarded additional support from Google.org’s AI for the Global Goals Impact Challenge, part of Google’s commitment to help accelerate progress towards the UN’s Sustainable Development Goals. Google.org provides us with technical and financial support, which will help us ensure our tool is best-in-class, available in multiple languages, and low-cost to operate. We hope our contributions can help further transform and improve how people, especially those in marginalized communities, access information from service providers. 

But what will make Ask A Question a game-changer in the landscape of AI tools designed for social impact? Let’s delve into the key goals:

1. Open-source

Many LLM-powered chat software have key features behind paywalls. Ask A Question began and will stay open source. The only cost will be the infrastructure. We hope that this will 1) make Ask A Question accessible to organizations with fewer resources; and 2) shape Ask A Question with contributions and feedback from the wider social sector and AI for Good community.

2. Designed to scale

Ask A Question is a set of containerized applications allowing for customizability, scalability, and deployment across diverse infrastructures. We will continue to expand it with cost-effectiveness in mind and aim for seamless integration with messaging platforms like WhatsApp and SMS. 

3. Accessibility

Ask A Question is being designed to handle messages in low-resource languages, ensuring that the tool is accessible to a wider range of users. We envision Ask A Question to provide a natural chat interface for users, with an option to send and receive voice notes for those who feel more comfortable that way.

4. Guardrails for AI safety

We take the safety of Ask A Question very seriously, especially considering its use of AI. We have rolled out a first version of guardrails to prevent misuse and LLM hallucinations. We will continue to improve upon them with support from Google.

We invite you to check out Ask A Question’s website to learn more about current and upcoming features, and how to deploy it yourself. Contributions to our Github repository are most welcome and please raise any questions about the project to tanmay.verma@idinsight.org.