Deploying New Technology To Improve Access to Sexual Health Education in Kenya with Artificial Intelligence
In August 2016 we deployed the first version of SophieBot, an AI persona to answer your questions on sexual health. The more we told our story the more excitement we received, the more press we got, and the more questions we got asked. This exposed a glaring problem. The Rule-Based technology we had to answer questions couldn't keep up with all the new and different unique ways our users were asking the same question.
Here is an example :
“Hmm let me think about it ” was the default response to anything outside the rules and it quickly became our worst answer. We incurred a fair bit of wrath from users asking new and unexpected questions.
We were faced with two choices. One, constantly race against the users and update our rules to answer new questions. Two, develop new technology that learns from previous conversations to handle new and unexpected questions. We chose the latter. We talk about that journey in this interview.
As we worked on building that technology, our rule-based model got asked tens of thousands of questions by 5000 users in 125 countries (yes I counted). On the other hand, the problem of credible information on sexual health persisted, with dire consequences:
- Kenya has the third-highest teen pregnancy rates
- 98 Adolescents Infected With HIV Weekly
- There are 133,455 adolescents living with HIV in Kenya
We are proud to announce we have made strides to solve this problem using recent advances in AI and Natural Language Processing.
We painstakingly cleaned all the questions we have been asked, paired them with answers, and trained our own Language Model. The same technology that:
- Lets Github generate new code on command for software developers
- Generates hilarious posts on Facebook for Bots of New York
- Tricked a google engineer into thinking its sentient
For ML researchers reading this it looks like this:
Today when someone asks new and unexpected questions. SophieBot answers the questions.
The technology is deployed and live. We invite you to try her out for yourself on :
We look forward to your feedback as we keep cleaning new questions, matching them with answers, and updating our model to improve its robustness.