Girl Effect builds products for girls in developing markets. We work in Rwanda, India, Ethiopia, South Africa and Tanzania. Our product portfolio contains magazines, television shows, music videos, websites, chatbots and social media channels, all designed to drive uptake of knowledge and change in attitude around key subject areas defined in girl effect’s theory of change.
We try to make products which put our users at the front and centre of the experience. In our digital products, this means allowing girls to share their feedback and guide their own user journeys, rather than be passive recipients of content. As a result of this, we have collected a lot of unstructured qualitative data from our users. In Big Sis, our chatbot product in South Africa, we are now using this data to enable girls to write their sex and relationships questions which we run gainst a dictionary we have created based on existing training data and provide relevant content suggestions.
The challenge that we face is that the submissions we get are in multiple languages, use lots of slang, and often contain misspellings. Building up dictionaries of intent has to reflect the actual language that these girls are using to describe specific subject matter areas. It also means that with each new language we have to begin the same process again. There are very few existing models that are able to cater for this type of engagement.
There are many different use cases we would like to explore to build on the conversational experiences which we are creating: continuing with our subject specific classification to a greater level of nuance and accuracy, but also looking at signs of change expressed by users as a result of their engagement, and also to be able to provide for natural moments of 'smalltalk' conversations which help users feel confortable and guide them through their experience.
In order to do this in a more scalable way, we need assistance from data scientists with an expertise in Natural Language Processing to develop appraches to identifying common patterns in the user submissions we have across the different languages and use cases we work with, to establish the extent to which it would be possible to build responses based on a common underlying set of interactions.
View our work here.