Researchers From Bangladesh University And UCLA Use AI To Develop A Framework (Text2App) To Create Android Apps From Text Descriptions

It uses pretrained RoBERTa for Natural Language Understanding and MIT App Inventor as a backend for creating apps

A team of Researchers at BUET (Bangladesh University of Engineering and Technology) and UCLA (University of California- Los Angeles) has created a framework that can be used to develop Android applications from text descriptions.

According to Masum Hasan, a researcher who carried out the study, their team wondered whether a full-fledged software could be built from natural language specification. Almost all the existing models for creating software based on text descriptions are based on end-to-end neural machine translation (NMT) models, which are similar to the one behind Google Translate. Usually, these models use NMT frameworks to translate human language into a source code.

Though some NMT-based methods for text to app translation have achieved significant results, most of them are still unable to generate large programs with several lines of code. The researchers surpassed this limitation by creating a new formal language that captures the complexity of app source code, thus representing it in a highly compact form.

Next, they developed a compiler that can convert the compact representation into actual source code. They trained a neural machine translation model to translate the natural language to an intermediate format that can further be compiled into an app.

Rather than creating raw source code, the researchers trained a translation model to generate representations in the intermediate language they made. Their framework allowed them to create fully functional mobile applications by transforming text descriptions into this formal language.

The framework called Text2App can create Android applications using instructions in English. The users need to describe a mobile application with a specific scope, and then the framework automatically creates the application.

This novel development could allow researchers to re-evaluate program and app generation by introducing a new method through which programs can be summarized or represented and is easier for AI technologies to understand. Reportedly, the AI models trained by the researchers exhibited a far higher capacity than other techniques to convert natural language into apps. The most significant achievement of the study is the development of the intermediate formal language that allows the conversion of text into an app and a compiler that finally creates the app.