Researchers at Bangladesh College of Engineering and Expertise (BUET) and College of California- Los Angeles (UCLA) just lately developed a framework that may create Android functions from textual content descriptions. This new app was introduced in a paper pre-published on arXiv.
“Our analysis group on the Division of CSE, BUET works on bridging the hole between human language and programming language to make pc programming extra accessible,” Masum Hasan, one of many researchers who carried out the examine, advised TechXplore. “We now have beforehand labored on pure language-based program restore, language-based code search, code doc technology, and so forth. In continuation, we questioned if absolutely functioning software program could possibly be constructed from pure language specification.”
Most current fashions for creating software program based mostly on textual content descriptions are based mostly on end-to-end neural machine translation (NMT) fashions, just like the one underpinning Google Translate. These fashions make the most of NMT frameworks to translate human language into supply code.
Whereas some NMT-based methods for textual content to app translation have achieved promising outcomes, most of them are unable to generate giant applications with a whole lot of strains of code. To beat this limitation, Hasan and his colleagues created a brand new formal language that captures the complexity of app supply code, representing it in a extremely compact type.
“We designed an intermediate formal language that may symbolize a consumer specified app in a fraction of the tokens—as a lot as 20 occasions much less,” Hasan stated. “We then developed a compiler that may convert the compact illustration into precise supply code and educated a neural machine translation mannequin to translate pure language to this intermediate format, which may then be compiled into an app.”
Basically, the researchers educated a translation mannequin to generate representations within the intermediate language they created, moderately than creating uncooked supply code, as most current textual content to app methods would. By remodeling textual content descriptions into this formal language, their framework allowed them to create absolutely fledged cellular functions, which proved unfeasible utilizing beforehand proposed strategies.
“Producing lengthy sequences is a identified limitation of present sequence technology fashions,” Hasan defined. “Supply code is often for much longer than pure language and the present finest fashions nonetheless wrestle to realize 20% accuracy in producing brief code snippets. This makes it impractical to create giant software program, comparable to cellular functions consisting of a whole lot of strains of codes, with present sequence technology networks.”
The framework created by Hasan and his colleagues, dubbed Text2App, can create Android functions utilizing directions in English. To make use of the system, customers merely want to explain a cellular utility with a particular scope and the framework mechanically creates it for them. A video demo of Text2App is now out there on GitHub, at: https://text2app.github.io/
Hasan stated. “Think about you need to memorize your grandmother’s particular recipe for pasta. As a substitute of remembering all the recipe step-by-step, you may solely study the steps which might be distinctive, and lookup the remainder of the steps from a cookbook. It will considerably scale back the quantity of effort you would need to spend. Equally, our intermediate app illustration solely captures the particular necessities supplied by the consumer via the pure language instruction and the compiler fills in the remaining.”
The current work by Hasan and his colleagues may enable researchers to re-think program and app technology, by introducing a brand new manner via which applications could be summarized or represented that’s simpler for AI know-how to grasp. Notably, the AI fashions educated by the researchers exhibited a far larger capability than different methods to transform pure language into apps.
“If we design programming languages or frameworks with AI in thoughts, our present AI applied sciences can carry out remarkably higher in code technology from human language,” Hasan stated. “Moreover, programming languages are made to be understood by compilers, and never totally by AIs. We display that counting on AI fashions alone for advanced duties like code technology is just not enough, and extra assist from parts like compilers will help an ideal deal.”
Essentially the most notable achievement of the examine carried out by this crew of researchers is the event of the intermediate formal language that permits the conversion of textual content into an app, in addition to a compiler that in the end creates the app. To this point, Hasan and his colleagues explored a mere fraction of its capabilities, however they’re now additionally asking the open-source neighborhood to contribute to its additional analysis and growth.
“With extra growth efforts, Text2App can grow to be a full-fledged pure language- based mostly app growth platform,” Hasan stated. “In our work, we additionally present that it’s attainable to create functions from surprisingly advanced directions with our system with the assistance of a big language mannequin, GPT-3. Nevertheless, extra ideas on this route are essential to make it absolutely dependable.”
A man-made neural community to accumulate grounded representations of robotic actions and language
Text2App: a framework for creating android apps from textual content descriptions. arXiv:2104.08301 [cs.CL]. arxiv.org/abs/2104.08301
Review4Repair: Code evaluate aided automated program repairing. arXiv:2010.01544 [cs.SE]. arxiv.org/abs/2010.01544
Language fashions are few-shot learners. arXiv:2005.14165 [cs.CL]. arxiv.org/abs/2005.14165
CoDesc: A big code-description parallel dataset. arXiv:2105.14220 [cs.CL]. arxiv.org/abs/2105.14220
© 2021 Science X Community
Text2App: A framework that creates Android apps from textual content descriptions (2021, June 2)
retrieved 5 June 2021
This doc is topic to copyright. Other than any truthful dealing for the aim of personal examine or analysis, no
half could also be reproduced with out the written permission. The content material is supplied for data functions solely.