Collective Intelligence for Smarter Neural Program Synthesis

We study the problem of automatically generating source code from different forms of user intents. Existing methods treating this problem as a language generating task of the neural network, known as Neural Program Synthesis (NPS). Most of these methods struggle with achieving high generating accuracy, one reason for that is the incompleteness and inaccuracy of user intents for a specific programming task. Inspired by the Swarm Intelligence (SI) and Collective Intelligence (CI) techniques, we proposed an automatic task-specific user intent merging framework combining both the bio-inspired algorithm in SI and CI merged from multiple developers. Empirically, we show that our approach is able to provide more accurate and adequate input for NPS, and our experiment on CI indicates that knowledge merging among isolated software developers in our approach has a significant influence on NPS.