The type of network that appears in the literature to be most commonly employed in natural language recognition (not necessarily processing) are called Recurrent Neural networks (RNNs). The varieties of that approach that show the greatest success in laboratory and commercial use are the Long Short-Term Memory (LSTM) and the Gated Recurrent Unit (GRU) designs.
The below are excellent articles. The first builds to the introduction of RNNs in Section 11. The second is more technical and deals with language as acoustic phenomena, which spoken words are.
Long Short-Term Memory Recurrent Neural Network Architectures for Large Scale Acoustic Modeling, Hasim Sak, et al, Interspeech 2014, Google, USA
A Primer on Neural Network Models for Natural Language Processing, Journal of Artificial Intelligence Research 57 (2016) 345–420, Submitted 9/15, published 11/16, Yoav Goldberg
Although a newer type of network design, the Attention approach, is touted to converge during training much faster that LSTM and the reasons given why it should are mathematically convincing, we haven't found any downloadable code or executable to test this claim.
Processing the output of the network is more related to the Eclipse plugin framework and the formatting facilities it contains. You want to specify your Abstract Symbol Tree (AST) as the network output and then use the Eclipse protocols and components to generate code according to a formatted so that it is consistent with developer or development team coding standards.
I also suggest taking a look at the Idea IDE. Personally, I use vim, Plugged, and the AST capabilities of clang++, EMMAScript, Java, and Python because IDEs have become cumbersome, slow, and buggy. I'd build the code generator as an independent app with exposed RestFUL query and control capabilities and then add plugins for Eclipse, Idea, Emacs, and Vim later.