The natural language programming (NLPRO) project: Turning text into executable code

In this paper we present the natural language programming (NLPRO) project (via ERC-StG-2015 grant 677352), where we strive to automatically translate requirements documents directly into the executable code of the systems they describe. To achieve this, we embrace the ambiguity of NL requirements and define a three-fold research agenda wherein we (i) formalize text-to-code translation as a structure prediction task, (ii) propose a formal semantic representation in terms of Live Sequence Charts (LSCs), and (iii) develop and comparatively evaluate novel sentence-based vs. discourse-based models for semantic parsing of requirements documents, and test their accuracy on various case studies. The empirical results of our first research cycle show that the discourse-based models consistently outperform the sentence-based models in constructing a system that reflects the requirements in the document. We conjecture that the formal representation of LSCs, the joint sentence-discourse modeling strategy, and the statistical learning component, are key ingredients for effectively tackling the NLPRO long-standing challenge.