Compositional Program Synthesis from Natural Language and Examples

Compositionality is a fundamental notion in computation whereby complex abstractions can be constructed from simpler ones, yet this property has so far escaped the paradigm of end-user programming from examples or natural language. Existing approaches restrict end users to only give holistic specifications of tasks, which limits the expressivity and scalability of these approaches to relatively simple programs in very restricted domains. In this paper we propose Compositional Program Synthesis (CPS): an approach in which tasks can be specified in a compositional manner through a combination of natural language and examples. We present a domain-agnostic program synthesis algorithm and demonstrate its application to an expressive string manipulation language. We evaluate our approach on complex tasks from online help forums that are beyond the scope of current state-of-the-art methods.