Generating Algorithmic Patterns from Semi-structured Input Using a Transition-Based Neural Network

Richárd Rikk, Teréz A. Várkonyi, Zsolt Richárd Borsi, Balázs Pintér, Tibor Gregorics

Abstract: Synthesizing program code from natural language is a challenging task as natural language utterances tend to be ambiguous and require substantial prior knowledge to interpret. Recent solutions approach these difficulties in different ways. Some models do not constrain their inputs and operate with a large variety of sentences, but tend to be less accurate. Others limit the space of possible inputs, requiring them to meet a fixed structure which makes them more similar to code than language. This paper offers a middle ground between these approaches. We train a transition-based neural network on descriptions of programming tasks generated by using context-free grammar and templates. We show that the model is able to generalize and can solve synthesis problems described in natural language.