Synthesizing Object Models from Natural Language Specifications

Abstract

Program synthesis has traditionally excelled in tasks with precise specifications such as input-output examples and formal constraints by using structured and algorithmic approaches based on enumerative search and type inference. However, traditional synthesis techniques have no mechanism of incorporating real-world knowledge, which is commonplace in software engineering. Motivated by this, we introduce a new synthesis task known as specification reification: synthesizing concrete realizations of vague, high-level application specifications. We focus on a specific instance of this: generating object models from natural language application descriptions. Towards this goal, we present three approaches for object model synthesis that leverage domain knowledge from the GPT-3 language model. In addition, we design a scoring metric to evaluate the success of synthesized object models on seven sample tasks such as classroom management and pet store applications. We demonstrate that our language-model-based synthesizers generate object models that are comparable in quality to human-generated ones.