3D Scene Retrieval from Text with Semantic Parsing

We examine the problem of providing a natural-language interface for retrieving 3D scenes from descriptions. We frame the natural language understanding component of this task as a semantic parsing problem, wherein we first generate a structured meaning representation of a description and then use the meaning representation to specify the space of acceptable scenes. Our model outperforms a one-vs.-all logistic regression classifier on synthetic data, and remains competitive on a large, real-world dataset. Furthermore, our model learns to favor meaning representations that take into account more of the natural language meaning and structure over those that ignore words or relations between them.