Abstract:

Semantic parsing aims to transform a natural language utterance into a structured representation that can be easily manipulated by a software (for example to query a database). As such, it is a central task in human-computer interfaces. It has recently been observed that sequence-to-sequence models struggle on settings that require compositional generalization.

In this talk, I will present our work on graph-based semantic parsing. We proposed two novel graph-based formulations of this problem. Both methods deliver novel state-of-art results on standard benchmarks that test for compositional generalization. Moreover, I will explain how to build losses for learning when the alignment between concepts and words are not annotated in the data, and briefly discuss theoretical limitation of token-separable losses that are commonly used in the literature (including the presented work).