Naacl 2021 Paper Review About Semantic Parsing

This page is a draft page, to take notes on recent text-to-SQL methods that is accepted in NAACL 2021.

The related papers are listed in random order.

[1] DuoRAT: Towards Simpler Text-to-SQL Models

[2] Structure-Grounded Pretraining for Text-to-SQL

[3] ShadowGNN: Graph Projection Neural Network for Text-to-SQL Parser

[4] Meta-Learning for Domain Generalization in Semantic Parsing

[5] Learning from Executions for Semantic Parsing

[6] Learning to Synthesize Data for Semantic Parsing

It seems like most of the SOTA semantic parsing papers are from one person: Bailin Wang. He is the author of RAT-SQL and [4], [5], and [6].

[4]

Intuition: Gradient Steps that improve source-domain performance should also improve target-domain performance

• DG-MAML (Domain Generalization with Model-Agnostic Meta-Learning)
  • a training algorithm that helps a parser acheive better domain generalization
  • training domain and test(eval) domain are different.
  • Meta-Train
    • SGD of loss from virtual source domain
  • Meta-Test
    • compute loss from virtual target domain
    • minimize the joint loss on both source and target domain -> require the gradient step beneficial to target domain
    • Can be viewed as regularization of gradient updates in additional to objective of conventional supervised learning
• Applying DG-MAML on RAT-SQL
  • evaluate on two zero-shot text-to-SQL - English/Chinese spider
  • achieves near SOTA on Spider, and SOTA on Chinese spider

Questions to ask TBU More notes TBU