Semantic Parsing, basically consists of two target

  • QA (for the sake of returned value)
  • Instruction (for the sake of side-effects)

1. Datasets

There’re 4 aspects of research

  1. for executable representations
  2. understanding in a situated environment
  3. generalizing to broader domains
  4. discourse or long-term language understanding

Executable Repr.

GeoQuery, Jobs (< 1K examples)

ATIS (27 tables and 160K entries, 5K examples)

CoNaLa: NL pseudocode to python (2500 curated examples, 600K noisy examples)

Django, HearthStone and Magic the Gathering, NL2Bash, CONCODE

WikiSQL (Zhong, Xiong, and Socher, 2017)

In A Situated Env.

SAIL: VR 3D NLVR: images

Referring expressing: Blocks world: move blocks around Room2Room: navigation in a realistic home

Broader domains

AMRBank WikiTableQuestions Free917, WebQuestions

Sequential Language Understanding



SQA (Iyyer et al. 2016)

Task-oriented dialogues over knowledge bases (Eric et al. 2017)

2. Constrained Decoding

neural models usually output syntactically or semantically errors

Token-based decoding:

Seq2Tree (Dong and Lapata, 2016): syntactically correct, but semantically error Sketch-Constrained Seq2Tree (Dong and Lapata, 2018)

Grammar-based decoding: Krishnamurthy et al. 2017 Neural Semantic Parsing with Type Constraints for Semi-Structured Tables

3. Training

Fully supervised (with LFs) v.s. weakly supervised

  • Maximum Marginal Likelihood method
  • Structured Learning methods
  • Reinforcement Learning methods
  • Hybrid ones…


Given $D = {x_i, w_i, y_i}_{i=1}^N$, where w is the world data, y is the correct answer.


for approximating Y:

heuristic search (usually bounded, by length or something else)

  • online search: during training, candidates varying but less efficient
  • offline search: consistent LFs are searched before training

Structured Learning

common in traditional semantic parsers (margin based or latent variable perceptron)

typically involve heuristic search over the state space

unlike MML, arbitrary cost function is available

training typically maximized margins or minimizes expected risks

RL method

with MML:

  • logic forms is also approximated, same with MML
  • but approximation is done using sampling

comparison with SL:

  • like SL, reward function is arbitrary
  • unlike SL, reward is directly maximized

using REINFORCE (policy gradient)

Bridging together

MML + RL: Guu et al. 2017

RL + SL: Iyyer et al. 2017

4. SP as Code Generation

  • single line
  • method level
  • class level

5. context-dependent

context consists of

  • previous instructions
  • previous interpretations
  • current world state
  • previous world state

Suhr et al. 2018