[TOC]

  1. Title: Grammar-Based Grounded Lexicon Learning
  2. Author: Jiayuan Mao
  3. Publish Year: 2021 NeurIPS
  4. Review Date: Dec 2021

Summary of paper

The paper extend the previous work “Neuro-Symbolic Concept Learner” by parsing the natural language questions using symbolic manner.

The core semantic parsing technique is Combinatory Categorical Grammar with CKY algorithm to prune unlikely expressions.

The full picture looks like this

image-20211224225350012

The detailed algorithm process looks like this

image-20211224225456819

How to derive concept embedding

e.g., “Shiny” concept is visually grounded

image-20211224230200070

image-20211224230930881

In the end, the concept embedding will freeze.

How to understand novel concept

actually, cannot….

but the author claimed that this model is very good at compositional generalisation task

essentially this is few-shot learning task, and the proposed neuro-symbolic model is good at few-shot learning task (of course)

image-20211224232553502


Some key terms

CKY-E algorithm

image-20211224225324045

Minor comments

Notes for Semantic Parsing with CCGs

CCG : Combinatory Categorial Grammar (CCG) is essential in this work.

We can learn it through Youtube videos https://www.youtube.com/watch?v=dOqe-ATkmeA&list=PLun-LUE1uLNvWi-qV-tRHohfHR90Y_cAk

Check the following notes

image-20211222173042879

image-20211222173105476

image-20211222173121482

Lambda Calculus

Base cases

image-20211222174146585

image-20211222174232631

example: check what logical form make sense

image-20211222175846916

Simply Typed Lambda Calculus

image-20211222175930455

image-20211222180013150

Constructing Lambda Calculus Expressions

image-20211224145641120

Combinatory Categorial Grammar

The grammar formalism that we will use is called combinatory categorical grammar CCG.

image-20211224145821724

CCG categories

image-20211224175624522

image-20211224175712783

image-20211224175802468

image-20211224175848037

Lexiconsimage-20211224180046672

image-20211224180121434

CCG Operations

image-20211224180302958

image-20211224180449204

image-20211224180644624

image-20211224180806512

image-20211224181613687

image-20211224181710351

image-20211224181803239

image-20211224181855551

How to parse with CCG, given an example

  1. step 1, use lexicon to match words and phrases with their categories
    • image-20211224182930407
  2. step 2, apply operation rules
    • image-20211224183119908
    • image-20211224183142345
  3. step 3, further operation rules, e.g., composition rule
    • image-20211224183323666
  4. step 4, coordinate composed adjectives
    • image-20211224183427652
  5. step 5, apply coordinated adjectives to noun
    • image-20211224183525807

Potential future work

a good baseline if we want to explore neuro-symbolic semantic parsing