kanishkamisra / world

world

Create mini-worlds consisting of concepts, their properties, their hierarchical organization, and ways to query for various related information.

This repo is in active development, so pls no judging code 😶‍🌫️

Requirements

All the following packages can be installed using pip

• semantic_memory*
• torch (> 1.8)
• nltk (3.7)
• numpy
• inflect
• pattern (optional, if you want to run src/negate.py)

*Note: run pip install semantic-memory -U just in case

Key datasets

Concept - Feature data

File: data/xcslb_compressed.csv

Fields:

• concept: noun concept
• feature: feature or property (verb-phrase)
• label: 1 if concept possesses property, 0 otherwise.
• category: category label as annotated by Devereaux et al., 2014

Note that all labels are set to 1 since the dataset in the repo is in compressed form (long-form), widening it would mean that every concept-feature pair that is absent in the compressed form would be associated with the label 0.

Concept - Feature Matrix

File: data/concept_matrix.txt

Rows = concepts, columns = individual features. Values = 1 if concept associated with feature, 0 otherwise.

Concept Lexicon

File: data/concept_senses.csv

Fields:

• category: category label as annotated by Devereaux et al., 2014
• concept: noun concept
• sensekey: Sense key in WordNET (to connect to a taxonomy)
• article: Article assignment of the noun (a/an/nothing).

Feature Lexicon

File: data/feature_lexicon.csv

Fields:

• feature: feature or property (verb-phrase)
• feature_type: Type of feature, as annotated by Devereaux et al., 2014. One of {taxonomic, encyclopedic, functional, visual perceptual, other perceptual}.
• negation: Negated verb-phrase of the property. Generated by running python negate.py in src/.

Using semantic-memory to generate worlds

Start off by running pip install semantic-memory -U, then follow along:

Creating a mini-world

from semantic_memory import memory

world = memory.Memory(
    concept_path="../data/concept_senses.csv",
    feature_path="../data/xcslb_compressed.csv",
    matrix_path="../data/concept_matrix.txt",
    feature_metadata="../data/feature_lexicon.csv",
)
world.create()

Get category members

Each world consists of a taxonomic organization of knowledge, which can be queried by using wordnet lemma names (e.g., animal = animal.n.01):

print(world.taxonomy['animal.n.01'])

# OUTPUT:
# Node animal.n.01
# Parent:organism.n.01
# Children: ['chordate.n.01', 'young.n.01', 'invertebrate.n.01', 'larva.n.01']

In most cases, the leaf nodes for every higher-order category are the main concepts in the dataset.

To get all birds (i.e., leaf nodes of the bird.n.01 category):

birds = world.taxonomy['bird.n.01'].descendants()
print(len(birds))

# OUTPUT:
# 36

print(birds[:5])

# OUTPUT:
# ['budgie', 'parakeet', 'buzzard', 'falcon', 'hawk']

Generating generics

Instructions for generating generics can be found in the demo notebook, src/Generics.ipynb.

TODO: exploring features, compute similarity, etc.

Citation

If you use the dataset and the tools used in this repository, please cite the following paper:

@inproceedings{misra2022property,
  title={A Property Induction Framework for Neural Language Models},
  author={Kanishka Misra and Julia Rayz and Allyson Ettinger},
  booktitle={Proceedings of the 44th Annual Conference of the Cognitive Science
Society},
  year={2022}
}