Linking Theories and Methods of Cognitive Control

This is the official repository for the paper Linking Theories and Methods in Cognitive Sciences via Joint Embedding of the Scientific Literature: The Example of Cognitive Control.

We performed automated text analyses on a large body of scientific texts (385705 scientific abstracts) and created a joint representation of cognitive control tasks and constructs.

Abstracts were first mapped into an embedding space using GPT-3 and Top2Vec models. Document embeddings were then used to identify a task-construct graph embedding that grounds constructs on tasks and supports nuanced meaning of the constructs by taking advantage of constrained random walks in the graph.

Setup

We recommend Conda/Mamba and DVC to set up a clean environment and download the data. You can create and activate the cogtext environment and automatically download the required data from CogText dataset on HuggingFace by running:

mamba env create --file environment.yml  # or use `conda`
mamba activate cogtext                   # activate the environment
dvc pull                                 # download the data

Notebooks

The main entry point of the project is the notebooks/ folder.

Note that Jupyter notebooks contain relative paths and are supposed to be run from the root of the project.

• 1 Data Collection (2023) uses the EFO ontology to search PubMed, aggregates abstracts as a single dataset, and stores the results in a compressed CSV file. If you already downloaded the CogText dataset, you can skip this step. Simply copy your downloaded file to data/pubmed/abstracts_2023.csv.gz.

• 2 Descriptive Statistics computes some basic statistics such as the number of tasks and constructs, co-occurrences, articles per each task or construct, etc. This notebook requires the data/pubmed/abstracts_2023.csv.gz file.

• 3 Document Embedding uses GPT-3 Embedding API (Ada) to transform the raw abstracts to vectorized embeddings.

• 4 Topic Embedding projects embeddings into a more interpretable topic space. The topic embedding uses UMAP and HDBSCAN to calculate the topic weights (as in Top2Vec).

• 5 Hypernomy visualizes construct hypernomy: inconsistent definitions of cognitive constructs across cognitive fields.

• 6 Hypergraph Visualization plots the task-construct hypergraph.

• 7 Link Prediction predicts the edges of the task-constructs hypergraph and learns Metapath2vec embedding of the graph nodes.

Acknowledgements

This research was supported by the Luxembourg National Research Fund (ATTRACT/2016/ID/11242114/DIGILEARN and INTER Mobility/2017-2/ID/11765868/ULALA).

Citation

To cite the paper use the following entry:

@misc{cogtext2022,
  author = {Morteza Ansarinia and
            Paul Schrater and
            Pedro Cardoso-Leite},
  title = {Linking Theories and Methods in Cognitive Sciences via Joint Embedding of the Scientific Literature: The Example of Cognitive Control},
  year = {2022},
  url = {https://arxiv.org/abs/2203.11016}
}