A Narrowing of AI research?

Code for the Narrowing of AI research paper (arXiv link) where we analyse the evolution of thematic diversity in the corpus of AI research in arXiv.

Setup

Environment

Run make create_environment to setup the correct conda environment, and then source activate narrowing_ai_research or conda_activate narrowing_ai_research to activate it.

Run conda install -c conda-forge graph-tool to install graph-tool.

Fetch data

Run make fetch.

This fetches data from the following sources:

• Figshare:
  • rXiv data (papers, institutions and categories)
  • topics
  • abstract vectors
• Scrapes arXiv ids from DeepMind and OpenAI websites

Data is stored in data/raw

It also fetches a topic model trained on arXiv abstracts identified as AI. The model is stored in models/topsbm

Processing data

Run make data

This:

• Cleans the fetched data
  • Create dates etc
  • Adds DeepMind and OpenAI labels to relevant papers
• Trains a word2vec model on the abstracts
• Finds the AI papers in the data and labels relevant datasets with an AI dummy
• Extracts topics form the topic model trained offline and removes generic and uninformative topics

All data outputs are stored in data/processed To save space, processed files in data/raw are removed.

Analysing the data

Run the scripts in narrowing_ai_research/paper to reproduce sections in the paper.

This:

• Reproduces all results in the paper including
  • Figures (in html and png) (in reports/figures)
  • Tables (in LaTeX) (in reports/tables)
  • Results (a text file in reports)
• You can also explore the results in the notebooks/paper folder.

The notebooks import functions from scripts in the narrowing_ai_research/paper folder.

The above uses the parametres in paper_config.yaml. Change for robustness tests etc.

Note that there are two steps in the pipeline that are not included in the repo. They are:

1. Fitting of hierarchical topic model (requires a machine w/ 64Gb RAM)
2. Creation of article embeddings

Step 1 is relatively easy to reproduce using the arxiv_tokenised.json tokenised abstracts generated while processing the data and the smbtm package that is already installed (for more infomation see here).

---

Project based on the Nesta cookiecutter data science project template.