This repository contains an implementation of the doc2vec algorithm.

Contact me if you have any questions or want to use the code.

• Python 3.6.3 or newer

This program requires specific folders and files to work:

├── documents
│   ├── doc_0.txt
│   ├── doc_1.txt
│   ├── ...
│   └── doc_n.txt
├── main.py
├── labels.json
└── .gitignore

Every document that should be taken into account has to be inside one directory

• default this directory is documents/ but can be set to any folder relative to main.py
• each file should simply contain the plain text of the document

All labels have to inside a json file of the following form:

{
  "doc_0.txt": "Amazon Invoice",
  "doc_1.txt": "News article",
  "...": "...",
  "doc_n.txt": "Amazon Invoice"
}

• default this file is labels.json but can be set to any file relative to main.py
• note that the file extension is also part of the key

• logs will be saved into doc2vec.log
• a 2d graph for visual feedback will be saved into graph.eps
• a JSON containing the 10 most similar documents for every document will be saved into most_similars.json
• note that in this json a document should be most similar to itself to see if the systems acted as expected

All packages can be installed using pip

• numpy
• scikit-learn
• gensim
• matplotlib
• smart-open

• Clone the repo
• Run python main.py --doc_dir=documents/ --label_file=labels.json

All hyperparameters can be set using parameters: python main.py --save_dir=stored_models/

A list of all hyperparameters and their use can be found using: python main.py --help