Self-Supervised Document Similarity Ranking (SDR) via Contextualized Language Models and Hierarchical Inference

This repo is the implementation for SDR.

 

Tested environment

• Python 3.7
• PyTorch 1.7
• CUDA 11.0

Lower CUDA and PyTorch versions should work as well.

 

Contents

• Installation
• Datasets
• Train with our datasets
• Hierarchical Inference
• Cite

License, Security, support and code of conduct specifications are under the Instructions directory.  

Installation

Run

bash instructions/installation.sh 

 

Datasets

The published datasets are:

• Video games
  • 21,935 articles
  • Expert annotated test set. 90 articles with 12 ground-truth recommendations.
  • Examples:
    • Grand Theft Auto - Mafia
    • Burnout Paradise - Forza Horizon 3
• Wines
  • 1635 articles
  • Crafted by a human sommelier, 92 articles with ~10 ground-truth recommendations.
  • Examples:
    • Pinot Meunier - Chardonnay
    • Dom Pérignon - Moët & Chandon

For more details and direct download see Wines and Video Games.

 

Training

The training process downloads the datasets automatically.

python train_doc_sim.py --dataset_name video_games

The code is based on PyTorch-Lightning, all PL hyperparameters are supported. (limit_train/val/test_batches, check_val_every_n_epoch etc.)

Tensorboard support

All metrics are being logged automatically and stored in

SDR/output/document_similarity/SDR/arch_SDR/dataset_name_<dataset>/<time_of_run>

Run tesnroboard --logdir=<path> to see the the logs.

 

Inference

The hierarchical inference described in the paper is implemented as a stand-alone service and can be used with any backbone algorithm (models/reco/hierarchical_reco.py).

 

python train_doc_sim.py --dataset_name <name> --resume_from_checkpoint <checkpoint> --test_only

Results

Citing & Authors

If you find this repository or the annotated datasets helpful, feel free to cite our publication -

SDR: Self-Supervised Document-to-Document Similarity Ranking viaContextualized Language Models and Hierarchical Inference

 Link will be uploaded soon!

Contact person: Dvir Ginzburg