Aspect-based Document Similarity for Research Papers

Implementation, trained models and result data for the paper Aspect-based Document Similarity for Research Papers (PDF on Arxiv). The supplemental material is available for download under GitHub Releases or Zenodo.

• Datasets are compatible with 🤗 Huggingface NLP library (now known as datasets).
• Models are available on 🤗 Huggingface Transformers models.

Demo

You can try our trained models directly on Google Colab on all papers available on Semantic Scholar (via DOI, ArXiv ID, ACL ID, PubMed ID):

Requirements

• Python 3.7
• CUDA GPU (for Transformers)

Datasets

• ACL Anthology Reference Corpus (ACL ARC)
• COVID-19 Open Research Dataset (CORD 19)

Installation

Create a new virtual environment for Python 3.7 with Conda:

conda create -n paper python=3.7
conda activate paper

Clone repository and install dependencies:

git clone https://github.com/malteos/aspect-document-similarity.git repo
cd repo
pip install -r requirements.txt

Experiments

To reproduce our experiments, follow these steps (if you just want to train and test the models, skip the first two steps):

Prepare

export DIR=./output

# ACL Anthology 
# Get parscit files from: https://acl-arc.comp.nus.edu.sg/archives/acl-arc-160301-parscit/)
sh ./sbin/download_parsecit.sh

# CORD-19
wget https://ai2-semanticscholar-cord-19.s3-us-west-2.amazonaws.com/historical_releases/cord-19_2020-03-13.tar.gz

# Get additional data (collected from Semantic Scholar API)
wget https://github.com/malteos/aspect-document-similarity/releases/download/1.0/acl_s2.tar
wget https://github.com/malteos/aspect-document-similarity/releases/download/1.0/cord19_s2.tar

Build datasets

# ACL
python -m acl.dataset save_dataset <input_dir> <parscit_dir> <output_dir>

# CORD-19
python -m cord19.dataset save_dataset <input_dir> <output_dir>

Use dataset

The datasets are built on the Huggingface NLP library (soon available on the official repository):

from nlp import load_dataset

# Training data for first CV split
train_dataset = load_dataset(
    './datasets/cord19_docrel/cord19_docrel.py',
    name='relations',
    split='fold_1_train'
)                   

Use models

from models.auto_modelling import AutoModelForMultiLabelSequenceClassification

# Load models with pretrained weights from Huggingface model hub
acl_model = AutoModelForMultiLabelSequenceClassification('malteos/aspect-acl-scibert-scivocab-uncased')
cord19_model = AutoModelForMultiLabelSequenceClassification('malteos/aspect-cord19-scibert-scivocab-uncased')

# Use the models in standard Huggingface fashion ...
# acl_model(input_ids, token_type_ids, ...)
# cord19_model(input_ids, token_type_ids, ...)

Train models

All models are trained with the trainer_cli.py script:

python trainer_cli.py --cv_fold $CV_FOLD \
    --output_dir $OUTPUT_DIR \
    --model_name_or_path $MODEL_NAME \
    --doc_id_col $DOC_ID_COL \
    --doc_a_col $DOC_A_COL \
    --doc_b_col $DOC_B_COL \
    --nlp_dataset $NLP_DATASET \
    --nlp_cache_dir $NLP_CACHE_DIR \
    --cache_dir $CACHE_DIR \
    --num_train_epochs $EPOCHS \
    --seed $SEED \
    --per_gpu_eval_batch_size $EVAL_BATCH_SIZE \
    --per_gpu_train_batch_size $TRAIN_BATCH_SIZE \
    --learning_rate $LR \
    --do_train \
    --save_predictions

The exact parameters are available in sbin/acl and sbin/cord19.

Evaluation

The results can be computed and viewed with a Jupyter notebook. Figures and tables from the paper are part of the notebook.

jupyter notebook evaluation.ipynb

Due to the space constraints some results could not be included in the paper. The full results for all methods and all test samples are available as CSV files under Releases (or via the Jupyter notebook).

How to cite

If you are using our code, please cite our paper:

@InProceedings{Ostendorff2020c,
  title = {Aspect-based Document Similarity for Research Papers},
  booktitle = {Proceedings of the 28th International Conference on Computational Linguistics (COLING 2020)},
  author = {Ostendorff, Malte and Ruas, Terry and Blume, Till and Gipp, Bela and Rehm, Georg},
  year = {2020},
  month = {Dec.},
}

License

MIT