Scientific papers typically organize contents in visual groups like text blocks or lines, and text within each group usually have the same semantics. We explore different approaches for injecting the group structure into the text classifiers, and build models that improves the accuracy or efficiency of the scientific text classification task.

After cloning the github repo, you can either install the vila library or just install the dependencies:

git clone git@github.com:allenai/VILA.git
cd VILA 
conda create -n vila python=3.6
pip install -e . # Install the `vila` library 
pip install -r requirements.txt # Only install the dependencies 

We tested the code and trained the models using Python≥3.6, PyTorch==1.7.1, and transformers==4.4.2.

Please check the detailed instructions in datasets/README.md.

We've uploaded a collection of pre-trained models to HuggingFace's model Hub:

MMDA is our newly designed toolkit that provides flexible supports for PDF document analysis. Please check the VILA predictor example here for more details.

In VILA repo, we also implemented a set of PDFPredictors. Please refer to the example code below:

import layoutparser as lp # For visualization 

from vila.pdftools.pdf_extractor import PDFExtractor
from vila.predictors import HierarchicalPDFPredictor
# Choose from SimplePDFPredictor,
# LayoutIndicatorPDFPredictor, 
# and HierarchicalPDFPredictor

pdf_extractor = PDFExtractor("pdfplumber")
page_tokens, page_images = pdf_extractor.load_tokens_and_image(f"path-to-your.pdf")

vision_model = lp.EfficientDetLayoutModel("lp://PubLayNet") 
pdf_predictor = HierarchicalPDFPredictor.from_pretrained("allenai/hvila-block-layoutlm-finetuned-docbank")

for idx, page_token in enumerate(page_tokens):
    blocks = vision_model.detect(page_images[idx])
    page_token.annotate(blocks=blocks)
    pdf_data = page_token.to_pagedata().to_dict()
    predicted_tokens = pdf_predictor.predict(pdf_data, page_token.page_size)
    lp.draw_box(page_images[idx], predicted_tokens, box_width=0, box_alpha=0.25)

VILA
├─ checkpoints  # For all trained weights 
│  └─ grotoap2  # For each dataset                                 
│     ├─ baseline  # For the experiment type, e.g., baseline, ivila, hvila, ...
│     │  └─ bert-base-uncased  # For the used base model, e.g., bert-base-uncased. 
│     │     ├─ checkpoint-199999                                
│     │     ├─ checkpoint-299999                                 
│     │     ├─ all_results.json                                       
│     │     └─ pytorch_model.bin                         
│     └─ ivila-BLK-row                           
│        └─ microsoft-layoutlm-base-uncased 
└─ data                                       
   ├─ docbank
   ├─ grotoap2
   └─ s2-vl

We provide the scripts to help you easily download and save the datasets into the aforementioned structure. Please follow the instructions in datasets/README.md.

All training scripts are in the ./scripts folder.

1. Train the baseline models

   cd scripts
   # bash train_baseline.sh [dataset-name] [base-model-name]
   bash train_baseline.sh grotoap2 bert-base-uncased
   bash train_baseline.sh docbank microsoft/layoutlm-base-uncased

2. Train the I-VILA models

   cd scripts
   # bash train_ivila.sh [dataset-name] [how-to-obtain-layout-indicators] [used-special-token] [base-model-name]
   bash train_ivila.sh grotoap2 row BLK microsoft/layoutlm-base-uncased 
     # Row is an alias for textline 
   bash train_ivila.sh docbank block SEP bert-base-uncased
     # We can also use the default special tokens like SEP 
   bash train_ivila.sh s2-vl sentence BLK roberta-base 
     # We can also extract the sentence breaks using spacy and use them as indicators.

3. Train the H-VILA models

   cd tools
   python create_hvila_model_base_weights.py 
   
   cd ../scripts
   # bash train_hvila.sh \
   #  [dataset-name] \
   #  [H-VILA-names] \
   #  [Group-Encoder-Output-Aggregation-Function] \
   #  [How-to-Obtain-Bounding-Box] \
   #  [Use-textline-or-block-as-the-group]
   
   bash train_hvila.sh \
     grotoap2 \
     weak-strong-layoutlm \
     average \
     first \
     row 

   In the above example, we use the:

   1. average of the group encoder outputs for all tokens as the group representation
   2. the bounding box of the first token as the group's bounding box
   3. textline (or row) to construct the groups

The evaluation toolkit can generate a detailed report for the prediction accuracy (marco F1 scores) and Visual Layout consistency (group entropy) for the prediction files test_predictions.csv produced by the training scripts.

1. Generate reports for a group of experiments for a specific dataset

  cd tools
  python generate-eval.py --dataset_name grotoap2 --experiment_name baseline
  # It will create a _reports folder in ../checkpoints/grotoap2/baseline and store the 
  # scores in the report.csv file. 

2. Generate reports for all experiments for a specific dataset

  cd tools
  python generate-eval.py --dataset_name grotoap2
  # It will create reports for all experiments in the ../checkpoints/grotoap2/ folder
  # Also it will aggregate all the results and save them in ../checkpoints/grotoap2/_reports 

3. Generate reports for per-class accuracy

  cd tools
  python generate-eval.py --dataset_name grotoap2 --experiment_name baseline --store_per_class
  # In additiona to the report.csv file, it will also generate a report_per_class.csv
  # table in the corresponding folder. 

Note: this evaluation toolkit might take a long time to run as calculating the group entropy may take long.

In order to support the AutoModel API, we changed the default transformers requirements to >=4.5 instead of 4.4.2. If you're working on reproducing the results, you might want to downgrade the transformers version to 4.4.2 to get the matching results.

@article{shen2022vila,
  title={VILA: Improving structured content extraction from scientific PDFs using visual layout groups},
  author={Shen, Zejiang and Lo, Kyle and Wang, Lucy Lu and Kuehl, Bailey and Weld, Daniel S and Downey, Doug},
  journal={Transactions of the Association for Computational Linguistics},
  volume={10},
  pages={376--392},
  year={2022},
  publisher={MIT Press}
}