This is the code for our EMNLP2020 paper: Multilevel Text Alignment with Cross-Document Attention

@inproceedings{Zhou2020Multilevel,
  author={Xuhui Zhou, Nikolaos Pappas, Noah A. Smith},
  title={Multilevel Text Alignment with Cross-Document Attention},
  booktitle={EMNLP},
  year={2020}
}

We also release a benchmark for Document Relation Prediction and Localization, where you can download the dataset used in our paper here!

conda env create -f environment.yml
conda activate dev

Our code has been run across many different python\ pytorch\ transformers versions (those are the essential packages for this projects). You can probably just use your own version of packages and run our code without further modification. Caution that we haven't tested out all models on different envs, so your results may vary from what we report in paper.

The following command will train a GRU-based hierachical attention network (HAN) classifier on the AAN corpus. Please modify the relative path accordingly to run the .sh script.

./run/acl_train.sh

• Please nvigate the files starting from .sh file which will point to relevant files in the /run folder.

• If a .sh file starts with acl, ai2_ab, ai2_g, and pla, it corresponds to AAN, OC, S2ORC, and PAN tasks described in our paper respectively.

The following command will train a BERT-based hierachical attention network (HAN) classifier on the AAN corpus. Please modify the relative path accordingly to run the .sh script.

./run/acl_train_bert.sh

First, download our dataset (Please always download the latest version).

You need to obtain the pre-trained contextualized embedding (.npy file as well as .index file) first to run the code. Though there are many ways to achieve that, we recommend using the following command (you can find get_rep.py in this repo):

export TRAIN_FILE='this does not matter'
export TEST_FILE='your data'

python get_rep.py \
    --output_dir=cite_models \
    --overwrite_output_dir \
    --model_type=bert \
    --per_gpu_eval_batch_size=10 \
    --model_name_or_path=bert-large-cased \
    --line_by_line \
    --train_data_file=$TRAIN_FILE \
    --special_eval \
    --eval_data_file=$TEST_FILE \
    --rep_name=./destination of your .npy \
    --mlm

The file takes in .txt, which is the sentences of documents as input and produce sentence embeddings .npy for each sentence. During the data reading phase, those sentence embeddings are re-organized into documents with sentence embeddings with the .index file. You can use the files in our datasets for this step.

If you want to run experiments on your own corpora, then you need to produce the aforementioned files. One example could be found at produce_txt.py, where the input is a .csv file and the output is a .txt and .index file.

Note that this step could be flexible. It should be fine as lone as you use a valid sentence tokenizer (I use nltk.tokenize.sent_tokenize) and record the position of the positive sentences (We use \001 to represent the positive sentence. For situations like pla with multiple positive sentences, we use \001 to circle the span of positive sentences. Our new version of data unifies how we represent the positive sentences, which is the version we use in our paper.) Instead of re-recognizing the positive sentences during the data reading phase, you could also genarate a .csv file to record the positions when you tokenize the sentences, which is the case in our plagirism detection experiments.

Warning: Please ensure you have GPU space exceeds 10 GB to run the fine-tuning version:

The following command will train a BERT-based hierachical attention network (HAN) classifier on the AAN corpus with BERT finetuning. Please modify the relative path accordingly to run the .sh script.

./BERT-HAN/run_ex_sent.sh

• For evaluating models, just find corresponding evaluation .sh scripts, they should be straightforward.
• For other scripts, they are used for variations of our experiments, some of them are reported in the paper while some are not. You can just ignore them or you can also read on.
• You can directly take a look at d_graph_hier_mat_model_g.py and graph_hier_mat_model_g.py for our implementation for Deep and Shallow CDA. This should be the place where magic happens.
• For evaluation scripts, we recommend taking a look at test_pla_mask_onestep.py and test_state_dict_mask_onestep.py for plagiarism task and citation recommendation task. They have some subtle difference for S2D task.