This repository contains the accompanying code for the paper:

"CDLM: Cross-Document Language Modeling." Avi Caciularu, Arman Cohan, Iz Beltagy, Matthew E Peters, Arie Cattan and Ido Dagan. In EMNLP Findings, 2021. [PDF]

The repository contains:

• Implementation of the CDMLM pretraining, based on the Huggingface code (in pretraining dir).
• Code for finetuning over cross-document coreference resolution (in cross_encoder dir).
• Code for finetuning over multi-document classification tasks (in CDA dir).
• Code for the attention analysis over the sampled ECB+ dataset (in attention_analysis dir).
• Code for finetuning over the multi-hop question answering task, using the HotpotQA dataset, including instructions, appears here.

You can either pretrain by yourself or use the pretrained CDLM model weights and tokenizer files, which are available on HuggingFace.

Then, use

from transformers import AutoTokenizer, AutoModel
# load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained('biu-nlp/cdlm')
model = AutoModel.from_pretrained('biu-nlp/cdlm')

Please note that during our pretraining we used the document and sentence separators, which you might want to add to your data. The document and sentence separators are <doc-s>, </doc-s> (the last two tokens in the vocabulary), and <s>, </s>, respectively.

If you find our work useful, please cite the paper as:

@article{caciularu2021CDLM,
  title={CDLM: Cross-Document Language Modeling},
  author={Caciularu, Avi and Cohan, Arman and Beltagy, Iz and Peters, Matthew E and Cattan, Arie and Dagan, Ido},
  journal={Findings of the Association for Computational Linguistics: EMNLP 2021},
  year={2021}
}