Error Analysis of using BART for Multi-Document Summarization: A Study for English and German Language

Authors: Timo Johner, Abhik Jana, Chris Biemann
Language Technology Group, Dept. of Informatics, Universitat Hamburg, Germany

Paper: Link

Accepted at the 23rd Nordic Conference on Computational Linguistics (NoDaLiDa 2021), held from May 31st to June 2nd, 2021. Published in the NEALT Proceedings Series by Linköping University Electronic Press and in the ACL Anthology.

This repository describes the implementation of our approach proposed in the paper.

Datasets:

name	language	topics	type	paper	source
CNN/DailyMail	en	311,971	single-document	Link	retrieved from here
Multi-News	en	56,216	multi-document	Link	adaption for BART by Hokamp et al (2020)
auto-hMDS	de	2,100	multi-document	Link	not publicly available, can be reproduced here

Setup:

The checkpoint for the fine-tuned BART model on the German auto-hMDS dataset can be downloaded here. The checkpoint file can be used to reproduce our results with the following setup.

Fine-Tuning:

We used the BART model based on the fairseq library. More information can be found here.

For fine-tuning on the three datasets (see above) we used the following parameters:

  CUDA_VISIBLE_DEVICES=1 fairseq-train hMDS_2-bin \
    --restore-file bart.large/model.pt \
    --max-tokens 1024 \
    --task translation \
    --source-lang source --target-lang target \
    --truncate-source \
    --layernorm-embedding \
    --share-all-embeddings \
    --share-decoder-input-output-embed \
    --reset-optimizer --reset-dataloader --reset-meters \
    --required-batch-size-multiple 1 \
    --arch bart_large \
    --criterion label_smoothed_cross_entropy \
    --label-smoothing 0.1 \
    --dropout 0.1 --attention-dropout 0.1 \
    --weight-decay 0.01 --optimizer adam --adam-betas "(0.9, 0.999)" --adam-eps 1e-08 \
    --clip-norm 0.1 \
    --lr-scheduler polynomial_decay --lr 3e-05 \
    --update-freq 1  \
    --skip-invalid-size-inputs-valid-test \
    --find-unused-parameters;

Citation:

If you find this paper interesting, please cite:

@inproceedings{#add,
    title = "Error Analysis of using BART for Multi-Document Summarization: A Study for English and German Language",
    author = "Johner, Timo and
    Abhik, Jana and
    Chris Biemann
    booktitle = "#add",
    month = #add,
    year = "2021",
    address = "Online",
    publisher = "#add",
    url = "#add",
    pages = "#add",
}