Skip to content

isekulic/longformer-marco

master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
src
 
 
 
 
 
 

Longformer for MS MARCO document ranking task

About

We employ Longformer, a BERT-like model for long documents, on the MS MARCO document re-ranking dataset. More details about our model and experimental setting can be found in our paper.

Learning setting

Due to the computing limitations, the hyperparameters were not optimised. We default to the following hyperparameters:

--lr=3e-05
--max_seq_len=4096
--num_warmup_steps=2500

For each query, we randomly sample 10 negative documents from the top 100 documents retrieved in the initial retrieval step.

Training the model

To train the model, first download all of the necessary data, as described in data/README.md. File names should match the filenames in MarcoDataset.py.

You can then train with:

python run_longformer_marco.py

You can check all available hyperparameters with:

python run_longformer_marco.py --help

Results

Dev Test
MRR@100 0.3366 0.307

The work is done by Ivan Sekulic (Università della Svizzera italiana), Amir Soleimani (University of Amsterdam), Mohammad Aliannejadi (University of Amsterdam), and Fabio Crestani (Università della Svizzera italiana).

Citing

Please consider citing our paper if you use our code or models:

@misc{sekuli2020longformer,
title={Longformer for MS MARCO Document Re-ranking Task},
author={Ivan Sekulić and Amir Soleimani and Mohammad Aliannejadi and Fabio Crestani},
year={2020},
eprint={2009.09392},
archivePrefix={arXiv},
primaryClass={cs.IR}
}

About

Longformer for MS MARCO document re-ranking task.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages