arglinking

Repository for ACL 2020 paper: Multi-Sentence Argument Linking [pdf].

Code

This repository uses a combination of files from the MISO repository used internally by Benjamin Van Durme's lab and components from allennlp. MISO is largely a fork of allennlp.

This code has been verified to run with python 3.6.9 and the packages listed in requirements.txt.

How to train models

For RAMS, GVDB, and BNB:

CUDA_VISIBLE_DEVICES=-1
python -u -m miso.commands.train <path/to/params.yaml>

The params file for the RAMS model is located at params/rams/gold_triggers+gold_arguments/rams.yaml. The params file for the BNB model is located at params/gc2012/gold_triggers+gold_arguments/gc2012.yaml. They contain gpu and cuda_device fields that you can modify depending on whether you are using CPU or GPU.

Various experimental settings from Table 2 may be set in the data reader (miso/data/dataset_readers/rams.py) and model (miso/models/arglinking.py) as designated with comments containing !!!!! (caveats/important items to be aware of), baseline (trigger representation baselines), and EXPERIMENT (use of distractor spans vs. full enumeration of spans).

How to evaluate models

For RAMS, GVDB, and BNB:

MODEL_DIR=<path/to/model>
OUTPUT_DIR=${MODEL_DIR}/results/test

CUDA_VISIBLE_DEVICES=-1
python -u -m miso.commands.evaluate ${MODEL_DIR}/model.tar.gz
       --weights-file ${MODEL_DIR}/best.th
       --input-file <path to test/*.json>
       --embeddingfile <path to/cached_bert_9_12.hdf5>
       --outputdir ${OUTPUT_DIR}
       --output-file ${OUTPUT_DIR}/test_metrics.json
       --batch-size 1

The suffix 9_12 in cached_bert_9_12.hdf5 refers to using BERT layers 9--12 (the top four layers). These can be computed using scripts/cache_bert_hf.py, which uses the Transformers library. See the header comment in that file for more details.

You may also need to set set -f to disable wildcard/glob expansion.

Data and Models

The RAMS dataset is available at https://nlp.jhu.edu/rams/. More information about the data is available on that webpage.

We plan to release the other datasets or splits used in the paper at a later date.

We also make available our RAMS and GVDB models.

Citation

If you wish to cite our work:

@inproceedings{ebner-etal-2020-multi,
  title={Multi-Sentence Argument Linking},
  author={Seth Ebner and Patrick Xia and Ryan Culkin and Kyle Rawlins and Benjamin {Van Durme}},
  year={2020},
  booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
}