An Empirical Study of Span Representation in Argumentation Structure Parsing

Citation

@InProceedings{P19-1464,
  author = "Kuribayashi, Tatsuki
	and Ouchi, Hiroki
	and Inoue, Naoya
	and Reisert, Paul
    	and Miyoshi, Toshinori
    	and Suzuki, Jun
    	and Inui, Kentaro"
  title = 	"An Empirical Study of Span Representation in Argumentation Structure Parsing",
  booktitle = 	"Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
  year = 	"2019",
  publisher = 	"Association for Computational Linguistics",
  pages = 	"4691-4698",
  location = 	"Florence, Italy",
  url = 	"https://www.aclweb.org/anthology/P19-1464"
}

• conference paper: https://www.aclweb.org/anthology/P19-1464

Prerequirement

• python=3.6.5
• cuda 9.0 (if the cuda version differs from your environment, change the cupy version in Pipfile)
• pipenv
• GloVe(6B, 300d)

Installation

pipenv install
pipenv shell

Preprocess

bash preprocess.sh [DEVICE_ID]

• Set DEVICE_ID if you use GPU for encoding the texts with ELMo. If not, set -1 to DEVICE_ID.

Usage

If you run the training on CPU, unset the --device and -g options.

Experiments on the persuasive essay corpus

• ELMo, LSTM+dist
  src/train.py --dataset PE --device -g [DEVICE_ID] --seed 39 --use-elmo 1 --elmo-path work/PE4ELMo.hdf5 --reps-type contextualized -ed 300 -hd 256 --optimizer Adam --lr 0.001 --ac-type-alpha 0.25 --link-type-alpha 0.25 --batchsize 16 --epoch 500 --dropout 0.5 --dropout-lstm 0.1 --lstm-ac --lstm-shell --lstm-ac-shell --lstm-type --elmo-layers avg -o [OUT_DIRECTORY]

• GloVe, LSTM+dist
  src/train.py --dataset PE --device -g [DEVICE_ID] --seed 39 --glove-dir [GLOVE_PATH] --use-elmo 0 --reps-type contextualized -ed 300 -hd 256 --optimizer Adam --lr 0.001 --ac-type-alpha 0.25 --link-type-alpha 0.25 --batchsize 16 --epoch 500 --dropout 0.5 --dropout-lstm 0.1 --lstm-ac --lstm-shell --lstm-ac-shell --lstm-type -o [OUT_DIRECTORY]

Training on the arg-microtext corpus (10 iterations of 5 fold-CV)

• ELMo, LSTM+dist
  scripts/exp_MT.sh [DEVICE_ID] [OUT_DIRECTORY]

Check the performance of each model

src/results.py -d [DIRECTORY_OF_OUTPUT_FILES]

Inference and analysis

python src/infer_analysis.py -d [DIRECTORY_OF_OUTPUT_FILES] -g [DEVICE_ID] --device --data test

LICENSE

MIT License