norikinishida / DiscourseConstituencyInduction-ViterbiEM

Unsupervised discourse constituency parsing using Viterbi EM (Nishida and Nakayama, 2020)

DiscourseConstituencyInduction-ViterbiEM

(c) 2020 Noriki Nishida

This is an implementation of an unsupervised discourse constituency parser described in the paper:

Noriki Nishida and Hideki Nakayama. 2020. Unsupervised Discourse Constituency Parsing Using Viterbi EM. Transactions of the Association for Computational Linguistics, vol.8, pp.215-230.

Task Definition

• Unsupervised discourse constituency parsing based on Rhetorical Structure Theory
• Input: EDUs, syntactic features, sentence/paragraph boundaries
• Output: Unlabeled RST-style constituent tree

Setup

Requirements

• numpy
• spacy >= 2.1.9
• chainer >= 6.1.0
• multiset
• jsonlines
• pyprind

Clone this repository and create directories to store preprocessed data and outputs

$ git clone https://github.com/norikinishida/DiscourseConstituencyInduction-ViterbiEM
$ cd ./DiscourseConstituencyInduction-ViterbiEM
$ mkdir ./data
$ mkdir ./results

Edit ./run_preprocessing.sh as follows:

STORAGE=./data

Edit ./config/path.ini as follows:

data = "./data"
results = "./results"
pretrained_word_embeddings = "/path/to/your/pretrained_word_embeddings"
rstdt = "/path/to/rst_discourse_treebank/data/RSTtrees-WSJ-main-1.0"
ptbwsj = "/path/to/LDC99T42/treebank_3/raw/wsj"

Clone other libraries

$ mkdir ./tmp
$ cd ./tmp
$ pip install pandas
$ pip install scikit-learn
$ pip install gensim
$ pip install nltk
$ git clone https://github.com/norikinishida/utils.git
$ git clone https://github.com/norikinishida/treetk.git
$ cp -r ./utils/utils ..
$ cp -r ./treetk/treetk ..

Preprocessing

./run_preprocessing.sh

• The following directories will be generated:

  • ./data/rstdt/wsj/{train,test} (preprocessed RST-DT)
  • ./data/ptbwsj_wo_rstdt (preprocessed PTB-WSJ)
  • ./data/rstdt-vocab (vocabularies)

• NOTE: We rewrote this part from scratch using spaCy to make the codes much simpler than the previous ones. (2020/05/11)

Training

• Training data: RST-DT training set

python main.py --gpu 0 --model spanbasedmodel2 --initial_tree_sampling RB2_RB_LB --config ./config/hyperparams_2.ini --name trial1 --actiontype train --max_epoch 15

• The following files will be generated:
  • ./results/spanbasedmodel2.RB2_RB_LB.hyperparams_2.aug_False.trial1.training.log
  • ./results/spanbasedmodel2.RB2_RB_LB.hyperparams_2.aug_False.trial1.training.jsonl
  • ./results/spanbasedmodel2.RB2_RB_LB.hyperparams_2.aug_False.trial1.model
  • ./results/spanbasedmodel2.RB2_RB_LB.hyperparams_2.aug_False.trial1.valid_pred.ctrees (optional)
  • ./results/spanbasedmodel2.RB2_RB_LB.hyperparams_2.aug_False.trial1.valid_gold.ctrees (optional)
  • ./results/spanbasedmodel2.RB2_RB_LB.hyperparams_2.aug_False.trial1.validation.jsonl (optional)

Evaluation

• Metrics: RST PARSEVAL by Morey et al. (2018)
• Test data: RST-DT test set

python main.py --gpu 0 --model spanbasedmodel2 --initial_tree_sampling RB2_RB_LB --config ./config/hyperparams_2.ini --name trial1 --actiontype evaluate

• The following files will be generated:
  • ./results/spanbasedmodel2.RB2_RB_LB.hyperparams_2.aug_False.trial1.evaluation.ctrees
  • ./results/spanbasedmodel2.RB2_RB_LB.hyperparams_2.aug_False.trial1.evaluation.json

Citation

If you use the code in research publications, please cite:

@article{nishida2020unsupervised,
    author={Nishida, Noriki and Nakayama, Hideki},
    title={Unsupervised Discourse Constituency Parsing Using Viterbi EM},
    journal={Transactions of the Association for Computational Linguistics},
    volume={8},
    number={},
    pages={215-230},
    year={2020},
    doi={10.1162/tacl\_a\_00312},
    URL={https://doi.org/10.1162/tacl_a_00312},
}