Compositional Phrase Alignment and Beyond

by Yuki Arase (Osaka University)

This repository provides an implementation of the phrase alignment method based on the constrained tree edit distance.

ESPADA is now available online!

Yuki Arase and Jun'ichi Tsujii. 2020. Compositional Phrase Alignment and Beyond. in Proc. of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1611-1623.

Prerequisites

Please make sure to install all the dependent libraries in requirements.txt

SPADE and ESPADA datasets are downloadable from LDC (Linguistic Data Consortium)

• SPADE: Test set
• ESPADA: Training and dev sets

Please place these corpora (xml files) in the data directory (the current repository provides just samples for debugging).

How to

Model names

The following labels correspond to models described in the paper.

• BERTTrainer: BERT model with simple fine-tuning
• BERT1F: The proposed model (BERT+SimMatrix)
• BERT1E: The proposed model using [CLS] instead of SimMatrix (BERT+[CLS])

Trained models

Trained models are distributed at Zenodo. Note that ALIR and ALIP values of these models are slightly different from what was reported in the paper because the paper reports the average performance of 10 models initialized with random seeds.

Fine-tune the BERT model

Pleaes set hyper-parameters as you want.

python ./fine-tune_bert.py --out_dir ../model/ --model_type BERT1F --pooling mean --train_epoch 100 --early_stop 5 --margin 1.0 --lr 3e-05 --ft_bert

Alignment with the constrained tree edit distance algorithm

python ./main.py --out_dir ../out/ --model_dir ../model/ --model_name BERT1F_TripletMarginLoss_margin-1.0_lr-3e-05_mean_100_ft-bert-base-uncased.pkl --pooling mean --null_thresh 0.8

If you want to output alignments, please flag --decode option.

Alignment with naive thresholding

python ./baseline_wo_ted.py --out_dir ../out/ --model_dir ../model/ --model_name BERT1F_TripletMarginLoss_margin-1.0_lr-3e-05_mean_100_ft-bert-base-uncased.pkl --pooling mean --null_thresh 0.6

Alignment with FastText

Download a FastText model you like, and specify the path to the model --model_dir

python ./main.py --out_dir ../out/ --model_name FastText --model_dir ../fasttext/crawl-300d-2M-subword.bin --pooling mean --null_thresh 0.8
python ./baseline_wo_ted.py --out_dir ../out/ --model_name FastText --model_dir ../fasttext/crawl-300d-2M-subword.bin --pooling mean --null_thresh 0.75

Alignment of your own dataset

1. Parse your dataset with Enju parser. Make sure to flag -xml to obtain outputs in an xml format
2. Rename your source and target xml files to follow the rule: s-(\d+).xml and t-(\d+).xml ((\d+) is the index of a pair). E.g., s-001.xml and t-001.xml
3. Place your xml files to ../data/name_of_your_xml_dir/ and change the path in decode() function in main.py
4. Run main.py flagging --decode
5. Alignment results will be saved in the output directory: index.txt are alignment results without postprocessing and PP_index.txt are alignments with postprocessing.

For your reference, data/Enju/ provides simple example inputs (Enju xml files) and out/alignment/ provides corresponding alignment outputs.

Citation

When you use our codes in your projects, please cite the following paper.

Yuki Arase and Jun'ichi Tsujii. 2020. Compositional Phrase Alignment and Beyond. in Proc. of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1611-1623.

@inproceedings{arase-tsujii-2020-compositional,
    title = "Compositional Phrase Alignment and Beyond",
    author = "Arase, Yuki  and
      Tsujii, Jun{'}ichi",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.125",
    doi = "10.18653/v1/2020.emnlp-main.125",
    pages = "1611--1623"
}