SAPPHIRE: Simple Aligner for Phrasal Paraphrase with Hierarchical Representation (Beta)

SAPPHIRE is a simple monolingual phrase aligner based on word embeddings.

We explain the details of SAPPHIRE in the following paper. [PDF]

@inproceedings{yoshinaka-etal-2020,
    author      = {Yoshinaka, Masato and Kajiwara, Tomoyuki and Arase, Yuki},
    title       = {SAPPHIRE: Simple Aligner for Phrasal Paraphrase with Hierarchical Representaion},
    booktitle   = {Proceedings of the 12th International Conference on Language Resources and Evaluation (LREC 2020)},
    year        = {2020},
    pages       = {6861--6867},
    url         = {https://www.aclweb.org/anthology/2020.lrec-1.847/}
}

Description

SAPPHIRE depends only on a pre-trained word embedding. Therefore, it is easily transferable to specific domains and different languages. This tool is designed for a pre-trained model of fastText. (Of course, it is easy to replace the word embedding.)

Requirements

• Python 3.6 or newer
• NumPy & SciPy
• fasttext

Installation

1. Download the pre-trained model of fastText (or prepare your model of fastText) and move it to model directory.

$ curl -O https://dl.fbaipublicfiles.com/fasttext/vectors-english/wiki-news-300d-1M-subword.bin.zip
$ unzip wiki-news-300d-1M-subword.bin.zip
$ mv wiki-news-300d-1M-subword.bin model/

Docker

1. Build the Docker image:

$ docker build -t sapphire .

2. Run a container:

$ docker run -it --rm -v ${PWD}/model:/work/model sapphire:latest /bin/bash
# python
>>> from sapphire import Sapphire

Local installation

1. Install requirements:

$ pip install -r requirements.txt

2. Install SAPPHIRE using develop option (that allows you to add scripts for other word representations):

$ python setup.py develop

Usage

Interactive mode

$ python run_sapphire.py model/wiki-news-300d-1M-subword.bin

To stop SAPPHIRE, enter Ctrl-C when inputting a sentence.

Usage of the SAPPHIRE module

>>> import fasttext
>>> from sapphire import Sapphire
>>> model = fasttext.FastText.load_model(path_to_your_model)
>>> aligner = Sapphire(model)

If you change the hyper-parameters,

>>> aligner.set_params(lambda_=0.6, delta=0.6, alpha=0.01, hungarian=False)

After preparing a tokenized sentence pair (tokenized_sentence_a: list and tokenized_sentence_b: list),

>>> _, alignment = aligner(tokenized_sentence_a, tokenized_sentence_b)
>>> alignment
[(1, 3, 2, 3), (8, 9, 5, 6), (13, 13, 8, 8), (27, 27, 9, 9)]

• Phrase pair is represented as .
• Outputs of SAPPHIRE are 1-indexed alignments.