This is the implementation of the paper On the Evaluation Metrics for Paraphrase Generation.

• Overview
• Requirements
• Data
• Run
• Toolkit
• Citation

In this work we present ParaScore, ParaScore is an evaluation metric for Paraphrase Evaluation. It possesses the merits of reference-based and reference-free metrics and explicitly models lexical divergence. There are two versions of ParaScore: reference-based and reference-free version.

(1) The reference-based version takes a reference, a source and a candidate as input, then it returns a score that indicates to what extent the candidate is good to be a paraphrase of source.

(2) The reference-free version is used for scenarios where references are unavailable. It takes a source and a candidate as input, then returns the score to as paraphrase quality indicator.

You can find more details of this work in our paper.

To run our code, please install all the dependency packages by using the following command:

pip install -r requirements.txt

NOTE: Different versions of packages (like pytorch, transformers, etc.) may lead to little different results from the paper. However, the trend should still hold no matter what versions of packages you use.

In a standard paraphrase evaluation paradigm, there're source sentences, references, and candidates.

In our paper, two benchmarks are selected: BQ-Para and Twitter-Para. Specifically, BQ-Para is the first Chinese paraphrase evaluation benchmark built by us, and Twitter-Para is adopted from Code or Paper.

There're two pickles in each dataset, one contains source-reference pairs and the other one contains source-candidate pairs. The IO of benchmarks can refer to data_utils.py.

Here's a simple script for running our codes.

python run.py \
    --metric parascore \
    --model_type bert-base-uncased \
    --batch_size 16 \
    --dataset_name twitter \
    --data_dir your data dir \
    --setting need \
    --extend True \

We further explain some of the script's arguments:

• metric: the metric name, including bleu | meteor | rouge | greedy | average | extrema | bert_score | parascore| ibleu
• model_type: the pre-trained model name when using PLM-based metrics like BERTScore.
• dataset_name: select dataset
  • bq: BQ-para dataset (Chinese)
  • twitter: Twitter-para dataset (English)
• setting: reference-based or reference-free version
  • need: reference-based
  • free: reference-free
• extend: whether the dataset is extended

this codebase supports pre-trained models in Huggingface's transformers. You can check Huggingface's website for available models and pass models with their names to --model_type. Some examples would be bert-large-uncased, roberta-base, roberta-large, etc.

We also prepare a convinient toolkit for fast usage for our ParaScore, which is called parascore. You can install it as follows:

pip install parascore==1.0.5

Some documents and tutorial can refer to the homepage for parascore_toolkit package. Here is a simple tutorial for parascore:

from parascore import ParaScorer
scorer = ParaScorer(lang="en", model_type = 'bert-base-uncased')
cands = ["A young person is skating."]
sources = ["There's a child on a skateboard."]
refs = ["A kid is skateboarding."]
score = scorer.base_score(cands, sources, refs, batch_size=16)
print(score)
[0.8152960109710693]

Please cite our paper if you use ParaScore in your work:

@inproceedings{shen2022evaluation,
  title={On the Evaluation Metrics for Paraphrase Generation},
  author={Shen, Lingfeng and Liu, Lemao and Jiang, Haiyun and Shi, Shuming},
  booktitle={Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing},
  pages={3178--3190},
  year={2022}
}