This is the Repo for the paper: BARTScore: Evaluating Generated Text as Text Generation
- 2021.09.29 Paper gets accepted to NeurIPS 2021 🎉
- 2021.08.18 Release code
- 2021.06.28 Release online evaluation Demo
- 2021.06.25 Release online Explainable Leaderboard for Meta-evaluation
- 2021.06.22 Code will be released soon
There is a recent trend that leverages neural models for automated evaluation in different ways, as shown in Fig.1.
(a) Evaluation as matching task. Unsupervised matching metrics aim to measure the semantic equivalence between the reference and hypothesis by using a token-level matching functions in distributed representation space (e.g. BERT) or discrete string space (e.g. ROUGE).
(b) Evaluation as regression task. Regression-based metrics (e.g. BLEURT) introduce a parameterized regression layer, which would be learned in a supervised fashion to accurately predict human judgments.
(c) Evaluation as ranking task. Ranking-based metrics (e.g. COMET) aim to learn a scoring function that assigns a higher score to better hypotheses than to worse ones.
(d) Evaluation as generation task. In this work, we formulate evaluating generated text as a text generation task from pre-trained language models.
Basic requirements for all the libraries are in the requirements.txt.
Our trained BARTScore (on ParaBank2) can be downloaded here. Example usage is shown below.
# To use the CNNDM version BARTScore
>>> from bart_score import BARTScorer
>>> bart_scorer = BARTScorer(device='cuda:0', checkpoint='facebook/bart-large-cnn')
>>> bart_scorer.score(['This is interesting.'], ['This is fun.'], batch_size=4) # generation scores from the first list of texts to the second list of texts.
[out]
[-2.510652780532837]
# To use our trained ParaBank version BARTScore
>>> from bart_score import BARTScorer
>>> bart_scorer = BARTScorer(device='cuda:0', checkpoint='facebook/bart-large-cnn')
>>> bart_scorer.load(path='bart.pth')
>>> bart_scorer.score(['This is interesting.'], ['This is fun.'], batch_size=4)
[out]
[-2.336203098297119]We also provide multi-reference support. Please make sure you have the same number of references for each test sample. The usage is shown below.
>>> from bart_score import BARTScorer
>>> bart_scorer = BARTScorer(device='cuda:0', checkpoint='facebook/bart-large-cnn')
>>> srcs = ["I'm super happy today.", "This is a good idea."]
>>> tgts = [["I feel good today.", "I feel sad today."], ["Not bad.", "Sounds like a good idea."]] # List[List of references for each test sample]
>>> bart_scorer.multi_ref_score(srcs, tgts, agg="max", batch_size=4) # agg means aggregation, can be mean or max
[out]
[-2.5008113384246826, -1.626236081123352]To reproduce the results for each task, please see the README.md in each folder: D2T (data-to-text), SUM (summarization), WMT (machine translation). Once you get the scored pickle file in the right path (in each dataset folder), you can use them to conduct analysis.
For analysis, we provide SUMStat, D2TStat and WMTStat in analysis.py that can conveniently run analysis. An example of using SUMStat is shown below. Detailed usage can refer to analysis.ipynb.
>>> from analysis import SUMStat
>>> stat = SUMStat('SUM/REALSumm/final_p.pkl')
>>> stat.evaluate_summary('litepyramid_recall')
[out]
Human metric: litepyramid_recall
metric spearman kendalltau
------------------------------------------------- ---------- ------------
rouge1_r 0.497526 0.407974
bart_score_cnn_hypo_ref_de_id est 0.49539 0.392728
bart_score_cnn_hypo_ref_de_Videlicet 0.491011 0.388237
...If you want to train your custom BARTScore with paired data, we provide the scripts and detailed instructions in the train folder. Once you got your trained model (for example, my_bartscore folder). You can use your custom BARTScore as shown below.
>>> from bart_score import BARTScorer
>>> bart_scorer = BARTScorer(device='cuda:0', checkpoint='my_bartscore')
>>> bart_scorer.score(['This is interesting.'], ['This is fun.'])Since we are using the average log-likelihood for target tokens, the calculated scores will be smaller than 0 (the probability is between 0 and 1, so the log of it should be negative). The higher the log-likelihood, the higher the probability.
To give an example, if SummaryA gets a score of -1 while SummaryB gets a score of -100, this means that the model thinks SummaryA is better than summaryB.
Please cite our work if you find it useful.
@misc{yuan2021bartscore,
title={BARTScore: Evaluating Generated Text as Text Generation},
author={Weizhe Yuan and Graham Neubig and Pengfei Liu},
year={2021},
eprint={2106.11520},
archivePrefix={arXiv},
primaryClass={cs.CL}
}