GEM-metrics

Automatic metrics for GEM benchmark tasks. Can also be used standalone for evaluation of various natural language generation tasks.

Installation

GEM-metrics require recent Python 3, virtualenv or similar is recommended. To install, simply run:

git clone https://github.com/GEM-benchmark/GEM-metrics
cd GEM-metrics
pip install -r requirements.txt -r requirements-heavy.txt

If you want to just run the metrics from console (and don't need direct access to the source code), you can just run:

pip install 'gem-metrics[heavy] @ git+https://github.com/GEM-benchmark/GEM-metrics.git'

Note that some NLTK stuff may be downloaded upon first run into a subdirectory where the code is located, so make sure you have write access when you run this. Also note that all the required Python libraries are around 3 GB in size when installed.

If you don't need trained metrics (BLEURT, BERTScore, NUBIA, QuestEval), you can ignore the “heavy” part, i.e. only install dependencies from requirements.txt or only use gem-metrics instead of gem-metrics[heavy] if installing without checkout. That way, your installed libraries will be ~300 MB.

Script Usage

To compute all default metrics for a file, run:

<script> [-r references.json] outputs.json

Where <script> is either ./run_metrics.py (if you created a checkout) or gem_metrics if you installed directly via pip.

See test_data for example JSON file formats.

For calculating basic metrics with the unit test data, run:

./run_metrics.py -s test_data/unit_tests/sources.json  -r test_data/unit_tests/references.json test_data/unit_tests/predictions.json

Use ./run_metrics.py -h to see all available options.

By default, the “heavy” metrics (BERTScore, BLEURT, NUBIA and QuestEval) aren't computed. Use --heavy-metrics to compute them.

Library Usage

You can compute metrics for the same JSON format as shown in test_data, or you can work with plain lists of texts (or lists of lists of texts in the case of multi-reference data).

Import GEM-metrics as a library:

import gem_metrics

To load data from JSON files:

preds = gem_metrics.texts.Predictions('path/to/pred-file.json')
refs = gem_metrics.texts.References('path/to/ref-file.json')

To prepare plain lists (assuming the same order):

preds = gem_metrics.texts.Predictions(list_of_predictions)
refs = gem_metrics.texts.References(list_of_references)  # input may be list of lists for multi-ref

Then compute the desired metrics:

result = gem_metrics.compute(preds, refs, metrics_list=['bleu', 'rouge'])  # add list of desired metrics here

List of supported metrics

Referenceless:

• local_recall -- LocalRecall
• msttr -- MSTTR
• ngrams -- n-gram statistics
• ttr -- TTR

Reference-based:

• bertscore -- BERTScore (heavy)
• bleu -- BLEU
• bleurt -- BLEURT (heavy)
• chrf -- CHRF
• cider -- CIDER
• meteor -- Meteor (heavy)
• moverscore -- MoverScore (heavy)
• nist -- NIST
• nubia -- NUBIA (heavy)
• prism -- Prism
• questeval -- QuestEval (heavy)
• rouge -- ROUGE
• ter -- TER
• wer -- WER
• yules_i -- Yules_I

Source + reference based:

• sari -- SARI

License

Licensed under the MIT license.