Evaluation code for various unsupervised automated metrics for NLG (Natural Language Generation). It takes as input a hypothesis file, and one or more references files and outputs values of metrics. Rows across these files should correspond to the same example.
- BLEU
- METEOR
- ROUGE
- CIDEr
- SkipThought cosine similarity
- Embedding Average cosine similarity
- Vector Extrema cosine similarity
- Greedy Matching score
Install Java 1.8.0 (or higher). Then run:
# Install the Python dependencies.
pip install git+https://github.com/Maluuba/nlg-eval.git@master
# If using macOS High Sierra or higher, run this before run setup, to allow multithreading
# export OBJC_DISABLE_INITIALIZE_FORK_SAFETY=YES
# Simple setup:
# Download required data (e.g. models, embeddings) and external code files.
nlg-eval --setup
# If you don't like the default path (~/.cache/nlgeval) for the downloaded data,
# then specify a path where you want the files to be downloaded.
# The value for the data path is stored in ~/.config/nlgeval/rc.json and can be overwritten by
# setting the NLGEVAL_DATA environment variable.
nlg-eval --setup ${data_path}
Once setup has completed, the metrics can be evaluated with a Python API or in the command line.
Examples of the Python API can be found in test_nlgeval.py.
nlg-eval --hypothesis=examples/hyp.txt --references=examples/ref1.txt --references=examples/ref2.txt
where each line in the hypothesis file is a generated sentence and the corresponding lines across the reference files are ground truth reference sentences for the corresponding hypothesis.
from nlgeval import compute_metrics
metrics_dict = compute_metrics(hypothesis='examples/hyp.txt',
references=['examples/ref1.txt', 'examples/ref2.txt'])
from nlgeval import compute_individual_metrics
metrics_dict = compute_individual_metrics(references, hypothesis)
where references
is a list of ground truth reference text strings and
hypothesis
is the hypothesis text string.
from nlgeval import NLGEval
nlgeval = NLGEval() # loads the models
metrics_dict = nlgeval.compute_individual_metrics(references, hypothesis)
where references
is a list of ground truth reference text strings and
hypothesis
is the hypothesis text string.
from nlgeval import NLGEval
nlgeval = NLGEval() # loads the models
metrics_dict = nlgeval.compute_metrics(references, hypothesis)
where references
is a list of lists of ground truth reference text strings and
hypothesis
is a list of hypothesis text strings. Each inner list in references
is one set of references for the hypothesis (a list of single reference strings for
each sentence in hypothesis
in the same order).
If you use this code as part of any published research, please cite the following paper:
Shikhar Sharma, Layla El Asri, Hannes Schulz, and Jeremie Zumer. "Relevance of Unsupervised Metrics in Task-Oriented Dialogue for Evaluating Natural Language Generation" arXiv preprint arXiv:1706.09799 (2017)
@article{sharma2017nlgeval,
author = {Sharma, Shikhar and El Asri, Layla and Schulz, Hannes and Zumer, Jeremie},
title = {Relevance of Unsupervised Metrics in Task-Oriented Dialogue for Evaluating Natural Language Generation},
journal = {CoRR},
volume = {abs/1706.09799},
year = {2017},
url = {http://arxiv.org/abs/1706.09799}
}
Running
nlg-eval --hypothesis=examples/hyp.txt --references=examples/ref1.txt --references=examples/ref2.txt
gives
Bleu_1: 0.550000
Bleu_2: 0.428174
Bleu_3: 0.284043
Bleu_4: 0.201143
METEOR: 0.295797
ROUGE_L: 0.522104
CIDEr: 1.242192
SkipThoughtsCosineSimilarity: 0.626149
EmbeddingAverageCosineSimilarity: 0.884690
VectorExtremaCosineSimilarity: 0.568696
GreedyMatchingScore: 0.784205
If you have issues with Meteor then you can try lowering the mem
variable in meteor.py
CIDEr by default (with idf parameter set to "corpus" mode) computes IDF values using the reference sentences provided. Thus, CIDEr score for a reference dataset with only 1 image (or example for NLG) will be zero. When evaluating using one (or few) images, set idf to "coco-val-df" instead, which uses IDF from the MSCOCO Vaildation Dataset for reliable results. This has not been adapted in this code. For this use-case, apply patches from vrama91/coco-caption.
To mount an already prepared data directory to a Docker container or share it between
users, you can set the NLGEVAL_DATA
environment variable to let nlg-eval know
where to find its models and data. E.g.
NLGEVAL_DATA=~/workspace/nlg-eval/nlgeval/data
This variable overrides the value provided during setup (stored in ~/.config/nlgeval/rc.json
)
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.
See LICENSE.md.