def test_rouge(self):
# Test the first several instances in the TAC 2008 data to ensure that
# our computation of ROUGE matches the values released by NIST
instances = ReferenceBasedDatasetReader().read(_summaries_file_path)
metrics_list = JsonlReader(_metrics_file_path, Metrics).read()
metric_names = ['rouge-1', 'rouge-2', 'rouge-3', 'rouge-4', 'rouge-l', 'rouge-su4', 'rouge-w-1.2']
rouge = Rouge(max_ngram=4,
use_porter_stemmer=True,
remove_stopwords=False,
compute_rouge_l=True,
skip_bigram_gap_length=4,
wlcs_weight=1.2)
peer_instances, peer_metrics = self._filter_by_type(instances, metrics_list, 'peer')
reference_instances, reference_metrics = self._filter_by_type(instances, metrics_list, 'reference')
num_to_check = 25
actual_metrics_dicts = score_instances(peer_instances[:num_to_check], [rouge])
for expected_metrics in peer_metrics[:num_to_check]:
instance_id = expected_metrics.instance_id
summarizer_id = expected_metrics.summarizer_id
actual_metrics = actual_metrics_dicts[instance_id][summarizer_id]
for metric in metric_names:
assert actual_metrics.metrics[metric] == expected_metrics.metrics[metric]
assert actual_metrics.metrics[metric + '_jk'] == expected_metrics.metrics[metric + '_jk']
actual_metrics_dicts = score_instances(reference_instances[:num_to_check], [rouge])
for expected_metrics in reference_metrics[:num_to_check]:
instance_id = expected_metrics.instance_id
summarizer_id = expected_metrics.summarizer_id
actual_metrics = actual_metrics_dicts[instance_id][summarizer_id]
for metric in metric_names:
assert actual_metrics.metrics[metric + '_jk'] == expected_metrics.metrics[metric + '_jk']
def test_task2_rouge(self):
# Test the first several instances in the DUC 2007 data to ensure that
# our computation of ROUGE matches the values released by NIST. In this year,
# they did not release the raw output for the non-jackknifing runs, so
# we cannot compare those scores
instances = ReferenceBasedDatasetReader(
_task2_summaries_file_path).read()
metrics_list = JsonlReader(_task2_metrics_file_path, Metrics).read()
metric_names = [
'rouge-1', 'rouge-2', 'rouge-3', 'rouge-4', 'rouge-l', 'rouge-su4',
'rouge-w-1.2'
]
rouge = Rouge(max_ngram=4,
use_porter_stemmer=True,
remove_stopwords=False,
compute_rouge_l=True,
skip_bigram_gap_length=4,
wlcs_weight=1.2)
peer_instances, peer_metrics = self._filter_by_type(
instances, metrics_list, 'peer')
reference_instances, reference_metrics = self._filter_by_type(
instances, metrics_list, 'reference')
num_to_check = 25
actual_metrics_dicts = score_instances(peer_instances[:num_to_check],
[rouge])
for expected_metrics in peer_metrics[:num_to_check]:
instance_id = expected_metrics.instance_id
summarizer_id = expected_metrics.summarizer_id
actual_metrics = actual_metrics_dicts[instance_id][summarizer_id]
for metric in metric_names:
assert actual_metrics.metrics[metric + '_jk'] == pytest.approx(
expected_metrics.metrics[metric + '_jk'], abs=1e-3)
actual_metrics_dicts = score_instances(
reference_instances[:num_to_check], [rouge])
for expected_metrics in reference_metrics[:num_to_check]:
instance_id = expected_metrics.instance_id
summarizer_id = expected_metrics.summarizer_id
actual_metrics = actual_metrics_dicts[instance_id][summarizer_id]
for metric in metric_names:
assert actual_metrics.metrics[
metric + '_jk'] == expected_metrics.metrics[metric + '_jk']