def process_files(
gold_fp: ty.TextIO, sys_fp: ty.TextIO, add_sys_mentions=True
) -> ty.Iterable[str]:
gold_clusters = clusters_from_json(gold_fp)
sys_clusters = clusters_from_json(sys_fp)
if add_sys_mentions:
gold_mentions: ty.Set[ty.Hashable] = set(m for c in gold_clusters for m in c)
sys_mentions: ty.Set[ty.Hashable] = set(m for c in sys_clusters for m in c)
extra_mentions = sys_mentions - gold_mentions
gold_clusters.extend(set([m]) for m in extra_mentions)
for name, metric in METRICS.items():
R, P, F = metric(gold_clusters, sys_clusters)
yield f"{name}:\tR={R}\tP={P}\tF₁={F}\n"
conll_score = scores.conll2012(gold_clusters, sys_clusters)
yield f"CoNLL-2012 average score: {conll_score}\n"
def test_conll2012_basic(Key, Response):
score = scores.conll2012(Key, Response)
assert round(score, 2) == pytest.approx(0.46)