def test_evaluate(self):
"""Test that comparing an SDP graph against itself returns perfect F1"""
for _, ref, sdp_id in read_test_sdp():
scores = evaluate(ref, ref)
self.assertAlmostEqual(scores.average_f1(), 1)
def evaluate_sdp(*args, **kwargs):
from semstr.evaluation import sdp
return sdp.evaluate(*args, **kwargs)
def convert_and_evaluate(self, passage, ref):
converted = to_sdp(passage)
scores = evaluate(converted, ref)
self.assertAlmostEqual(scores.average_f1(), 1, msg="\n" + "\n".join(converted))