def calculate_overlap_scores(self):
for qa in self.original_qas:
ans = qa['predicted_answer']
gt_ans = qa['answers'][0]['text']
# store all word overlap metrics between predicted answer and original answer
qa['exact_match'] = exact_match_metric(ans, gt_ans)
qa['f1_match'] = f1_metric(ans, gt_ans)
qa['recall_match'] = recall_metric(ans, gt_ans)
qa['precision_match'] = precision_metric(ans, gt_ans)
qa['unanswerable'] = qa['predicted_answer'] == ''
def find_parent(sq):
# find precision of predicted answer with reference general question answer
gq_precisions = [
precision_metric(sq['predicted_answer'],
gq['answers'][0]['text'])
for gq in general_questions
]
if np.max(gq_precisions) > 0.5:
# precision is high enough, return the top point
return np.argmax(gq_precisions)
else:
# if precision is too low, resort to the answer positioning heuristic
current_pos = sq['answers'][0]['answer_start']
for i, gq in enumerate(general_questions):
if gq['answers'][0]['answer_start'] > current_pos:
return i - 1
return len(general_questions) - 1