def compute_f1_for_retrieved_passage()

in utils/evaluate_retrieval.py [0:0]

• 11 lines of code
• 4 McCabe index (conditional complexity)

def compute_f1_for_retrieved_passage(line: str) -> dict:
    """
    Given a serialized JSON line, with fields 'content' and 'answer', find the closest span matching answer,
    update the deserialized dict with the span and F1 score, and return the dict.
    """
    data = json.loads(line)
    content, answer = data['content'], data['answer']

    # If there is no answer, although the closest extractive answer is '', in the MRR and recall@k functions below
    # we do not count any passage for these questions as relevant.
    if len(answer) < 1:
        data['heuristic_answer'] = ''
        data['f1'] = compute_f1(answer, '')
        return data

    best_span, best_f1 = find_closest_span_match(content, answer)

    data['heuristic_answer'] = best_span
    data['f1'] = best_f1

    return data