def _summary_level_lcs()

in modeling/qr_eval/qr/rouge/rouge_scorer.py [0:0]

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

def _summary_level_lcs(ref_sent, can_sent):
    """ROUGE: Summary-level LCS, section 3.2 in ROUGE paper.

  Args:
    ref_sent: list of tokenized reference sentences
    can_sent: list of tokenized candidate sentences

  Returns:
    summary level ROUGE score
  """
    if not ref_sent or not can_sent:
        return scoring.Score(precision=0, recall=0, fmeasure=0)

    m = sum(map(len, ref_sent))
    n = sum(map(len, can_sent))
    if not n or not m:
        return scoring.Score(precision=0, recall=0, fmeasure=0)

    # get token counts to prevent double counting
    token_cnts_r = collections.Counter()
    token_cnts_c = collections.Counter()
    for s in ref_sent:
        # s is a list of tokens
        token_cnts_r.update(s)
    for s in can_sent:
        token_cnts_c.update(s)

    hits = 0
    for r in ref_sent:
        lcs = _union_lcs(r, can_sent)
        # Prevent double-counting:
        # The paper describes just computing hits += len(_union_lcs()),
        # but the implementation prevents double counting. We also
        # implement this as in version 1.5.5.
        for t in lcs:
            if token_cnts_c[t] > 0 and token_cnts_r[t] > 0:
                hits += 1
                token_cnts_c[t] -= 1
                token_cnts_r[t] -= 1

    recall = hits / m
    precision = hits / n
    fmeasure = scoring.fmeasure(precision, recall)
    return scoring.Score(precision=precision, recall=recall, fmeasure=fmeasure)