eval_entity_coref.py 文件源码

def partial_match(x, y):
    """
    Return whether these two mentions have a partial match in WordNet synset.
    :param x: the first mention
    :param y: the second mention
    :return: Whether they are aligned
    """

    # Allow partial matching
    if fuzz.partial_ratio(' ' + x + ' ', ' ' + y + ' ') == 100:
        return True

    x_words = [w for w in x.split() if not is_stop(w)]
    y_words = [w for w in y.split() if not is_stop(w)]

    if len(x_words) == 0 or len(y_words) == 0:
        return False

    x_synonyms = [set([lemma.lower().replace('_', ' ') for synset in wn.synsets(w) for lemma in synset.lemma_names()])
                  for w in x_words]
    y_synonyms = [set([lemma.lower().replace('_', ' ') for synset in wn.synsets(w) for lemma in synset.lemma_names()])
                  for w in y_words]

    # One word - check whether there is intersection between synsets
    if len(x_synonyms) == 1 and len(y_synonyms) == 1 and \
                    len([w for w in x_synonyms[0].intersection(y_synonyms[0]) if not is_stop(w)]) > 0:
        return True

    # More than one word - align words from x with words from y
    cost = -np.vstack([np.array([len([w for w in s1.intersection(s2) if not is_stop(w)]) for s1 in x_synonyms])
                       for s2 in y_synonyms])
    m = Munkres()
    cost = pad_to_square(cost)
    indices = m.compute(cost)

    # Compute the average score of the alignment
    average_score = np.mean([-cost[row, col] for row, col in indices])

    if average_score >= 0.75:
        return True

    return False