한국어 문서 추출요약 AI 경진대회

안녕하세요. 데이콘입니다.

본 대회에서 사용하는 평가 코드입니다.

import argparse
import pandas as pd
from tqdm import tqdm
from rouge_metric import Rouge

class RougeScorer:
    def __init__(self):
        self.rouge_evaluator = Rouge(
            metrics=["rouge-n", "rouge-l"],
            max_n=2,
            limit_length=True,
            length_limit=1000,
            length_limit_type="words",
            use_tokenizer=True,
            apply_avg=True,
            apply_best=False,
            alpha=0.5,  # Default F1_score
            weight_factor=1.2,
        )

    def compute_rouge(self, ref_df, hyp_df):
        #ref_df = pd.read_csv(ref_path)
        #hyp_df = pd.read_csv(hyp_path)
        hyp_df.iloc[:,1] = hyp_df.iloc[:,1].fillna(' ')
        ids = ref_df['id']
        hyp_df = hyp_df[hyp_df['id'].isin(ids)]
        hyp_df.index = ref_df.index

        ref_df = ref_df.sort_values(by=["id"])
        hyp_df = hyp_df.sort_values(by=["id"])
        ref_df["id"] = ref_df["id"].astype(int)
        hyp_df["id"] = hyp_df["id"].astype(int)

        hyps = [tuple(row) for row in hyp_df.values]
        refs = [tuple(row) for row in ref_df.values]

        reference_summaries = []
        generated_summaries = []

        for ref_tp, hyp_tp in zip(refs, hyps):
            ref_id, ref = ref_tp
            hyp_id, hyp = hyp_tp

            assert ref_id == hyp_id

            reference_summaries.append(ref)
            generated_summaries.append(hyp)

        scores = self.rouge_evaluator.get_scores(generated_summaries, reference_summaries)
        str_scores = self.format_rouge_scores(scores)
        #self.save_rouge_scores(str_scores)
        return str_scores

    def save_rouge_scores(self, str_scores):
        with open("rouge_scores.txt", "w") as output:
            output.write(str_scores)

    def format_rouge_scores(self, scores):
    	return "{:.3f},{:.3f},{:.3f}".format(
            scores["rouge-1"]["f"],
            scores["rouge-2"]["f"],
            scores["rouge-l"]["f"],
        )

rouge_metric.py

import os
import re
import platform
import itertools
import collections
import pkg_resources  # pip install py-rouge
from io import open


if platform.system() == "Windows":
    try:
        from eunjeon import Mecab
    except:
        print("please install eunjeon module")
else:  # Ubuntu일 경우
    from konlpy.tag import Mecab




class Rouge:
    DEFAULT_METRICS = {"rouge-n"}
    DEFAULT_N = 1
    STATS = ["f", "p", "r"]
    AVAILABLE_METRICS = {"rouge-n", "rouge-l", "rouge-w"}
    AVAILABLE_LENGTH_LIMIT_TYPES = {"words", "bytes"}
    REMOVE_CHAR_PATTERN = re.compile("[^A-Za-z0-9가-힣]")


    def __init__(
        self,
        metrics=None,
        max_n=None,
        limit_length=True,
        length_limit=1000,
        length_limit_type="words",
        apply_avg=True,
        apply_best=False,
        use_tokenizer=True,
        alpha=0.5,
        weight_factor=1.0,
    ):
        self.metrics = metrics[:] if metrics is not None else Rouge.DEFAULT_METRICS
        for m in self.metrics:
            if m not in Rouge.AVAILABLE_METRICS:
                raise ValueError("Unknown metric '{}'".format(m))


        self.max_n = max_n if "rouge-n" in self.metrics else None
        # Add all rouge-n metrics
        if self.max_n is not None:
            index_rouge_n = self.metrics.index("rouge-n")
            del self.metrics[index_rouge_n]
            self.metrics += ["rouge-{}".format(n) for n in range(1, self.max_n + 1)]
        self.metrics = set(self.metrics)


        self.limit_length = limit_length
        if self.limit_length:
            if length_limit_type not in Rouge.AVAILABLE_LENGTH_LIMIT_TYPES:
                raise ValueError("Unknown length_limit_type '{}'".format(length_limit_type))


        self.length_limit = length_limit
        if self.length_limit == 0:
            self.limit_length = False
        self.length_limit_type = length_limit_type


        self.use_tokenizer = use_tokenizer
        if use_tokenizer:
            self.tokenizer = Mecab()


        self.apply_avg = apply_avg
        self.apply_best = apply_best
        self.alpha = alpha
        self.weight_factor = weight_factor
        if self.weight_factor <= 0:
            raise ValueError("ROUGE-W weight factor must greater than 0.")


    def tokenize_text(self, text):
        if self.use_tokenizer:
            return self.tokenizer.morphs(text)
        else:
            return text


    @staticmethod
    def split_into_sentences(text):
        return text.split("\n")


    @staticmethod
    def _get_ngrams(n, text):
        ngram_set = collections.defaultdict(int)
        max_index_ngram_start = len(text) - n
        for i in range(max_index_ngram_start + 1):
            ngram_set[tuple(text[i : i + n])] += 1
        return ngram_set


    @staticmethod
    def _split_into_words(sentences):
        return list(itertools.chain(*[_.split() for _ in sentences]))


    @staticmethod
    def _get_word_ngrams_and_length(n, sentences):
        assert len(sentences) > 0
        assert n > 0


        tokens = Rouge._split_into_words(sentences)
        return Rouge._get_ngrams(n, tokens), tokens, len(tokens) - (n - 1)


    @staticmethod
    def _get_unigrams(sentences):
        assert len(sentences) > 0


        tokens = Rouge._split_into_words(sentences)
        unigram_set = collections.defaultdict(int)
        for token in tokens:
            unigram_set[token] += 1
        return unigram_set, len(tokens)


    @staticmethod
    def _compute_p_r_f_score(
        evaluated_count,
        reference_count,
        overlapping_count,
        alpha=0.5,
        weight_factor=1.0,
    ):
        precision = 0.0 if evaluated_count == 0 else overlapping_count / float(evaluated_count)
        if weight_factor != 1.0:
            precision = precision ** (1.0 / weight_factor)
        recall = 0.0 if reference_count == 0 else overlapping_count / float(reference_count)
        if weight_factor != 1.0:
            recall = recall ** (1.0 / weight_factor)
        f1_score = Rouge._compute_f_score(precision, recall, alpha)
        return {"f": f1_score, "p": precision, "r": recall}


    @staticmethod
    def _compute_f_score(precision, recall, alpha=0.5):
        return (
            0.0
            if (recall == 0.0 or precision == 0.0)
            else precision * recall / ((1 - alpha) * precision + alpha * recall)
        )


    @staticmethod
    def _compute_ngrams(evaluated_sentences, reference_sentences, n):
        if len(evaluated_sentences) <= 0 or len(reference_sentences) <= 0:
            raise ValueError("Collections must contain at least 1 sentence.")


        evaluated_ngrams, _, evaluated_count = Rouge._get_word_ngrams_and_length(
            n, evaluated_sentences
        )
        reference_ngrams, _, reference_count = Rouge._get_word_ngrams_and_length(
            n, reference_sentences
        )


        # Gets the overlapping ngrams between evaluated and reference
        overlapping_ngrams = set(evaluated_ngrams.keys()).intersection(set(reference_ngrams.keys()))
        overlapping_count = 0
        for ngram in overlapping_ngrams:
            overlapping_count += min(evaluated_ngrams[ngram], reference_ngrams[ngram])


        return evaluated_count, reference_count, overlapping_count


    @staticmethod
    def _compute_ngrams_lcs(evaluated_sentences, reference_sentences, weight_factor=1.0):
        def _lcs(x, y):
            m = len(x)
            n = len(y)
            vals = collections.defaultdict(int)
            dirs = collections.defaultdict(int)


            for i in range(1, m + 1):
                for j in range(1, n + 1):
                    if x[i - 1] == y[j - 1]:
                        vals[i, j] = vals[i - 1, j - 1] + 1
                        dirs[i, j] = "|"
                    elif vals[i - 1, j] >= vals[i, j - 1]:
                        vals[i, j] = vals[i - 1, j]
                        dirs[i, j] = "^"
                    else:
                        vals[i, j] = vals[i, j - 1]
                        dirs[i, j] = "<"


            return vals, dirs


        def _wlcs(x, y, weight_factor):
            m = len(x)
            n = len(y)
            vals = collections.defaultdict(float)
            dirs = collections.defaultdict(int)
            lengths = collections.defaultdict(int)


            for i in range(1, m + 1):
                for j in range(1, n + 1):
                    if x[i - 1] == y[j - 1]:
                        length_tmp = lengths[i - 1, j - 1]
                        vals[i, j] = (
                            vals[i - 1, j - 1]
                            + (length_tmp + 1) ** weight_factor
                            - length_tmp ** weight_factor
                        )
                        dirs[i, j] = "|"
                        lengths[i, j] = length_tmp + 1
                    elif vals[i - 1, j] >= vals[i, j - 1]:
                        vals[i, j] = vals[i - 1, j]
                        dirs[i, j] = "^"
                        lengths[i, j] = 0
                    else:
                        vals[i, j] = vals[i, j - 1]
                        dirs[i, j] = "<"
                        lengths[i, j] = 0


            return vals, dirs


        def _mark_lcs(mask, dirs, m, n):
            while m != 0 and n != 0:
                if dirs[m, n] == "|":
                    m -= 1
                    n -= 1
                    mask[m] = 1
                elif dirs[m, n] == "^":
                    m -= 1
                elif dirs[m, n] == "<":
                    n -= 1
                else:
                    raise UnboundLocalError("Illegal move")


            return mask


        if len(evaluated_sentences) <= 0 or len(reference_sentences) <= 0:
            raise ValueError("Collections must contain at least 1 sentence.")


        evaluated_unigrams_dict, evaluated_count = Rouge._get_unigrams(evaluated_sentences)
        reference_unigrams_dict, reference_count = Rouge._get_unigrams(reference_sentences)


        # Has to use weight factor for WLCS
        use_WLCS = weight_factor != 1.0
        if use_WLCS:
            evaluated_count = evaluated_count ** weight_factor
            reference_count = 0


        overlapping_count = 0.0
        for reference_sentence in reference_sentences:
            reference_sentence_tokens = reference_sentence.split()
            if use_WLCS:
                reference_count += len(reference_sentence_tokens) ** weight_factor
            hit_mask = [0 for _ in range(len(reference_sentence_tokens))]


            for evaluated_sentence in evaluated_sentences:
                evaluated_sentence_tokens = evaluated_sentence.split()


                if use_WLCS:
                    _, lcs_dirs = _wlcs(
                        reference_sentence_tokens,
                        evaluated_sentence_tokens,
                        weight_factor,
                    )
                else:
                    _, lcs_dirs = _lcs(reference_sentence_tokens, evaluated_sentence_tokens)
                _mark_lcs(
                    hit_mask,
                    lcs_dirs,
                    len(reference_sentence_tokens),
                    len(evaluated_sentence_tokens),
                )


            overlapping_count_length = 0
            for ref_token_id, val in enumerate(hit_mask):
                if val == 1:
                    token = reference_sentence_tokens[ref_token_id]
                    if evaluated_unigrams_dict[token] > 0 and reference_unigrams_dict[token] > 0:
                        evaluated_unigrams_dict[token] -= 1
                        reference_unigrams_dict[ref_token_id] -= 1


                        if use_WLCS:
                            overlapping_count_length += 1
                            if (
                                ref_token_id + 1 < len(hit_mask) and hit_mask[ref_token_id + 1] == 0
                            ) or ref_token_id + 1 == len(hit_mask):
                                overlapping_count += overlapping_count_length ** weight_factor
                                overlapping_count_length = 0
                        else:
                            overlapping_count += 1


        if use_WLCS:
            reference_count = reference_count ** weight_factor


        return evaluated_count, reference_count, overlapping_count


    def get_scores(self, hypothesis, references):
        if isinstance(hypothesis, str):
            hypothesis, references = [hypothesis], [references]


        if type(hypothesis) != type(references):
            raise ValueError("'hyps' and 'refs' are not of the same type")


        if len(hypothesis) != len(references):
            raise ValueError("'hyps' and 'refs' do not have the same length")
        scores = {}
        has_rouge_n_metric = (
            len([metric for metric in self.metrics if metric.split("-")[-1].isdigit()]) > 0
        )
        if has_rouge_n_metric:
            scores.update(self._get_scores_rouge_n(hypothesis, references))
            # scores = {**scores, **self._get_scores_rouge_n(hypothesis, references)}


        has_rouge_l_metric = (
            len([metric for metric in self.metrics if metric.split("-")[-1].lower() == "l"]) > 0
        )
        if has_rouge_l_metric:
            scores.update(self._get_scores_rouge_l_or_w(hypothesis, references, False))
            # scores = {**scores, **self._get_scores_rouge_l_or_w(hypothesis, references, False)}


        has_rouge_w_metric = (
            len([metric for metric in self.metrics if metric.split("-")[-1].lower() == "w"]) > 0
        )
        if has_rouge_w_metric:
            scores.update(self._get_scores_rouge_l_or_w(hypothesis, references, True))
            # scores = {**scores, **self._get_scores_rouge_l_or_w(hypothesis, references, True)}


        return scores


    def _get_scores_rouge_n(self, all_hypothesis, all_references):
        metrics = [metric for metric in self.metrics if metric.split("-")[-1].isdigit()]


        if self.apply_avg or self.apply_best:
            scores = {metric: {stat: 0.0 for stat in Rouge.STATS} for metric in metrics}
        else:
            scores = {
                metric: [{stat: [] for stat in Rouge.STATS} for _ in range(len(all_hypothesis))]
                for metric in metrics
            }


        for sample_id, (hypothesis, references) in enumerate(zip(all_hypothesis, all_references)):
            assert isinstance(hypothesis, str)
            has_multiple_references = False
            if isinstance(references, list):
                has_multiple_references = len(references) > 1
                if not has_multiple_references:
                    references = references[0]


            # Prepare hypothesis and reference(s)
            hypothesis = self._preprocess_summary_as_a_whole(hypothesis)
            references = (
                [self._preprocess_summary_as_a_whole(reference) for reference in references]
                if has_multiple_references
                else [self._preprocess_summary_as_a_whole(references)]
            )


            # Compute scores
            for metric in metrics:
                suffix = metric.split("-")[-1]
                n = int(suffix)


                # Aggregate
                if self.apply_avg:
                    # average model
                    total_hypothesis_ngrams_count = 0
                    total_reference_ngrams_count = 0
                    total_ngrams_overlapping_count = 0


                    for reference in references:
                        (
                            hypothesis_count,
                            reference_count,
                            overlapping_ngrams,
                        ) = Rouge._compute_ngrams(hypothesis, reference, n)
                        total_hypothesis_ngrams_count += hypothesis_count
                        total_reference_ngrams_count += reference_count
                        total_ngrams_overlapping_count += overlapping_ngrams


                    score = Rouge._compute_p_r_f_score(
                        total_hypothesis_ngrams_count,
                        total_reference_ngrams_count,
                        total_ngrams_overlapping_count,
                        self.alpha,
                    )


                    for stat in Rouge.STATS:
                        scores[metric][stat] += score[stat]
                else:
                    # Best model
                    if self.apply_best:
                        best_current_score = None
                        for reference in references:
                            (
                                hypothesis_count,
                                reference_count,
                                overlapping_ngrams,
                            ) = Rouge._compute_ngrams(hypothesis, reference, n)
                            score = Rouge._compute_p_r_f_score(
                                hypothesis_count,
                                reference_count,
                                overlapping_ngrams,
                                self.alpha,
                            )
                            if best_current_score is None or score["r"] > best_current_score["r"]:
                                best_current_score = score


                        for stat in Rouge.STATS:
                            scores[metric][stat] += best_current_score[stat]
                    # Keep all
                    else:
                        for reference in references:
                            (
                                hypothesis_count,
                                reference_count,
                                overlapping_ngrams,
                            ) = Rouge._compute_ngrams(hypothesis, reference, n)
                            score = Rouge._compute_p_r_f_score(
                                hypothesis_count,
                                reference_count,
                                overlapping_ngrams,
                                self.alpha,
                            )
                            for stat in Rouge.STATS:
                                scores[metric][sample_id][stat].append(score[stat])


        # Compute final score with the average or the the max
        if (self.apply_avg or self.apply_best) and len(all_hypothesis) > 1:
            for metric in metrics:
                for stat in Rouge.STATS:
                    scores[metric][stat] /= len(all_hypothesis)


        return scores


    def _get_scores_rouge_l_or_w(self, all_hypothesis, all_references, use_w=False):
        metric = "rouge-w" if use_w else "rouge-l"
        if self.apply_avg or self.apply_best:
            scores = {metric: {stat: 0.0 for stat in Rouge.STATS}}
        else:
            scores = {
                metric: [{stat: [] for stat in Rouge.STATS} for _ in range(len(all_hypothesis))]
            }


        for sample_id, (hypothesis_sentences, references_sentences) in enumerate(
            zip(all_hypothesis, all_references)
        ):
            assert isinstance(hypothesis_sentences, str)
            has_multiple_references = False
            if isinstance(references_sentences, list):
                has_multiple_references = len(references_sentences) > 1
                if not has_multiple_references:
                    references_sentences = references_sentences[0]


            # Prepare hypothesis and reference(s)
            hypothesis_sentences = self._preprocess_summary_per_sentence(hypothesis_sentences)
            references_sentences = (
                [
                    self._preprocess_summary_per_sentence(reference)
                    for reference in references_sentences
                ]
                if has_multiple_references
                else [self._preprocess_summary_per_sentence(references_sentences)]
            )


            # Compute scores
            # Aggregate
            if self.apply_avg:
                # average model
                total_hypothesis_ngrams_count = 0
                total_reference_ngrams_count = 0
                total_ngrams_overlapping_count = 0


                for reference_sentences in references_sentences:
                    (
                        hypothesis_count,
                        reference_count,
                        overlapping_ngrams,
                    ) = Rouge._compute_ngrams_lcs(
                        hypothesis_sentences,
                        reference_sentences,
                        self.weight_factor if use_w else 1.0,
                    )
                    total_hypothesis_ngrams_count += hypothesis_count
                    total_reference_ngrams_count += reference_count
                    total_ngrams_overlapping_count += overlapping_ngrams


                score = Rouge._compute_p_r_f_score(
                    total_hypothesis_ngrams_count,
                    total_reference_ngrams_count,
                    total_ngrams_overlapping_count,
                    self.alpha,
                    self.weight_factor if use_w else 1.0,
                )
                for stat in Rouge.STATS:
                    scores[metric][stat] += score[stat]
            else:
                # Best model
                if self.apply_best:
                    best_current_score = None
                    best_current_score_wlcs = None
                    for reference_sentences in references_sentences:
                        (
                            hypothesis_count,
                            reference_count,
                            overlapping_ngrams,
                        ) = Rouge._compute_ngrams_lcs(
                            hypothesis_sentences,
                            reference_sentences,
                            self.weight_factor if use_w else 1.0,
                        )
                        score = Rouge._compute_p_r_f_score(
                            total_hypothesis_ngrams_count,
                            total_reference_ngrams_count,
                            total_ngrams_overlapping_count,
                            self.alpha,
                            self.weight_factor if use_w else 1.0,
                        )


                        if use_w:
                            reference_count_for_score = reference_count ** (
                                1.0 / self.weight_factor
                            )
                            overlapping_ngrams_for_score = overlapping_ngrams
                            score_wlcs = (
                                overlapping_ngrams_for_score / reference_count_for_score
                            ) ** (1.0 / self.weight_factor)


                            if (
                                best_current_score_wlcs is None
                                or score_wlcs > best_current_score_wlcs
                            ):
                                best_current_score = score
                                best_current_score_wlcs = score_wlcs
                        else:
                            if best_current_score is None or score["r"] > best_current_score["r"]:
                                best_current_score = score


                    for stat in Rouge.STATS:
                        scores[metric][stat] += best_current_score[stat]
                # Keep all
                else:
                    for reference_sentences in references_sentences:
                        (
                            hypothesis_count,
                            reference_count,
                            overlapping_ngrams,
                        ) = Rouge._compute_ngrams_lcs(
                            hypothesis_sentences,
                            reference_sentences,
                            self.weight_factor if use_w else 1.0,
                        )
                        score = Rouge._compute_p_r_f_score(
                            hypothesis_count,
                            reference_count,
                            overlapping_ngrams,
                            self.alpha,
                            self.weight_factor,
                        )


                        for stat in Rouge.STATS:
                            scores[metric][sample_id][stat].append(score[stat])


        # Compute final score with the average or the the max
        if (self.apply_avg or self.apply_best) and len(all_hypothesis) > 1:
            for stat in Rouge.STATS:
                scores[metric][stat] /= len(all_hypothesis)


        return scores


    def _preprocess_summary_as_a_whole(self, summary):
        sentences = Rouge.split_into_sentences(summary)


        # Truncate
        if self.limit_length:
            # By words
            if self.length_limit_type == "words":
                summary = " ".join(sentences)
                all_tokens = summary.split()  # Counting as in the perls script
                summary = " ".join(all_tokens[: self.length_limit])


            # By bytes
            elif self.length_limit_type == "bytes":
                summary = ""
                current_len = 0
                for sentence in sentences:
                    sentence = sentence.strip()
                    sentence_len = len(sentence)


                    if current_len + sentence_len < self.length_limit:
                        if current_len != 0:
                            summary += " "
                        summary += sentence
                        current_len += sentence_len
                    else:
                        if current_len > 0:
                            summary += " "
                        summary += sentence[: self.length_limit - current_len]
                        break
        else:
            summary = " ".join(sentences)


        summary = Rouge.REMOVE_CHAR_PATTERN.sub(" ", summary.lower()).strip()


        tokens = self.tokenize_text(Rouge.REMOVE_CHAR_PATTERN.sub(" ", summary))
        preprocessed_summary = [" ".join(tokens)]


        return preprocessed_summary


    def _preprocess_summary_per_sentence(self, summary):
        sentences = Rouge.split_into_sentences(summary)


        # Truncate
        if self.limit_length:
            final_sentences = []
            current_len = 0
            # By words
            if self.length_limit_type == "words":
                for sentence in sentences:
                    tokens = sentence.strip().split()
                    tokens_len = len(tokens)
                    if current_len + tokens_len < self.length_limit:
                        sentence = " ".join(tokens)
                        final_sentences.append(sentence)
                        current_len += tokens_len
                    else:
                        sentence = " ".join(tokens[: self.length_limit - current_len])
                        final_sentences.append(sentence)
                        break
            # By bytes
            elif self.length_limit_type == "bytes":
                for sentence in sentences:
                    sentence = sentence.strip()
                    sentence_len = len(sentence)
                    if current_len + sentence_len < self.length_limit:
                        final_sentences.append(sentence)
                        current_len += sentence_len
                    else:
                        sentence = sentence[: self.length_limit - current_len]
                        final_sentences.append(sentence)
                        break
            sentences = final_sentences


        final_sentences = []
        for sentence in sentences:
            sentence = Rouge.REMOVE_CHAR_PATTERN.sub(" ", sentence.lower()).strip()


            tokens = self.tokenize_text(Rouge.REMOVE_CHAR_PATTERN.sub(" ", sentence))


            sentence = " ".join(tokens)


            final_sentences.append(sentence)


        return final_sentences