web_document_filtering.py

import json
import re
from collections import Counter

import fasttext
import kenlm
import numpy as np
import sentencepiece

from obelics.utils import SPECIAL_CHARACTERS


class FilteringFunctions:
    @staticmethod
    def check_format(image_metadata, valid_formats):
        if "format" in image_metadata:
            if image_metadata["format"] not in valid_formats:
                return False
        return True

    @staticmethod
    def check_size_image(
        image_metadata,
        original_width_min_cutoff,
        original_width_max_cutoff,
        original_height_min_cutoff,
        original_height_max_cutoff,
        rendered_width_min_cutoff,
        rendered_width_max_cutoff,
        rendered_height_min_cutoff,
        rendered_height_max_cutoff,
        aspect_ratio_max_cutoff,
    ):
        if not (original_width_min_cutoff <= image_metadata["original_width"] <= original_width_max_cutoff):
            return False
        if not (original_height_min_cutoff <= image_metadata["original_height"] <= original_height_max_cutoff):
            return False
        if "rendered_width" in image_metadata:
            if not (rendered_width_min_cutoff <= image_metadata["rendered_width"] <= rendered_width_max_cutoff):
                return False
        if "rendered_height" in image_metadata:
            if not (rendered_height_min_cutoff <= image_metadata["rendered_height"] <= rendered_height_max_cutoff):
                return False
        if not (
            1 / aspect_ratio_max_cutoff
            <= image_metadata["original_width"] / image_metadata["original_height"]
            <= aspect_ratio_max_cutoff
        ):
            return False
        return True

    @staticmethod
    def check_number_images(number_images, number_images_min_cutoff, number_images_max_cutoff):
        if (number_images < number_images_min_cutoff) or (number_images > number_images_max_cutoff):
            return False
        return True

    @staticmethod
    def remove_empty_el_from_list(list_):
        return [el for el in list_ if el]

    @staticmethod
    def remove_non_printing_characters(text, non_printing_characters_re):
        return non_printing_characters_re.sub("", text)

    @staticmethod
    def standardize_whitespace(
        text,
        whitespace=[
            " ",
            " ",
            " ",
            " ",
            " ",
            "　",
            " ",
            " ",
            " ",
            " ",
            "",
            "„",
            " ",
        ],
    ):
        """There are different whitespace characters."""
        whitespace = set(whitespace)
        text = "".join([char if char not in whitespace else " " for char in text])
        return text

    @staticmethod
    def split_on_whitespace(
        text,
        new_line=False,
        tab=False,
    ):
        """This method also removes concatenated spaces."""
        sep = [" "] + new_line * ["\n"] + tab * ["\t"]
        sep = "|".join(sep)
        split_text = re.split(sep, text)
        split_text = FilteringFunctions.remove_empty_el_from_list(split_text)
        return split_text

    @staticmethod
    def strip(text, strip_characters):
        """Way faster than text.strip(strip_characters)
        since strip_characters is a set instead of a str,
        and it contains a lot of elements (all the emojis)."""
        if not text:
            return text
        beg_ind = 0
        end_ind = len(text)
        for i in range(len(text)):
            if text[i] in strip_characters:
                beg_ind += 1
            else:
                break
        for i in range(1, len(text) + 1):
            if text[-i] in strip_characters:
                end_ind -= 1
            else:
                break
        text_stripped = text[beg_ind:end_ind]
        return text_stripped

    @staticmethod
    def get_words_from_text(text, lower_case=True, strip_words=True, strip_characters=SPECIAL_CHARACTERS):
        """Get words from a text. Non reversible since the text
        is split on multiple characters, words are stripped of
        special characters and characters are converted to lower case.
        Useful to compute ratios, like the stopword ratio."""
        if strip_words and strip_characters is None:
            raise ValueError("strip_characters must be provided if strip_words is True.")
        words = FilteringFunctions.split_on_whitespace(text=text, new_line=True, tab=True)
        if lower_case:
            words = [word.lower() for word in words]
        if strip_words:
            words = [FilteringFunctions.strip(word, strip_characters) for word in words]
            words = FilteringFunctions.remove_empty_el_from_list(words)
        return words

    @staticmethod
    def check_number_words(text, strip_characters, number_words_min_cutoff, number_words_max_cutoff):
        words = FilteringFunctions.get_words_from_text(
            text=text, lower_case=True, strip_words=True, strip_characters=strip_characters
        )
        number_words = len(words)
        if (number_words < number_words_min_cutoff) or (number_words > number_words_max_cutoff):
            return False
        return True

    @staticmethod
    def compute_character_repetition_ratio(text, character_repetition_length):
        def get_freq_character_ngrams(text, n):
            character_ngrams = [text[i : i + n] for i in range(len(text) - n + 1)]
            freq_character_ngrams = Counter(character_ngrams)
            return freq_character_ngrams

        freq_character_ngrams = get_freq_character_ngrams(text=text, n=character_repetition_length)
        if len(freq_character_ngrams) == 0:
            return 0
        freq_character_ngrams = list(freq_character_ngrams.values())
        freq_character_ngrams = sorted(freq_character_ngrams, reverse=True)
        val_one = len([el for el in freq_character_ngrams if el == 1])
        num_rep_character_ngrams = min(
            int(np.sqrt(len(freq_character_ngrams))),
            len(freq_character_ngrams) - val_one,
        )
        character_repetition_ratio = sum(freq_character_ngrams[:num_rep_character_ngrams]) / sum(freq_character_ngrams)
        return character_repetition_ratio

    @staticmethod
    def check_character_repetition_ratio(
        text,
        character_repetition_length,
        character_repetition_max_cutoff,
    ):
        character_repetition_ratio = FilteringFunctions.compute_character_repetition_ratio(
            text=text, character_repetition_length=character_repetition_length
        )
        if character_repetition_ratio > character_repetition_max_cutoff:
            return False
        return True

    @staticmethod
    def compute_word_repetition_ratio(text, strip_characters, word_repetition_length):
        def get_freq_word_ngrams(text, strip_characters, n):
            words = FilteringFunctions.get_words_from_text(
                text=text, lower_case=True, strip_words=True, strip_characters=strip_characters
            )
            word_ngrams = [" ".join(words[i : i + n]) for i in range(len(words) - n + 1)]
            freq_word_ngrams = Counter(word_ngrams)
            return freq_word_ngrams

        freq_word_ngrams = get_freq_word_ngrams(text=text, strip_characters=strip_characters, n=word_repetition_length)
        if len(freq_word_ngrams) == 0:
            return 0
        freq_word_ngrams = list(freq_word_ngrams.values())
        word_repetition_ratio = sum(freq for freq in freq_word_ngrams if freq > 1) / sum(freq_word_ngrams)
        return word_repetition_ratio

    @staticmethod
    def check_word_repetition_ratio(
        text,
        strip_characters,
        word_repetition_length,
        word_repetition_max_cutoff,
    ):
        word_repetition_ratio = FilteringFunctions.compute_word_repetition_ratio(
            text=text, strip_characters=strip_characters, word_repetition_length=word_repetition_length
        )
        cond = word_repetition_ratio <= word_repetition_max_cutoff
        return cond

    @staticmethod
    def compute_special_character_ratio(text, special_characters):
        if len(text) == 0:
            return 0
        special_character_ratio = len([char for char in text if char in special_characters]) / len(text)
        return special_character_ratio

    @staticmethod
    def check_special_character_ratio(text, special_characters, special_character_ratio_max_cutoff):
        special_character_ratio = FilteringFunctions.compute_special_character_ratio(
            text=text, special_characters=special_characters
        )
        if special_character_ratio > special_character_ratio_max_cutoff:
            return False
        return True

    @staticmethod
    def compute_stopword_ratio(text, strip_characters, stopwords):
        words = FilteringFunctions.get_words_from_text(
            text=text, lower_case=True, strip_words=True, strip_characters=strip_characters
        )
        if not words:
            return 0
        stopword_ratio = len([word for word in words if word in stopwords]) / len(words)
        return stopword_ratio

    @staticmethod
    def check_stopword_ratio(text, strip_characters, stopwords, stopword_ratio_min_cutoff):
        stopword_ratio = FilteringFunctions.compute_stopword_ratio(
            text=text, strip_characters=strip_characters, stopwords=stopwords
        )
        if stopword_ratio < stopword_ratio_min_cutoff:
            return False
        return True

    @staticmethod
    def compute_flagged_word_ratio(text, strip_characters, flagged_words):
        words = FilteringFunctions.get_words_from_text(
            text=text, lower_case=True, strip_words=True, strip_characters=strip_characters
        )
        if not words:
            return 0
        flagged_word_ratio = len([word for word in words if word in flagged_words]) / len(words)
        return flagged_word_ratio

    @staticmethod
    def check_flagged_word_ratio(
        text,
        strip_characters,
        flagged_words,
        flagged_word_ratio_max_cutoff,
    ):
        flagged_word_ratio = FilteringFunctions.compute_flagged_word_ratio(
            text=text,
            strip_characters=strip_characters,
            flagged_words=flagged_words,
        )
        if flagged_word_ratio > flagged_word_ratio_max_cutoff:
            return False
        return True

    @staticmethod
    def compute_punctuation_ratio(text, punctuation, min_nb_words=-1):
        # The regepx is used to extract all words and punctuation marks from the text string.
        # `[\w']+`: This matches any sequence of one or more word characters (letters, digits, or underscores) or apostrophes.
        # The square brackets indicate a character class, which means "match any of these characters".
        # The backslash before the w means to escape it as it has a special meaning in regex,
        # and the plus sign means "one or more occurrences". We choose to ignore other special characters because in practise,
        # it's a good enough approximation to filter web documents that don't contain long sequences of (English) text.
        # `|`: This is the "or" operator in regex. It separates the two alternative patterns that the regex can match.
        # `[punctuation]`: This matches any one of the punctuation marks specified inside the square brackets.
        # This is inspired from https://stackoverflow.com/a/367292/6226208
        punc_splitters = "".join(punctuation)
        pattern = r"{}".format(f"[\w']+|[{punc_splitters}]")
        punctuation_splitted_words = re.findall(pattern, text)
        if not punctuation_splitted_words:
            return 0
        if min_nb_words > 0:
            if len(punctuation_splitted_words) < min_nb_words:
                # If the sequence is too short, we don't want to rely on punctuation to filter it out.
                return 1.0
        punctuation_ratio = len([word for word in punctuation_splitted_words if word in punctuation]) / len(
            punctuation_splitted_words
        )
        return punctuation_ratio

    @staticmethod
    def check_punctuation_ratio(
        text,
        punctuation,
        punctuation_ratio_min_cutoff,
        min_nb_words=-1,
    ):
        punctuation_ratio = FilteringFunctions.compute_punctuation_ratio(
            text=text,
            punctuation=punctuation,
            min_nb_words=min_nb_words,
        )
        if punctuation_ratio < punctuation_ratio_min_cutoff:
            return False
        return True

    @staticmethod
    def compute_common_word_ratio(text, strip_characters, common_words):
        words = FilteringFunctions.get_words_from_text(
            text=text, lower_case=True, strip_words=True, strip_characters=strip_characters
        )
        if not words:
            return 0
        common_word_ratio = len([word for word in words if word in common_words]) / len(words)
        return common_word_ratio

    @staticmethod
    def check_common_word_ratio(
        text,
        strip_characters,
        common_words,
        common_word_ratio_min_cutoff,
    ):
        common_word_ratio = FilteringFunctions.compute_common_word_ratio(
            text=text,
            strip_characters=strip_characters,
            common_words=common_words,
        )
        if common_word_ratio < common_word_ratio_min_cutoff:
            return False
        return True

    @staticmethod
    def compute_lang_id_pred_score(text, lang_id_model):
        text = text.lower().replace("\n", " ")
        tot_pred_lang_id = lang_id_model.predict(text)
        pred_lang_id = tot_pred_lang_id[0][0].replace("__label__", "")
        score_pred_lang_id = tot_pred_lang_id[1][0]
        return pred_lang_id, score_pred_lang_id

    @staticmethod
    def check_lang_id(text, lang_id_model, target_lang_id, lang_id_min_cutoff):
        pred_lang_id, score_pred_lang_id = FilteringFunctions.compute_lang_id_pred_score(
            text=text, lang_id_model=lang_id_model
        )
        if (pred_lang_id != target_lang_id) or (score_pred_lang_id < lang_id_min_cutoff):
            return False
        return True

    @staticmethod
    def replace_digits_with_zeros(text, digits_re):
        return digits_re.sub("0", text)

    @staticmethod
    def replace_unicode_punctuation(text, unicode_punctuation):
        return "".join(unicode_punctuation.get(c, c) for c in text)

    @staticmethod
    def normalization(
        text,
        remove_non_printing_characters,
        strip,
        lower_case,
        standardize_whitespace,
        replace_digits_with_zeros,
        replace_unicode_punctuation,
        non_printing_characters_re,
        digits_re,
        unicode_punctuation,
    ):
        if remove_non_printing_characters:
            text = FilteringFunctions.remove_non_printing_characters(
                text=text, non_printing_characters_re=non_printing_characters_re
            )
        if strip:
            text = text.strip()
        if not text:
            return text
        if lower_case:
            text = text.lower()
        if standardize_whitespace:
            text = FilteringFunctions.standardize_whitespace(text=text)
        if replace_digits_with_zeros:
            text = FilteringFunctions.replace_digits_with_zeros(text=text, digits_re=digits_re)
        if replace_unicode_punctuation:
            text = FilteringFunctions.replace_unicode_punctuation(text=text, unicode_punctuation=unicode_punctuation)
        return text

    @staticmethod
    def tokenization(text, sentencepiece_model, join_on_whitespace):
        text_tokenized = sentencepiece_model.encode_as_pieces(text)
        if join_on_whitespace:
            text_tokenized = " ".join(text_tokenized)
        return text_tokenized

    @staticmethod
    def compute_perplexity_score(
        text, non_printing_characters_re, digits_re, unicode_punctuation, sentencepiece_model, kenlm_model
    ):
        text = FilteringFunctions.normalization(
            text=text,
            remove_non_printing_characters=True,
            strip=True,
            lower_case=False,
            standardize_whitespace=True,
            replace_digits_with_zeros=True,
            replace_unicode_punctuation=True,
            non_printing_characters_re=non_printing_characters_re,
            digits_re=digits_re,
            unicode_punctuation=unicode_punctuation,
        )
        text = FilteringFunctions.tokenization(
            text=text, sentencepiece_model=sentencepiece_model, join_on_whitespace=True
        )
        doc_log_score, doc_length = 0, 0
        for line in text.split("\n"):
            log_score = kenlm_model.score(line)
            length = len(line.split()) + 1
            doc_log_score += log_score
            doc_length += length
        pp_score = 10.0 ** (-doc_log_score / doc_length)
        pp_score = round(pp_score, 1)
        return pp_score

    @staticmethod
    def check_perplexity_score(
        text,
        non_printing_characters_re,
        digits_re,
        unicode_punctuation,
        sentencepiece_model,
        kenlm_model,
        perplexity_score_max_cutoff,
    ):
        perplexity_score = FilteringFunctions.compute_perplexity_score(
            text=text,
            non_printing_characters_re=non_printing_characters_re,
            digits_re=digits_re,
            unicode_punctuation=unicode_punctuation,
            sentencepiece_model=sentencepiece_model,
            kenlm_model=kenlm_model,
        )
        if perplexity_score > perplexity_score_max_cutoff:
            return False
        return True


class WebDocumentFilteringNodeLevel:
    # Needed to make multiprocessing work
    __slots__ = (
        "cond_check_format",
        "valid_formats",
        "cond_check_size_image",
        "original_width_min_cutoff",
        "original_width_max_cutoff",
        "original_height_min_cutoff",
        "original_height_max_cutoff",
        "rendered_width_min_cutoff",
        "rendered_width_max_cutoff",
        "rendered_height_min_cutoff",
        "rendered_height_max_cutoff",
        "aspect_ratio_max_cutoff",
        "cond_remove_non_printing_characters",
        "non_printing_characters_re",
        "cond_standardize_whitespace",
        "cond_check_number_words_node_level",
        "strip_characters",
        "number_words_node_level_min_cutoff",
        "number_words_node_level_max_cutoff",
        "cond_check_character_repetition_ratio_node_level",
        "character_repetition_length_node_level",
        "character_repetition_node_level_max_cutoff",
        "cond_check_word_repetition_ratio_node_level",
        "word_repetition_length_node_level",
        "word_repetition_node_level_max_cutoff",
        "cond_check_special_character_ratio_node_level",
        "special_character_ratio_node_level_max_cutoff",
        "cond_check_stopword_ratio_node_level",
        "stopwords",
        "stopword_ratio_node_level_min_cutoff",
        "cond_check_flagged_word_ratio_node_level",
        "flagged_words",
        "flagged_word_ratio_node_level_max_cutoff",
        "cond_check_punctuation_ratio_node_level",
        "min_number_words_to_check_punctuation_ratio_node_level",
        "punctuation",
        "punctuation_ratio_node_level_min_cutoff",
        "cond_check_common_word_ratio_node_level",
        "path_common_words",
        "common_words",
        "common_word_ratio_node_level_min_cutoff",
        "cond_check_lang_id_node_level",
        "path_lang_id_model",
        "lang_id_model",
        "lang_id_node_level_min_cutoff",
        "cond_check_perplexity_score_node_level",
        "digits_re",
        "unicode_punctuation",
        "path_sentencepiece_model",
        "sentencepiece_model",
        "path_kenlm_model",
        "kenlm_model",
        "perplexity_score_node_level_max_cutoff",
    )

    def __init__(
        self,
        cond_check_format,
        valid_formats,
        cond_check_size_image,
        original_width_min_cutoff,
        original_width_max_cutoff,
        original_height_min_cutoff,
        original_height_max_cutoff,
        rendered_width_min_cutoff,
        rendered_width_max_cutoff,
        rendered_height_min_cutoff,
        rendered_height_max_cutoff,
        aspect_ratio_max_cutoff,
        cond_remove_non_printing_characters,
        non_printing_characters_re,
        cond_standardize_whitespace,
        cond_check_number_words_node_level,
        strip_characters,
        number_words_node_level_min_cutoff,
        number_words_node_level_max_cutoff,
        cond_check_character_repetition_ratio_node_level,
        character_repetition_length_node_level,
        character_repetition_node_level_max_cutoff,
        cond_check_word_repetition_ratio_node_level,
        word_repetition_length_node_level,
        word_repetition_node_level_max_cutoff,
        cond_check_special_character_ratio_node_level,
        special_character_ratio_node_level_max_cutoff,
        cond_check_stopword_ratio_node_level,
        stopwords,
        stopword_ratio_node_level_min_cutoff,
        cond_check_flagged_word_ratio_node_level,
        flagged_words,
        flagged_word_ratio_node_level_max_cutoff,
        cond_check_punctuation_ratio_node_level,
        min_number_words_to_check_punctuation_ratio_node_level,
        punctuation,
        punctuation_ratio_node_level_min_cutoff,
        cond_check_common_word_ratio_node_level,
        path_common_words,
        common_word_ratio_node_level_min_cutoff,
        cond_check_lang_id_node_level,
        path_lang_id_model,
        lang_id_node_level_min_cutoff,
        cond_check_perplexity_score_node_level,
        digits_re,
        unicode_punctuation,
        path_sentencepiece_model,
        path_kenlm_model,
        perplexity_score_node_level_max_cutoff,
    ):
        self.cond_check_format = cond_check_format
        self.valid_formats = valid_formats

        self.cond_check_size_image = cond_check_size_image
        self.original_width_min_cutoff = original_width_min_cutoff
        self.original_width_max_cutoff = original_width_max_cutoff
        self.original_height_min_cutoff = original_height_min_cutoff
        self.original_height_max_cutoff = original_height_max_cutoff
        self.rendered_width_min_cutoff = rendered_width_min_cutoff
        self.rendered_width_max_cutoff = rendered_width_max_cutoff
        self.rendered_height_min_cutoff = rendered_height_min_cutoff
        self.rendered_height_max_cutoff = rendered_height_max_cutoff
        self.aspect_ratio_max_cutoff = aspect_ratio_max_cutoff

        self.cond_remove_non_printing_characters = cond_remove_non_printing_characters
        self.non_printing_characters_re = non_printing_characters_re

        self.cond_standardize_whitespace = cond_standardize_whitespace

        self.cond_check_number_words_node_level = cond_check_number_words_node_level
        self.strip_characters = strip_characters
        self.number_words_node_level_min_cutoff = number_words_node_level_min_cutoff
        self.number_words_node_level_max_cutoff = number_words_node_level_max_cutoff

        self.cond_check_character_repetition_ratio_node_level = cond_check_character_repetition_ratio_node_level
        self.character_repetition_length_node_level = character_repetition_length_node_level
        self.character_repetition_node_level_max_cutoff = character_repetition_node_level_max_cutoff

        self.cond_check_word_repetition_ratio_node_level = cond_check_word_repetition_ratio_node_level
        self.word_repetition_length_node_level = word_repetition_length_node_level
        self.word_repetition_node_level_max_cutoff = word_repetition_node_level_max_cutoff

        self.cond_check_special_character_ratio_node_level = cond_check_special_character_ratio_node_level
        self.special_character_ratio_node_level_max_cutoff = special_character_ratio_node_level_max_cutoff

        self.cond_check_stopword_ratio_node_level = cond_check_stopword_ratio_node_level
        self.stopwords = stopwords
        self.stopword_ratio_node_level_min_cutoff = stopword_ratio_node_level_min_cutoff

        self.cond_check_flagged_word_ratio_node_level = cond_check_flagged_word_ratio_node_level
        self.flagged_words = flagged_words
        self.flagged_word_ratio_node_level_max_cutoff = flagged_word_ratio_node_level_max_cutoff

        self.cond_check_punctuation_ratio_node_level = cond_check_punctuation_ratio_node_level
        self.min_number_words_to_check_punctuation_ratio_node_level = (
            min_number_words_to_check_punctuation_ratio_node_level
        )
        self.punctuation = punctuation
        self.punctuation_ratio_node_level_min_cutoff = punctuation_ratio_node_level_min_cutoff

        self.cond_check_common_word_ratio_node_level = cond_check_common_word_ratio_node_level
        self.path_common_words = path_common_words
        with open(path_common_words) as f:
            self.common_words = json.load(f)
        self.common_word_ratio_node_level_min_cutoff = common_word_ratio_node_level_min_cutoff

        self.cond_check_lang_id_node_level = cond_check_lang_id_node_level
        self.path_lang_id_model = path_lang_id_model
        if cond_check_lang_id_node_level:
            self.lang_id_model = fasttext.load_model(path_lang_id_model)
        self.lang_id_node_level_min_cutoff = lang_id_node_level_min_cutoff

        self.cond_check_perplexity_score_node_level = cond_check_perplexity_score_node_level
        self.digits_re = digits_re
        self.unicode_punctuation = unicode_punctuation
        self.path_sentencepiece_model = path_sentencepiece_model
        if cond_check_perplexity_score_node_level:
            self.sentencepiece_model = sentencepiece.SentencePieceProcessor()
            self.sentencepiece_model.load(path_sentencepiece_model)
        self.path_kenlm_model = path_kenlm_model
        if cond_check_perplexity_score_node_level:
            self.kenlm_model = kenlm.Model(path_kenlm_model)
        self.perplexity_score_node_level_max_cutoff = perplexity_score_node_level_max_cutoff

    def __call__(self, web_document):
        texts = web_document["texts"]
        images = web_document["images"]
        metadata = json.loads(web_document["metadata"])

        indices_to_remove = set()

        for ind, (text, image, meta) in enumerate(zip(texts, images, metadata)):
            if image is not None:
                meta["original_width"], meta["original_height"] = image.size
                for side in ["width", "height"]:
                    if (
                        (f"rendered_{side}" in meta)
                        and isinstance(meta[f"rendered_{side}"], str)
                        and ("%" in meta[f"rendered_{side}"])
                    ):
                        try:
                            meta[f"rendered_{side}"] = round(
                                meta[f"original_{side}"] * int(meta[f"rendered_{side}"].replace("%", "")) / 100
                            )
                        except Exception:
                            del meta[f"rendered_{side}"]
                if image.format:
                    meta["format"] = image.format.lower()

                if self.cond_check_format:
                    if not FilteringFunctions.check_format(image_metadata=meta, valid_formats=self.valid_formats):
                        indices_to_remove.add(ind)
                        continue

                if self.cond_check_size_image:
                    if not FilteringFunctions.check_size_image(
                        image_metadata=meta,
                        original_width_min_cutoff=self.original_width_min_cutoff,
                        original_width_max_cutoff=self.original_width_max_cutoff,
                        original_height_min_cutoff=self.original_height_min_cutoff,
                        original_height_max_cutoff=self.original_height_max_cutoff,
                        rendered_width_min_cutoff=self.rendered_width_min_cutoff,
                        rendered_width_max_cutoff=self.rendered_width_max_cutoff,
                        rendered_height_min_cutoff=self.rendered_height_min_cutoff,
                        rendered_height_max_cutoff=self.rendered_height_max_cutoff,
                        aspect_ratio_max_cutoff=self.aspect_ratio_max_cutoff,
                    ):
                        indices_to_remove.add(ind)
                        continue

                metadata[ind] = meta

            elif text is not None:
                if self.cond_remove_non_printing_characters:
                    text = FilteringFunctions.remove_non_printing_characters(
                        text=text, non_printing_characters_re=self.non_printing_characters_re
                    )

                if self.cond_standardize_whitespace:
                    text = FilteringFunctions.standardize_whitespace(text=text)

                paragraphs = text.split("\n\n")
                paragraphs_indices_to_remove = set()

                for ind_par, paragraph in enumerate(paragraphs):
                    if paragraph == "END_OF_DOCUMENT_TOKEN_TO_BE_REPLACED":
                        continue

                    if self.cond_check_number_words_node_level:
                        if not FilteringFunctions.check_number_words(
                            text=paragraph,
                            strip_characters=self.strip_characters,
                            number_words_min_cutoff=self.number_words_node_level_min_cutoff,
                            number_words_max_cutoff=self.number_words_node_level_max_cutoff,
                        ):
                            paragraphs_indices_to_remove.add(ind_par)
                            continue

                    if self.cond_check_character_repetition_ratio_node_level:
                        if not FilteringFunctions.check_character_repetition_ratio(
                            text=paragraph,
                            character_repetition_length=self.character_repetition_length_node_level,
                            character_repetition_max_cutoff=self.character_repetition_node_level_max_cutoff,
                        ):
                            paragraphs_indices_to_remove.add(ind_par)
                            continue

                    if self.cond_check_word_repetition_ratio_node_level:
                        if not FilteringFunctions.check_word_repetition_ratio(
                            text=paragraph,
                            strip_characters=self.strip_characters,
                            word_repetition_length=self.word_repetition_length_node_level,
                            word_repetition_max_cutoff=self.word_repetition_node_level_max_cutoff,
                        ):
                            paragraphs_indices_to_remove.add(ind_par)
                            continue

                    if self.cond_check_special_character_ratio_node_level:
                        if not FilteringFunctions.check_special_character_ratio(
                            text=paragraph,
                            special_characters=self.strip_characters,
                            special_character_ratio_max_cutoff=self.special_character_ratio_node_level_max_cutoff,
                        ):
                            paragraphs_indices_to_remove.add(ind_par)
                            continue

                    if self.cond_check_stopword_ratio_node_level:
                        if not FilteringFunctions.check_stopword_ratio(
                            text=paragraph,
                            strip_characters=self.strip_characters,
                            stopwords=self.stopwords,
                            stopword_ratio_min_cutoff=self.stopword_ratio_node_level_min_cutoff,
                        ):
                            paragraphs_indices_to_remove.add(ind_par)
                            continue

                    if self.cond_check_flagged_word_ratio_node_level:
                        if not FilteringFunctions.check_flagged_word_ratio(
                            text=paragraph,
                            strip_characters=self.strip_characters,
                            flagged_words=self.flagged_words,
                            flagged_word_ratio_max_cutoff=self.flagged_word_ratio_node_level_max_cutoff,
                        ):
                            paragraphs_indices_to_remove.add(ind_par)
                            continue

                    if self.cond_check_punctuation_ratio_node_level:
                        if not FilteringFunctions.check_punctuation_ratio(
                            text=paragraph,
                            punctuation=self.punctuation,
                            punctuation_ratio_min_cutoff=self.punctuation_ratio_node_level_min_cutoff,
                            min_nb_words=self.min_number_words_to_check_punctuation_ratio_node_level,
                        ):
                            paragraphs_indices_to_remove.add(ind_par)
                            continue

                    if self.cond_check_common_word_ratio_node_level:
                        if not FilteringFunctions.check_common_word_ratio(
                            text=paragraph,
                            strip_characters=self.strip_characters,
                            common_words=self.common_words,
                            common_word_ratio_min_cutoff=self.common_word_ratio_node_level_min_cutoff,
                        ):
                            paragraphs_indices_to_remove.add(ind_par)
                            continue

                    if self.cond_check_lang_id_node_level:
                        if not FilteringFunctions.check_lang_id(
                            text=paragraph,
                            lang_id_model=self.lang_id_model,
                            target_lang_id="en",
                            lang_id_min_cutoff=self.lang_id_node_level_min_cutoff,
                        ):
                            paragraphs_indices_to_remove.add(ind_par)
                            continue

                    if self.cond_check_perplexity_score_node_level:
                        if not FilteringFunctions.check_perplexity_score(
                            text=paragraph,
                            non_printing_characters_re=self.non_printing_characters_re,
                            digits_re=self.digits_re,
                            unicode_punctuation=self.unicode_punctuation,
                            sentencepiece_model=self.sentencepiece_model,
                            kenlm_model=self.kenlm_model,
                            perplexity_score_max_cutoff=self.perplexity_score_node_level_max_cutoff,
                        ):
                            paragraphs_indices_to_remove.add(ind_par)
                            continue

                paragraphs = [
                    el for ind_par, el in enumerate(paragraphs) if ind_par not in paragraphs_indices_to_remove
                ]
                if not paragraphs:
                    indices_to_remove.add(ind)
                else:
                    texts[ind] = "\n\n".join(paragraphs)

            else:
                indices_to_remove.add(ind)

        web_document["texts"] = [el for ind, el in enumerate(texts) if ind not in indices_to_remove]
        web_document["images"] = [el for ind, el in enumerate(images) if ind not in indices_to_remove]
        web_document["metadata"] = json.dumps([el for ind, el in enumerate(metadata) if ind not in indices_to_remove])

        return web_document

    # Needed to make multiprocessing work
    def __reduce__(self):
        return (
            self.__class__,
            (
                self.cond_check_format,
                self.valid_formats,
                self.cond_check_size_image,
                self.original_width_min_cutoff,
                self.original_width_max_cutoff,
                self.original_height_min_cutoff,
                self.original_height_max_cutoff,
                self.rendered_width_min_cutoff,
                self.rendered_width_max_cutoff,
                self.rendered_height_min_cutoff,
                self.rendered_height_max_cutoff,
                self.aspect_ratio_max_cutoff,
                self.cond_remove_non_printing_characters,
                self.non_printing_characters_re,
                self.cond_standardize_whitespace,
                self.cond_check_number_words_node_level,
                self.strip_characters,
                self.number_words_node_level_min_cutoff,
                self.number_words_node_level_max_cutoff,
                self.cond_check_character_repetition_ratio_node_level,
                self.character_repetition_length_node_level,
                self.character_repetition_node_level_max_cutoff,
                self.cond_check_word_repetition_ratio_node_level,
                self.word_repetition_length_node_level,
                self.word_repetition_node_level_max_cutoff,
                self.cond_check_special_character_ratio_node_level,
                self.special_character_ratio_node_level_max_cutoff,
                self.cond_check_stopword_ratio_node_level,
                self.stopwords,
                self.stopword_ratio_node_level_min_cutoff,
                self.cond_check_flagged_word_ratio_node_level,
                self.flagged_words,
                self.flagged_word_ratio_node_level_max_cutoff,
                self.cond_check_punctuation_ratio_node_level,
                self.min_number_words_to_check_punctuation_ratio_node_level,
                self.punctuation,
                self.punctuation_ratio_node_level_min_cutoff,
                self.cond_check_common_word_ratio_node_level,
                self.path_common_words,
                self.common_word_ratio_node_level_min_cutoff,
                self.cond_check_lang_id_node_level,
                self.path_lang_id_model,
                self.lang_id_node_level_min_cutoff,
                self.cond_check_perplexity_score_node_level,
                self.digits_re,
                self.unicode_punctuation,
                self.path_sentencepiece_model,
                self.path_kenlm_model,
                self.perplexity_score_node_level_max_cutoff,
            ),
        )


class WebDocumentFilteringDocLevel:
    def __init__(
        self,
        cond_check_number_images,
        number_images_min_cutoff,
        number_images_max_cutoff,
        cond_check_number_words_doc_level,
        strip_characters,
        number_words_doc_level_min_cutoff,
        number_words_doc_level_max_cutoff,
        cond_check_character_repetition_ratio_doc_level,
        character_repetition_length_doc_level,
        character_repetition_doc_level_max_cutoff,
        cond_check_word_repetition_ratio_doc_level,
        word_repetition_length_doc_level,
        word_repetition_doc_level_max_cutoff,
        cond_check_special_character_ratio_doc_level,
        special_character_ratio_doc_level_max_cutoff,
        cond_check_stopword_ratio_doc_level,
        stopwords,
        stopword_ratio_doc_level_min_cutoff,
        cond_check_flagged_word_ratio_doc_level,
        flagged_words,
        flagged_word_ratio_doc_level_max_cutoff,
        cond_check_punctuation_ratio_doc_level,
        punctuation,
        punctuation_ratio_doc_level_min_cutoff,
        cond_check_common_word_ratio_doc_level,
        path_common_words,
        common_word_ratio_doc_level_min_cutoff,
        cond_check_lang_id_doc_level,
        path_lang_id_model,
        lang_id_doc_level_min_cutoff,
        cond_check_perplexity_score_doc_level,
        non_printing_characters_re,
        digits_re,
        unicode_punctuation,
        path_sentencepiece_model,
        path_kenlm_model,
        perplexity_score_doc_level_max_cutoff,
    ):
        self.cond_check_number_images = cond_check_number_images
        self.number_images_min_cutoff = number_images_min_cutoff
        self.number_images_max_cutoff = number_images_max_cutoff

        self.cond_check_number_words_doc_level = cond_check_number_words_doc_level
        self.strip_characters = strip_characters
        self.number_words_doc_level_min_cutoff = number_words_doc_level_min_cutoff
        self.number_words_doc_level_max_cutoff = number_words_doc_level_max_cutoff

        self.cond_check_character_repetition_ratio_doc_level = cond_check_character_repetition_ratio_doc_level
        self.character_repetition_length_doc_level = character_repetition_length_doc_level
        self.character_repetition_doc_level_max_cutoff = character_repetition_doc_level_max_cutoff

        self.cond_check_word_repetition_ratio_doc_level = cond_check_word_repetition_ratio_doc_level
        self.word_repetition_length_doc_level = word_repetition_length_doc_level
        self.word_repetition_doc_level_max_cutoff = word_repetition_doc_level_max_cutoff

        self.cond_check_special_character_ratio_doc_level = cond_check_special_character_ratio_doc_level
        self.special_character_ratio_doc_level_max_cutoff = special_character_ratio_doc_level_max_cutoff

        self.cond_check_stopword_ratio_doc_level = cond_check_stopword_ratio_doc_level
        self.stopwords = stopwords
        self.stopword_ratio_doc_level_min_cutoff = stopword_ratio_doc_level_min_cutoff

        self.cond_check_flagged_word_ratio_doc_level = cond_check_flagged_word_ratio_doc_level
        self.flagged_words = flagged_words
        self.flagged_word_ratio_doc_level_max_cutoff = flagged_word_ratio_doc_level_max_cutoff

        self.cond_check_punctuation_ratio_doc_level = cond_check_punctuation_ratio_doc_level
        self.punctuation = punctuation
        self.punctuation_ratio_doc_level_min_cutoff = punctuation_ratio_doc_level_min_cutoff

        self.cond_check_common_word_ratio_doc_level = cond_check_common_word_ratio_doc_level
        self.path_common_words = path_common_words
        with open(path_common_words) as f:
            self.common_words = json.load(f)
        self.common_word_ratio_doc_level_min_cutoff = common_word_ratio_doc_level_min_cutoff

        self.cond_check_lang_id_doc_level = cond_check_lang_id_doc_level
        self.path_lang_id_model = path_lang_id_model
        if cond_check_lang_id_doc_level:
            self.lang_id_model = fasttext.load_model(path_lang_id_model)
        self.lang_id_doc_level_min_cutoff = lang_id_doc_level_min_cutoff

        self.cond_check_perplexity_score_doc_level = cond_check_perplexity_score_doc_level
        self.non_printing_characters_re = non_printing_characters_re
        self.digits_re = digits_re
        self.unicode_punctuation = unicode_punctuation
        self.path_sentencepiece_model = path_sentencepiece_model
        if cond_check_perplexity_score_doc_level:
            self.sentencepiece_model = sentencepiece.SentencePieceProcessor()
            self.sentencepiece_model.load(path_sentencepiece_model)
        self.path_kenlm_model = path_kenlm_model
        if cond_check_perplexity_score_doc_level:
            self.kenlm_model = kenlm.Model(path_kenlm_model)
        self.perplexity_score_doc_level_max_cutoff = perplexity_score_doc_level_max_cutoff

    def __call__(self, web_document):
        texts = web_document["texts"]
        images = web_document["images"]

        full_text = "\n\n".join([text for text in texts if text])
        all_images = [image for image in images if image]

        if self.cond_check_number_images:
            if not FilteringFunctions.check_number_images(
                number_images=len(all_images),
                number_images_min_cutoff=self.number_images_min_cutoff,
                number_images_max_cutoff=self.number_images_max_cutoff,
            ):
                return False

        if self.cond_check_number_words_doc_level:
            if not FilteringFunctions.check_number_words(
                text=full_text,
                strip_characters=self.strip_characters,
                number_words_min_cutoff=self.number_words_doc_level_min_cutoff,
                number_words_max_cutoff=self.number_words_doc_level_max_cutoff,
            ):
                return False

        if self.cond_check_character_repetition_ratio_doc_level:
            if not FilteringFunctions.check_character_repetition_ratio(
                text=full_text,
                character_repetition_length=self.character_repetition_length_doc_level,
                character_repetition_max_cutoff=self.character_repetition_doc_level_max_cutoff,
            ):
                return False

        if self.cond_check_word_repetition_ratio_doc_level:
            if not FilteringFunctions.check_word_repetition_ratio(
                text=full_text,
                strip_characters=self.strip_characters,
                word_repetition_length=self.word_repetition_length_doc_level,
                word_repetition_max_cutoff=self.word_repetition_doc_level_max_cutoff,
            ):
                return False

        if self.cond_check_special_character_ratio_doc_level:
            if not FilteringFunctions.check_special_character_ratio(
                text=full_text,
                special_characters=self.strip_characters,
                special_character_ratio_max_cutoff=self.special_character_ratio_doc_level_max_cutoff,
            ):
                return False

        if self.cond_check_stopword_ratio_doc_level:
            if not FilteringFunctions.check_stopword_ratio(
                text=full_text,
                strip_characters=self.strip_characters,
                stopwords=self.stopwords,
                stopword_ratio_min_cutoff=self.stopword_ratio_doc_level_min_cutoff,
            ):
                return False

        if self.cond_check_flagged_word_ratio_doc_level:
            if not FilteringFunctions.check_flagged_word_ratio(
                text=full_text,
                strip_characters=self.strip_characters,
                flagged_words=self.flagged_words,
                flagged_word_ratio_max_cutoff=self.flagged_word_ratio_doc_level_max_cutoff,
            ):
                return False

        if self.cond_check_punctuation_ratio_doc_level:
            if not FilteringFunctions.check_punctuation_ratio(
                text=full_text,
                punctuation=self.punctuation,
                punctuation_ratio_min_cutoff=self.punctuation_ratio_doc_level_min_cutoff,
            ):
                return False

        if self.cond_check_common_word_ratio_doc_level:
            if not FilteringFunctions.check_common_word_ratio(
                text=full_text,
                strip_characters=self.strip_characters,
                common_words=self.common_words,
                common_word_ratio_min_cutoff=self.common_word_ratio_doc_level_min_cutoff,
            ):
                return False

        if self.cond_check_lang_id_doc_level:
            if not FilteringFunctions.check_lang_id(
                text=full_text,
                lang_id_model=self.lang_id_model,
                target_lang_id="en",
                lang_id_min_cutoff=self.lang_id_doc_level_min_cutoff,
            ):
                return False

        if self.cond_check_perplexity_score_doc_level:
            if not FilteringFunctions.check_perplexity_score(
                text=full_text,
                non_printing_characters_re=self.non_printing_characters_re,
                digits_re=self.digits_re,
                unicode_punctuation=self.unicode_punctuation,
                sentencepiece_model=self.sentencepiece_model,
                kenlm_model=self.kenlm_model,
                perplexity_score_max_cutoff=self.perplexity_score_doc_level_max_cutoff,
            ):
                return False

        return True

    # Needed to make multiprocessing work
    def __reduce__(self):
        return (
            self.__class__,
            (
                self.cond_check_number_images,
                self.number_images_min_cutoff,
                self.number_images_max_cutoff,
                self.cond_check_number_words_doc_level,
                self.strip_characters,
                self.number_words_doc_level_min_cutoff,
                self.number_words_doc_level_max_cutoff,
                self.cond_check_character_repetition_ratio_doc_level,
                self.character_repetition_length_doc_level,
                self.character_repetition_doc_level_max_cutoff,
                self.cond_check_word_repetition_ratio_doc_level,
                self.word_repetition_length_doc_level,
                self.word_repetition_doc_level_max_cutoff,
                self.cond_check_special_character_ratio_doc_level,
                self.special_character_ratio_doc_level_max_cutoff,
                self.cond_check_stopword_ratio_doc_level,
                self.stopwords,
                self.stopword_ratio_doc_level_min_cutoff,
                self.cond_check_flagged_word_ratio_doc_level,
                self.flagged_words,
                self.flagged_word_ratio_doc_level_max_cutoff,
                self.cond_check_punctuation_ratio_doc_level,
                self.punctuation,
                self.punctuation_ratio_doc_level_min_cutoff,
                self.cond_check_common_word_ratio_doc_level,
                self.path_common_words,
                self.common_word_ratio_doc_level_min_cutoff,
                self.cond_check_lang_id_doc_level,
                self.path_lang_id_model,
                self.lang_id_doc_level_min_cutoff,
                self.cond_check_perplexity_score_doc_level,
                self.non_printing_characters_re,
                self.digits_re,
                self.unicode_punctuation,
                self.path_sentencepiece_model,
                self.path_kenlm_model,
                self.perplexity_score_doc_level_max_cutoff,
            ),
        )