Skip to content
Snippets Groups Projects
SimTexter.py 9.73 KiB
Newer Older
Frederik Arnold's avatar
Frederik Arnold committed
from sim.Match import Match
from sim.MatchSegment import MatchSegment
from sim.Text import Text
import re
from sim.Token import Token
from rapidfuzz import fuzz, process
from datasketch import MinHash, MinHashLSH

class SimTexter:

    def __init__(self, min_match_length):
        self.min_match_length = min_match_length
        self.cache = {}

    def compare(self, input_texts):

        (texts, tokens) = self.read_input(input_texts)

        mts_tags = {}
        forward_references = {}
        existing_tags = []
        lsh = MinHashLSH(threshold=0.80, num_perm=128)

        for i in range(0, len(texts)):
            (mts_tags, forward_references, existing_tags, lsh) = self.make_forward_references(i, texts[i], tokens, mts_tags, existing_tags, forward_references, lsh)

        similarities = self.get_similarities(tokens, texts, 0, 1, forward_references)

        return similarities

        # self.print_similarities(similarities, input_texts)

    def read_input(self, input_texts):

        texts = []
        tokens = []

        for input_text in input_texts:
            nr_of_characters = len(input_text)
            nr_of_words = len(input_text.split())
            file_name = 'dummy'
            tk_start_pos = len(tokens)

            tokens.extend(self.tokenize_text(input_text))
            tk_end_pos = len(tokens)
            text = Text('Text', nr_of_characters, nr_of_words, file_name, tk_start_pos, tk_end_pos)
            texts.append(text)

        return texts, tokens

    def tokenize_text(self, input_text):
        cleaned_text = self.clean_text(input_text)

        tokens = []

        for match in re.finditer("[^\\s]+", cleaned_text):
            token = self.clean_word(match.group())

            if len(token) > 0:
                text_begin_pos = match.start()
                text_end_pos = match.end()

                tokens.append(Token(token, text_begin_pos, text_end_pos))

        return tokens

    def clean_text(self, input_text):
        # TODO: optional machen

        input_text = re.sub("[.?!,;:/()'+\\-\\[\\]‚‘…]", " ", input_text)
        input_text = re.sub("[0-9]", " ", input_text)

        return input_text.lower()

    def clean_word(self, input_word):
        # TODO: Umlaute ersetzen, optional machen
        return input_word

    def make_forward_references(self, text_index, text, tokens, mts_tags, existing_tags, forward_references, lsh):
        text_begin_pos = text.tk_start_pos
        text_end_pos = text.tk_end_pos

        for i in range(text_begin_pos, text_end_pos - self.min_match_length):
            tag = ''

            for token in tokens[i: i + self.min_match_length]:
                tag = tag + token.text

            # TODO: geht das fuzzy??

            # if tag in mts_tags:
            #    forward_references[mts_tags[tag]] = i

            # mts_tags[tag] = i

            # if text_index == 0:
            #     existing_tags.append(tag)
            # else:
            #     best_existing_tag = process.extractOne(tag, existing_tags, scorer=fuzz.ratio, score_cutoff=80)
            #
            #     if best_existing_tag:
            #         forward_references[mts_tags[best_existing_tag[0]]] = i

            if text_index == 0:
                my_set = set(tag)
                min_hash = MinHash(num_perm=128)

                for d in my_set:
                    min_hash.update(d.encode('utf8'))

                lsh.insert(tag, min_hash, False)
            else:
                my_set = set(tag)
                min_hash = MinHash(num_perm=128)

                for d in my_set:
                    min_hash.update(d.encode('utf8'))

                result = lsh.query(min_hash)

                if result and len(result) > 0:
                    closest_match = self.get_closest_match(result, tag)
                    if closest_match:
                        forward_references[mts_tags[closest_match]] = i

            mts_tags[tag] = i

        return mts_tags, forward_references, existing_tags, lsh

    def get_similarities(self, tokens, texts, source_text_index, target_text_index, forward_references):
        source_token_start_pos = texts[source_text_index].tk_start_pos
        source_token_end_pos = texts[source_text_index].tk_end_pos

        similarities = []

        while source_token_start_pos + self.min_match_length <= source_token_end_pos:
            best_match = self.get_best_match(tokens, texts, source_text_index, target_text_index, source_token_start_pos,
                                        forward_references)

            if best_match and best_match.length > 0:
                source_character_start_pos = tokens[best_match.source_token_start_pos].start_pos
                source_character_end_pos = tokens[best_match.source_token_start_pos + best_match.length - 1].end_pos
                target_character_start_pos = tokens[best_match.target_token_start_pos].start_pos
                target_character_end_pos = tokens[best_match.target_token_start_pos + best_match.length - 1].end_pos

                similarities.append((MatchSegment(best_match.source_text_index, best_match.source_token_start_pos,
                                                  best_match.length, source_character_start_pos, source_character_end_pos),
                                     MatchSegment(best_match.target_text_index, best_match.target_token_start_pos,
                                                  best_match.length, target_character_start_pos, target_character_end_pos)))

                source_token_start_pos = source_token_start_pos + best_match.length
            else:
                source_token_start_pos = source_token_start_pos + 1

        return similarities

    def get_best_match(self, tokens, texts, source_text_index, target_text_index, source_token_start_pos, forward_references):
        best_match_length = 0
        token_pos = source_token_start_pos

        source_token_pos = 0
        target_token_pos = 0

        best_match_token_pos = 0

        best_match = None

        while 0 < token_pos < len(tokens):

            if token_pos < texts[target_text_index].tk_start_pos:
                if token_pos in forward_references:
                    token_pos = forward_references[token_pos]
                else:
                    token_pos = -1
                continue

            min_match_length = self.min_match_length

            if best_match_length > 0:
                min_match_length = best_match_length + 1

            source_token_pos = source_token_start_pos + min_match_length - 1
            target_token_pos = token_pos + min_match_length - 1

            if source_token_pos < texts[source_text_index].tk_end_pos and texts[
                target_text_index].tk_end_pos > target_token_pos >= source_token_pos + min_match_length:

                cnt = min_match_length

                while cnt > 0 and self.fuzzy_match(tokens[source_token_pos].text, tokens[target_token_pos].text) > 80:
                    source_token_pos = source_token_pos - 1
                    target_token_pos = target_token_pos - 1
                    cnt = cnt - 1

                if cnt > 0:
                    if token_pos in forward_references:
                        token_pos = forward_references[token_pos]
                    else:
                        token_pos = -1
                    continue
            else:
                if token_pos in forward_references:
                    token_pos = forward_references[token_pos]
                else:
                    token_pos = -1
                continue

            new_match_length = min_match_length
            source_token_pos = source_token_start_pos + min_match_length
            target_token_pos = token_pos + min_match_length

            while source_token_pos < texts[source_text_index].tk_end_pos and texts[
                target_text_index].tk_end_pos > target_token_pos > source_token_pos + \
                    new_match_length and self.fuzzy_match(tokens[source_token_pos].text,
                                                          tokens[target_token_pos].text) > 80:

                source_token_pos = source_token_pos + 1
                target_token_pos = target_token_pos + 1
                new_match_length = new_match_length + 1

            if new_match_length >= self.min_match_length and new_match_length > best_match_length:
                best_match_length = new_match_length
                best_match_token_pos = token_pos
                best_match = Match(source_text_index, source_token_start_pos, target_text_index, best_match_token_pos,
                                   best_match_length)

            if token_pos in forward_references:
                token_pos = forward_references[token_pos]
            else:
                token_pos = -1

        return best_match

    def fuzzy_match(self, input1, input2):

        # if input1 + input2 in self.cache:
        #    return self.cache[input1 + input2]

        # if abs(len(input1) - len(input2)) >= 3:
        #    self.cache[input1 + input2] = 0
        #    return 0

        ratio = fuzz.ratio(input1, input2)
        # self.cache[input1 + input2] = ratio
        return ratio

    def get_closest_match(self, candidates, word):
        if word in candidates:
            return word

        best_existing_tag = process.extractOne(word, candidates, scorer=fuzz.ratio, score_cutoff=80)

        if best_existing_tag:
            return best_existing_tag[0]

        return None

    def print_similarities(self, similarities, input_texts):
        for similarity_tuple in similarities:
            similarity_literature = similarity_tuple[0]
            similarity_scientific = similarity_tuple[1]

            print('{0}, {1}'.format(similarity_literature, similarity_scientific))
            print(input_texts[0][similarity_literature.character_start_pos:similarity_literature.character_end_pos])