challenging-america-word-ga.../prep_data.py
Jakub Henyk 4807f6e442 452627
2024-04-22 21:57:49 +02:00

150 lines
3.8 KiB
Python

import csv
import re
from gensim.models import Word2Vec
import gensim.downloader as api
import numpy as np
from spellchecker import SpellChecker
import pandas as pd
folder = 'test-A'
filename = f"{folder}/in_1.csv"
data = []
data = pd.read_csv(f'{folder}/in.tsv',delimiter='\t', header=None, encoding='utf-8', quoting=csv.QUOTE_NONE, engine='python').values.tolist()
data_a = []
data_b = []
data_pair = []
for i in range(len(data)):
data_a.append(data[i][6])
try:
data_b.append(data[i][7])
except:
data_b.append('')
for i in range(len(data)):
data_pair.append([data_a[i], data_b[i]])
data_tabs = []
for x, y in data_pair:
cleaned_text_a = x.replace('\\t', '\t').replace('\\n', '\n').strip("[]")
cleaned_text_b = y.replace('\\t', '\t').replace('\\n', '\n').strip("[]")
data_tabs.append([cleaned_text_a, cleaned_text_b])
data_removed = []
for x, y in data_tabs:
text = re.sub(r'(?<!-)\n', ' ', x)
text = re.sub(r'[\n-]', '', text)
text = re.sub(r'[^a-zA-Z0-9\s]', '', text)
text = re.sub(r'\s+', ' ', text)
text_2 = re.sub(r'(?<!-)\n', ' ', y)
text_2 = re.sub(r'[\n-]', '', text_2)
text_2 = re.sub(r'[^a-zA-Z0-9\s]', '', text_2)
text_2 = re.sub(r'\s+', ' ', text_2)
data_removed.append([text, text_2])
model = api.load("word2vec-google-news-300")
def is_close_to_actual(word, threshold=0.5):
if word in model:
similarities = model.similar_by_word(word)
return any(similarity > threshold for _, similarity in similarities)
else:
return False
def remove_words(text, words_to_destroy):
pattern = r'\b(?:{})\b'.format('|'.join(words_to_destroy))
cleaned_text = re.sub(pattern, '', text, flags=re.IGNORECASE)
return cleaned_text
spell = SpellChecker()
data_cleared = []
i = 0
for x, y in data_removed:
words = x.split()
words_2 = y.split()
misspelled = spell.unknown(words + words_2)
text = remove_words(x, list(misspelled))
text_2 = remove_words(y, list(misspelled))
data_cleared.append([text, text_2])
if i % 20000 == 0:
print(f'{i/430000*100}%')
i += 1
data_cleared_2 = []
for x, y in data_cleared:
text = re.sub(r'(?<!-)\n', ' ', x)
text = re.sub(r'[\n-]', '', text)
text = re.sub(r'[^a-zA-Z0-9\s]', '', text)
text = re.sub(r'\s+', ' ', text)
text_2 = re.sub(r'(?<!-)\n', ' ', y)
text_2 = re.sub(r'[\n-]', '', text_2)
text_2 = re.sub(r'[^a-zA-Z0-9\s]', '', text_2)
text_2 = re.sub(r'\s+', ' ', text_2)
data_cleared_2.append([text, text_2])
with open(filename, 'w', newline='') as csvfile:
writer = csv.writer(csvfile)
writer.writerows(data_cleared_2)
"""import wordninja
from spellchecker import SpellChecker
spell = SpellChecker()
concatenated_misspelled = []
for x, y in data_removed:
words = x.split()
words_2 = y.split()
misspelled = spell.unknown(words + words_2)
concatenated_misspelled.append(list(misspelled))
data_corrected = []
i = 0
for x, y in data_removed:
text = x
text_2 = y
for word in flattened_concatenated_misspelled:
if is_close_to_actual(word, model):
corrected_word = spell.correction(word)
if corrected_word != None:
text = text.replace(word, corrected_word)
text_2 = text_2.replace(word, corrected_word)
else:
if len(word) > 6:
tokens = wordninja.split(word)
my_string = ' '.join(tokens)
text = text.replace(word, my_string)
text_2 = text_2.replace(word, my_string)
else:
text = text.replace(word, '')
text_2 = text_2.replace(word, '')
if i % 20000 == 0:
print(f'{i/430000*100}%')
i += 1
data_corrected.append([text, text_2])"""