app.py

import gradio as gr
import requests
import emoji
import re
import json
from thefuzz import process, fuzz
import numpy as np
import re


API_URL = "https://api-inference.huggingface.co/models/Dabid/test2"
headers = {"Authorization": "Bearer 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"}

def read_text(filename, filetype='txt'):
    words = []

    if filetype == 'txt':
        with open(filename + '.txt') as file:
            words = [line.rstrip() for line in file]
            words = list(set(words))
    elif filetype == 'json':
        with open(filename + '.json') as json_file:
            words = json.load(json_file)
    
    return words


contractions = read_text('contractions', 'json')
lookup_words = read_text('lookup_words')
obj_pronouns = read_text('obj_pronouns')
profanities = read_text('profanities', 'json')


def fuzzy_lookup(tweet):
    lookup_profanity = np.concatenate([np.hstack(list(profanities.values())), list(profanities.keys())])
    matches = dict()

    # Loop each word in tweet
    for word in tweet.split():
         # Only get digits and letters then lowercase
        word = re.sub("[^a-zA-Z0-9@]", "", word).lower()
        scores = []
        matched_words = []
        # If word > 4 chars
        if len(word) >= 4:
            # Get fuzzy ratio
            for lookup_word in lookup_words:
                score = fuzz.ratio(word, lookup_word)
                if score >= 70:
                    scores.append(score)
                    matched_words.append(lookup_word)
            if len(scores) > 0:
                max_score_index = np.argmax(scores)
                if matched_words[max_score_index] in lookup_profanity:
                    matches[word] = matched_words[max_score_index]
                

    for word, matched_profanity in matches.items():
        word_split = word.split(matched_profanity[-2:])
        for pronoun in obj_pronouns:
            if len(word_split) > 1:
                if pronoun == word_split[-1]:
                    matches[word] = matched_profanity + ' ' + pronoun
                    break

    # Replace each profanities by fuzzy lookup result
    for word, matched_profanity in matches.items():
        tweet = tweet.replace(word, matched_profanity)

    for profanity, prof_varations in profanities.items():
        if len(prof_varations) > 0:
            for prof_variant in prof_varations:
                tweet = tweet.replace(prof_variant, profanity)

    return tweet, matches


def preprocess(tweet):
    laugh_texts = ['hahaha', 'wahaha', 'hahaa', 'ahha', 'haaha', 'hahah', 'ahah', 'hha']
    symbols = ['@', '#']

    # Lowercase
    tweet = tweet.lower()

    # Remove emojis
    tweet = emoji.replace_emoji(tweet, replace='')

    # Replace elongated words 'grabeee' -> 'grabe' (not applicable on 2 corresponding letter)
    tweet = re.sub(r'(.)\1{2,}', r'\1', tweet)

    # Split sentence into list of words
    row_split = tweet.split()

    for index, word in enumerate(row_split):

        # Remove words with symbols (e.g. @username, #hashtags)
        if any(x in word for x in symbols):
            row_split[index] = ''

        # Remove links
        if 'http' in word:
            row_split[index] = ''

        # Unify laugh texts format to 'haha'
        if any(x in word for x in laugh_texts):
            row_split[index] = 'haha'

    # Combine list of words back to sentence
    combined_text = ' '.join(filter(None, row_split))

    # Check if output contains single word then return null
    if len(combined_text.split()) == 1:
        return combined_text

    # Filter needed characters
    combined_text = re.sub(r"[^A-Za-z ]+", '', combined_text)

    # Expand Contractions
    for i in contractions.items():
        combined_text = re.sub(rf"\b{i[0]}\b", i[1], combined_text)

    return combined_text


def query(payload):
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()


def predict(tweet):
    
    fuzzy_text, matches = fuzzy_lookup(tweet)
    processed_text = preprocess(fuzzy_text)
    output = query(processed_text)


    if 'error' in output:
        return output['error'], 'Error occured. Try again later.', {}
    elif len(matches) == 0:
        return 'No Profanity Found.', '', {}
    else:
        output = [tuple(i.values()) for i in output[0]]
        output = dict((x, y) for x, y in output)
        predicted_label = list(output.keys())[0]

        if predicted_label == 'Abusive':
            # Censor
            for base_word, _ in matches.items():
                mask = '*' * len(base_word)
                compiled = re.compile(re.escape(base_word), re.IGNORECASE)
                tweet = compiled.sub(mask, tweet)
                # tweet = tweet.replace(base_word, re.sub("[a-zA-Z0-9@]", "*", base_word))
            return output, tweet, json.dumps(matches)
        else:
            return output, tweet, json.dumps(matches)

# output, tweet, matches = predict('ul0L Sama ng ugali mo pre Tangina uL0l!!!')
# print(output, '\n', tweet, '\n', matches)

hf_writer = gr.HuggingFaceDatasetSaver('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', 'tagalog-profanity-feedbacks')


demo = gr.Interface(
    fn=predict,

    inputs=[gr.components.Textbox(lines=5, placeholder='Enter your input here', label='INPUT')],

    outputs=[gr.components.Label(num_top_classes=2, label="PREDICTION"),
             gr.components.Text(label='OUTPUT'),
             gr.components.JSON(label='DETECTED PROFANITIES')],

    examples=['Tangina mo naman sobrang yabang mo gago!!😠😤 @davidrafael',
              'Napakainit ngayong araw pakshet namaaan!!',
              'Napakabagal naman ng wifi tangina #PLDC #HelloDITO',
              'Bobo ka ba? napakadali lang nyan eh... 🤡',
              'Uy gago laptrip yung nangyare samen kanina HAHAHA😂😂'],

    allow_flagging="manual",
    flagging_callback=hf_writer,
    flagging_options=['Good bot', 'Bad bot']
)

demo.launch()