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
from string import punctuation
import nltk
nltk.download('words')
from nltk.corpus import words


API_URL = "https://api-inference.huggingface.co/models/Dabid/abusive-tagalog-profanity-detection"
headers = {"Authorization": "Bearer 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"}

def query(text):
    text = {"inputs": text}
    response = requests.post(API_URL, headers=headers, json=text)
    return response.json()

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')
eng_words = set(words.words())

# TODO check eng words that are tagalog profanities

def fuzzy_lookup(tweet):

    matched_profanity = dict()
    lookup_profanity = np.concatenate([np.hstack(list(profanities.values())), list(profanities.keys())])

    for word in tweet.split():
        if word in list(set(eng_words) - set(lookup_profanity)):
            break
        scores = []
        matched_words = []
        word = word.strip(punctuation)
        processed_word = re.sub("[^a-zA-Z0-9@]", "", word)

        if len(processed_word) >= 4:
            # Get fuzzy ratio
            for lookup_word in lookup_words:
                score = fuzz.ratio(processed_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:
                    matched_profanity[word] = matched_words[max_score_index]

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

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


def preprocess(tweet):
    
    tweet = tweet.lower()
    tweet = emoji.replace_emoji(tweet, replace='')

    # Elongated words conversion
    tweet = re.sub(r'(.)\1{2,}', r'\1', tweet)

    row_split = tweet.split()

    for index, word in enumerate(row_split):

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

        # Unify laugh texts format to 'haha'
        laugh_texts = ['hahaha', 'wahaha', 'hahaa', 'ahha', 'haaha', 'hahah', 'ahah', 'hha']
        if any(x in word for x in laugh_texts):
            row_split[index] = 'haha'

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

    # Fuzzy Lookup
    preprocessed_tweet, matches = fuzzy_lookup(preprocessed_tweet)

    if len(preprocessed_tweet.split()) == 1:
        return preprocessed_tweet, matches

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

    return preprocessed_tweet, matches 



def predict(tweet):
    
    preprocessed_tweet, matched_profanity = preprocess(tweet)

    prediction = query(preprocessed_tweet)

    if type(prediction) == dict: 
        print(prediction)
        error_message = prediction['error']
        return error_message

    
    if bool(matched_profanity) == False:
        return "No Profanity"
        
    
    prediction = [tuple(i.values()) for i in prediction[0]]
    prediction = dict((x, y) for x, y in prediction)
    
    print("\nTWEET:", tweet)
    print("DETECTED PROFANITY:", matched_profanity)
    print("LABELS:", prediction, "\n")

    return prediction


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")],

    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😂😂'],
)

demo.launch(debug=True)