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
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):
    payload = {"inputs": text}
    response = requests.post(API_URL, headers=headers, json=payload)
    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')
similar_words = read_text('similar_words')
addon_words = read_text('addon_words')
profanities = read_text('profanities', 'json')
lookup_profanity = np.concatenate([np.hstack(list(profanities.values())), list(profanities.keys())])
lookup_words = list(set(similar_words).union(set(lookup_profanity.tolist())))
eng_words = list(set(words.words()) - set(lookup_profanity))
punctuations = re.compile(r'^[^\w#@]+|[^\w#@]+$')

# TODO check eng words that are tagalog profanities

def fuzzy_lookup(tweet):

    matched_profanity = dict()

    # tweet = punctuations.sub('', tweet).lower()

    for word in tweet.split():

        word = punctuations.sub('', word).lower()
        base_word =  word
        word = re.sub(r'(.)\1{2,}', r'\1', word)
        
        if word in eng_words:
            continue
        
        for addon in addon_words:
            if word.startswith(addon):
                word = word[len(addon):]
            if word.endswith(addon):
                word = word[:-len(addon)]

        if word.startswith("@") or word.startswith("#"):
            word = word[1:]

        scores = []
        matched_words = []

        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:
                    for base_profanity, profanity_variations in profanities.items():
                        if matched_words[max_score_index] == base_profanity:
                            matched_profanity[base_word] = base_profanity
                            break
                        if matched_words[max_score_index] in profanity_variations:
                            matched_profanity[base_word] = base_profanity
                            break
                        
    return matched_profanity


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

    # Replace profanities
    for base_word, matched_word in profanities.items():
        tweet = tweet.replace(base_word, matched_word)

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

    row_split = tweet.split()

    for index, word in enumerate(row_split):

        # Seperate pronouns
        for addon in addon_words:
            if word.endswith(addon):
                row_split[index] = word[:-len(addon)] + " " + addon

        # 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))

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

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

    return preprocessed_tweet



def predict(tweet):
    
    profanities = fuzzy_lookup(tweet)

    if len(profanities) > 0:

        preprocessed_tweet = preprocess(tweet, profanities)

        prediction = query(preprocessed_tweet)

        if type(prediction) == dict: 
            print(prediction)
            error_message = prediction['error']
            return error_message, {}
        prediction = prediction[0][0]["label"]
        
        print("\nTWEET:", tweet)
        print("DETECTED PROFANITY:", list(profanities.keys()))
        print("LABEL:", prediction, "\n")

        return prediction, list(profanities.keys())
    
    return "No Profanity", {}


demo = gr.Interface(
    fn=predict,

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

    outputs=[gr.components.Text(label="PREDICTION"), gr.JSON(label="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="never",

    title="Tagalog Profanity Classifier"
)

demo.launch(debug=True)