app.py

import gradio as gr
from huggingface_hub import from_pretrained_fastai

from src.start_up.start_up_rbs import create_weak_rbs, create_strong_rbs
from src.start_up.start_up_gibberish import create_gibberish_detector

# Start up modules
#   Initiate model
learner = from_pretrained_fastai("nosdigitalmedia/dutch-youth-comment-classifier")
verdict_map = {
    0: 'Inappropriate',
    1: 'Allowed'
}

#   Initiate rule based systems (RBS)
weak_rbs = create_weak_rbs()
strong_rbs = create_strong_rbs()

#   Initiate gibberish detector
gibberish_detector = create_gibberish_detector()


# Define function to judge comment
def predict(input_text):

    # Get model predictions
    predictions = learner.predict(input_text)

    # Pass through weak RBS
    allows_weak, reasons_weak, highlights_weak = weak_rbs.allows(input_text)

    # Pass through strong RBS
    allows_strong, reasons_strong, highlights_strong = strong_rbs.allows(input_text)

    # Pass through gibberish detector
    is_gibberish_free = gibberish_detector.predict(input_text)

    # Construct json respons
    to_return = {
        'model': {
            'allowed': bool(int(predictions[0])),
            'verdict': verdict_map[int(predictions[0])],
            'highlights': [],
            'reasons': ['Machine learning model does not approve' for el in [predictions[0]] if not int(el)]
        },
        'gibberish': {
            'allowed': bool(is_gibberish_free),
            'verdict': verdict_map[int(is_gibberish_free)],
            'highlights': [],
            'reasons': ['Comment text contains gibberish' for el in [is_gibberish_free] if not el]
        },
        'weak_rbs': {
            'allowed': allows_weak,
            'verdict': verdict_map[allows_weak],
            'highlights': highlights_weak,
            'reasons': reasons_weak
        },
        'strong_rbs': {
            'allowed': allows_strong,
            'verdict': verdict_map[allows_strong],
            'highlights': highlights_strong,
            'reasons': reasons_strong
        },
    }

    return to_return


# Set up app
iface = gr.Interface(
    fn=predict,
    inputs="text",
    outputs="json",
    title="Dutch Youth Comment Classifier",
    description="<p>Input a comment and our model and rule based system will tell you if the comment is appropriate for "
                "youth. Four systems are used to judge the comment: "
                "<ol> <li> <code>model</code> refers to the Deep Learning model trained on moderator decisions as provided <a href=\"https://huggingface.co/nosdigitalmedia/dutch-youth-comment-classifier\">here</a>.  </li>"
                "<li>  <code>gibberish</code> refers to a markov-chain model predicting whether a comment is actual text or a random sequence of characters </li>"
                "<li> <code>weak_rbs</code> refers to a rule-based system that flags ambiguous comments that should be checked by a human. </li>"
                "<li> <code>strong_rbs</code> refers to a rule-based system that flags inappropriate comments that could be removed directly.</ol></p>",
    examples=[
        "Ik zag gisteren een puppy",
        "hljksdfghslkdjfghlsdkfghjslkdjhfg",
        "mijn email adress is abc@hotmail.com",
        "<H1>Dit is mijn website nu</H1>",
        "mijn nicht is een zuster",
        "jij bent een klootzak"
    ]
)
iface.launch(share=False)