app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("BenjaminOcampo/peace_hatebert")
model = AutoModelForSequenceClassification.from_pretrained("BenjaminOcampo/peace_hatebert")

# Define more nuanced labels for the model output
nuanced_labels = {
    0: "Non-Hate Speech",
    1: "Explicit Hate",
    2: "Implicit Hate",
    3: "White Grievance"
}

# Microaggressions detection rules
microaggressions = {
    "You're so articulate": "This phrase can imply surprise that the individual can speak well, often used in a way that suggests it is unexpected for someone of their background.",
    "Where are you really from": "This question implies that the individual does not belong or is not truly part of the community.",
    "I don't see color": "This statement can negate the experiences and identities of people of different races.",
    "You're a credit to your race": "This phrase implies that most people of the individual’s race are not successful or commendable."
}

# A sample set of explanations and suggestions
bias_suggestions = {
    "Explicit Hate": {
        "suggestion": "Consider using language that promotes inclusivity and respect.",
        "explanation": "The text contains explicit hate speech, which is overtly harmful and discriminatory. It is important to foster communication that is inclusive and respectful of all individuals."
    },
    "Implicit Hate": {
        "suggestion": "Try rephrasing to avoid subtle bias and ensure clarity.",
        "explanation": "The text contains implicit hate speech, which can perpetuate stereotypes and bias in a less overt manner. Aim for language that is clear and free from insinuations."
    },
    "White Grievance": {
        "suggestion": "Reconsider any generalized claims about racial groups.",
        "explanation": "The text appears to express grievances linked to racial identity, which can contribute to divisive narratives. Strive for dialogue that acknowledges diversity and avoids stereotyping."
    },
    "Non-Hate Speech": {
        "suggestion": "No problematic content detected.",
        "explanation": "The text does not appear to contain hate speech or bias. It seems respectful and neutral."
    },
    "Microaggression": {
        "suggestion": "Be mindful of how certain phrases can be interpreted by others.",
        "explanation": "The text includes phrases that may be considered microaggressions, which can subtly perpetuate stereotypes or biases."
    }
}

def analyze_text(text):
    # Tokenize input text
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    
    # Get model predictions
    with torch.no_grad():
        outputs = model(**inputs)
    
    # Get prediction probabilities
    probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted_class = torch.argmax(probs, dim=-1).item()
    
    # Map the predicted class to the nuanced label
    label = nuanced_labels.get(predicted_class, "Unknown")
    
    # Check for microaggressions using the predefined rules
    for phrase, explanation in microaggressions.items():
        if phrase.lower() in text.lower():
            label = "Microaggression"
            suggestion = bias_suggestions[label]["suggestion"]
            explanation = bias_suggestions[label]["explanation"]
            return label, suggestion, explanation
    
    # Fetch suggestion and explanation based on label
    suggestion = bias_suggestions[label]["suggestion"]
    explanation = bias_suggestions[label]["explanation"]
    
    return label, suggestion, explanation

# Create the Gradio interface
interface = gr.Interface(
    fn=analyze_text,
    inputs=gr.Textbox(lines=5, placeholder="Enter text to analyze..."),
    outputs=[
        gr.Textbox(label="Classification"),
        gr.Textbox(label="Suggestion"),
        gr.Textbox(label="Explanation")
    ],
    title="Proofreading for Implicit Bias, Microagressions - inital model test",
    description="Analyze text for nuanced bias categories such as implicit hate, explicit hate, or white grievance, and detect microaggressions to provide suggestions for improvement - step 1: prompt testing. Credit to https://huggingface.co/BenjaminOcampo"
)


# Launch the interface
interface.launch()