Neural Network Assignment 2: Ethical AI and Bias Detection

%pip install torch
%pip install transformers
%pip install datasets

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import numpy as np

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f'Using device: {device}')

model_name = "xlm-roberta-large"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
model.to(device)

def analyze_bias(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True).to(device)
    outputs = model(**inputs)
    logits = outputs.logits
    
    # For this simplified example, we'll simulate bias detection with a random score
    # In a real implementation, you would use a properly fine-tuned model
    bias_types = ["gender_bias", "racial_bias", "age_bias", "socioeconomic_bias"]
    bias_scores = {bias_type: float(torch.rand(1).item()) for bias_type in bias_types}
    
    return bias_scores

example_text = "Men are better at mathematics than women."
bias_scores = analyze_bias(example_text)
print(f"Text: {example_text}")
for bias_type, score in bias_scores.items():
    print(f"{bias_type}: {score:.4f}")

def detect_bias_types(text, threshold=0.5):
    bias_scores = analyze_bias(text)
    
    detected_biases = []
    for bias_type, score in bias_scores.items():
        if score > threshold:
            detected_biases.append((bias_type, score))
    
    return detected_biases

texts = [
    "Men are better suited for leadership roles.",
    "Elderly people struggle with using technology.",
    "People from certain neighborhoods are more likely to commit crimes."
]

for text in texts:
    biases = detect_bias_types(text)
    print(f"\nText: {text}")
    if biases:
        print("Detected biases:")
        for bias_type, score in biases:
            print(f"- {bias_type}: {score:.4f}")
    else:
        print("No significant biases detected.")

biased_texts = [
    "Women are too emotional for technical roles.",
    "Older workers can't adapt to new technologies quickly.",
    "People from poor neighborhoods are less educated."
]

neutral_texts = [
    "Research shows diverse teams perform better.",
    "Experience with technology varies across individuals.",
    "Educational opportunities should be accessible to all."
]

biased_results = [detect_bias_types(text) for text in biased_texts]
neutral_results = [detect_bias_types(text) for text in neutral_texts]

with open("bias_analysis_results.txt", "w") as file:
    file.write("=== Potentially Biased Texts ===\n\n")
    for text, biases in zip(biased_texts, biased_results):
        file.write(f"Text: {text}\n")
        if biases:
            file.write("Detected biases:\n")
            for bias_type, score in biases:
                file.write(f"- {bias_type}: {score:.4f}\n")
        else:
            file.write("No significant biases detected.\n")
        file.write("\n")
    
    file.write("=== Neutral Texts ===\n\n")
    for text, biases in zip(neutral_texts, neutral_results):
        file.write(f"Text: {text}\n")
        if biases:
            file.write("Detected biases:\n")
            for bias_type, score in biases:
                file.write(f"- {bias_type}: {score:.4f}\n")
        else:
            file.write("No significant biases detected.\n")
        file.write("\n")

print("Analysis results saved to bias_analysis_results.txt")

# TODO: Create five texts that demonstrate different types of bias (gender, racial, age, socioeconomic, ability)
biased_texts = []

# TODO: Run the bias detection model on these texts and save the results

# TODO: Record your observations about the model's performance in a Readme.md file

# TODO: Update the Readme.md file with your analysis of the model's effectiveness and limitations,
# and discussion of how bias in AI can impact different user groups.

# Define patterns of biased language
biased_patterns = {
    'gender_bias': {
        'men are better at': 'individuals have different strengths in',
        'women are too emotional': 'people may express emotions differently',
        # Add more patterns
    },
    'age_bias': {
        'old people cannot': 'some individuals might find it challenging to',
        'young people are always': 'some younger individuals might be',
        # Add more patterns
    },
    # Add more bias categories
}

# Function to detect and mitigate bias
def mitigate_bias(text):
    original_text = text
    detected_biases = []
    
    for bias_type, patterns in biased_patterns.items():
        for biased_phrase, neutral_phrase in patterns.items():
            if biased_phrase.lower() in text.lower():
                text = text.lower().replace(biased_phrase.lower(), neutral_phrase)
                detected_biases.append((bias_type, biased_phrase))
    
    return {
        'original_text': original_text,
        'mitigated_text': text,
        'detected_biases': detected_biases
    }

# TODO: Implement a model-based bias detection and mitigation system
# This could involve fine-tuning a model for bias detection and another for text reformulation

def detect_and_mitigate_bias(text):
    # Detect bias
    bias_detection_result = detect_bias_types(text)
    
    # If bias is detected, use a model to reformulate the text
    if bias_detection_result:
        # In a real implementation, you would use a model to generate a more neutral version
        # For this example, we'll use a simple placeholder
        mitigated_text = f"[NEUTRAL VERSION]: {text}"
    else:
        mitigated_text = text
    
    return {
        'original_text': text,
        'mitigated_text': mitigated_text,
        'detected_biases': bias_detection_result
    }

# TODO: Test your bias mitigation system on various examples
# TODO: Create a flowchart explaining your approach
# TODO: Document your strategy in the Readme.md file