app.py

import gradio as gr
import torch
from torch.nn.functional import softmax
import shap
import requests
from bs4 import BeautifulSoup
from sklearn.metrics.pairwise import cosine_similarity
from transformers import RobertaTokenizer,RobertaForSequenceClassification, pipeline
from IPython.core.display import HTML
model_dir = 'temp'
tokenizer = RobertaTokenizer.from_pretrained(model_dir)
model = RobertaForSequenceClassification.from_pretrained(model_dir)
#pipe = pipeline("text-classification", model="thugCodeNinja/robertatemp")
tokenizer1 = RobertaTokenizer.from_pretrained('roberta-base')
model1 = RobertaModel.from_pretrained('roberta-base')
pipe = pipeline("text-classification",model=model,tokenizer=tokenizer)
def process_text(input_text):
    if input_text:
        text = input_text
    inputs = tokenizer(text, return_tensors="pt")
    with torch.no_grad():
        logits = model(**inputs).logits
    probs = softmax(logits, dim=1)
    max_prob, predicted_class_id = torch.max(probs, dim=1)
    prob = str(round(max_prob.item() * 100, 2))
    label = model.config.id2label[predicted_class_id.item()]
    final_label='Human' if model.config.id2label[predicted_class_id.item()]=='LABEL_0' else 'Chat-GPT'
    processed_result = text
    def search(text):
        query = text
        api_key = 'AIzaSyClvkiiJTZrCJ8BLqUY9I38WYmbve8g-c8'
        search_engine_id = '53d064810efa44ce7'
        url = f'https://www.googleapis.com/customsearch/v1?key={api_key}&cx={search_engine_id}&q={query}'

        try:
            response = requests.get(url)
            data = response.json()
            return data
        except Exception as e:
            return {'error': str(e)}
    def get_article_text(url):
    try:
        response = requests.get(url)
        if response.status_code == 200:
            soup = BeautifulSoup(response.content, 'html.parser')
            # Extract text from the article content (you may need to adjust this based on the website's structure)
            article_text = ' '.join([p.get_text() for p in soup.find_all('p')])
            return article_text
    except Exception as e:
        print(f"An error occurred: {e}")
    return ''
    def find_plagiarism(text):
        search_results = search(text)
        if 'items' not in search_results:
            return []
        similar_articles = []
        for item in search_results['items']:
            link = item.get('link', '')
            article_text = get_article_text(link)
            if article_text:
            # Tokenize and encode the input text and the article text
                encoding1 = tokenizer(text, max_length=512, truncation=True, padding=True, return_tensors="pt")
                encoding2 = tokenizer(article_text, max_length=512, truncation=True, padding=True, return_tensors="pt")
            
            # Calculate embeddings using the model
                with torch.no_grad():
                    embedding1 = model(**encoding1).last_hidden_state.mean(dim=1)
                    embedding2 = model(**encoding2).last_hidden_state.mean(dim=1)
            
            # Calculate cosine similarity between the input text and the article text embeddings
            similarity = cosine_similarity(embedding1, embedding2)[0][0]
            similar_articles.append({'Link': link, 'Similarity': similarity})
            similar_articles = sorted(similar_articles, key=lambda x: x['Similarity'], reverse=True)
            threshold = 0.5  # Adjust the threshold as needed
            similar_articles = [article for article in similar_articles if article['Similarity'] > threshold]
            return similar_articles[:5]

    prediction = pipe([text])
    explainer = shap.Explainer(pipe)
    shap_values = explainer([text])
    shap_plot_html = HTML(shap.plots.text(shap_values, display=False)).data
    similar_articles = find_plagiarism(text)

    return processed_result, prob, final_label, shap_plot_html,similar_articles

text_input = gr.Textbox(label="Enter text")
outputs = [gr.Textbox(label="Processed text"), gr.Textbox(label="Probability"), gr.Textbox(label="Label"), gr.HTML(label="SHAP Plot"),gr.Dataframe(label="Similar Articles", headers=["Title", "Link"],row_count=5)]
title = "Group 2- ChatGPT text detection module"
description = '''Please upload text files and text input responsibly and await the explainable results. The approach in place includes finetuning a Roberta model for text classification.Once the classifications are done the decision is exaplined thorugh the SHAP text plot.
The probability is particularly explained by the attention plots through SHAP'''
gr.Interface(fn=process_text,title=title,description=description, inputs=[text_input], outputs=outputs).launch()