README.md

metadata

title: Narrative Stance Analyzer
emoji: 🧠
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false

Narrative Stance Analyzer

Overview

The Narrative Stance Analyzer is a public, interactive web application built using pretrained Transformer models to classify the narrative stance of text related to climate communication. The model predicts whether a given message reflects a believer, neutral, or denier perspective.

This project is designed to help researchers, communicators, and advocates better understand how narratives are framed in public discourse.

---

Use Case

Prompt:
“Given a campaign message, news headline, or paragraph, analyze its narrative stance as believer, neutral, or denier using emotion and framing cues.”

Example Input:

"Climate change is the biggest threat to our future. We must act now."

Output:

• Predicted Stance: believer
• Confidence scores from SBERT and DeBERTa
• Final ensemble prediction

---

Models Used

1. SBERT + MLP Classifier

• Embeddings from: all-MiniLM-L6-v2 (via Sentence-Transformers)
• Trained classifier: Logistic Regression with balanced class weights

2. DeBERTa (Base)

• Model: microsoft/deberta-v3-small
• Used for zero-shot classification (no fine-tuning)
• Tokenizer + model loaded from saved checkpoint

3. Ensemble Logic

• Final prediction is a weighted ensemble:
  • 60% SBERT + 40% DeBERTa confidence scores

---

How It Works

1. User enters a text message
2. The app:
   • Generates sentence embedding using SBERT
   • Tokenizes and classifies with DeBERTa
   • Combines both predictions into a final ensemble score
3. Returns the top label and confidence scores

---

🖥 Interface

Built with Gradio for fast prototyping and deployment on Hugging Face Spaces.

---

Running the App

Local Installation

git clone https://huggingface.co/spaces/yourusername/narrative-stance-app
cd narrative-stance-app
pip install -r requirements.txt
python app.py