🚔 AI Feedback Analytics Platform

Intelligent Text Analysis & Insights Generation System

A Streamlit-powered web application that transforms unstructured feedback into actionable insights using state-of-the-art AI/ML models.

Model Link

https://feedbackanalysis-kphosvhshtsnckpxsmgffh.streamlit.app/

---

📋 Table of Contents

• Overview
• Features
• AI/ML Models
• Tech Stack
• Installation
• Usage
• Configuration
• Project Structure
• API Reference
• Screenshots
• Model Working
• Future Improvements

---

🎯 Overview

The AI Feedback Analytics Platform processes unstructured text from feedback, news, and social media. It uses machine learning models and extractors to provide insights like sentiment analysis, entity extraction, and competency tagging.

---

✨ Features

🤖 AI-Powered Analysis

• Sentiment classification
• Summarization
• Question-answering
• Language detection

📊 Custom Intelligence

• Entity extraction
• Competency tagging
• Recognition scoring

🎨 Dashboard

• Real-time analytics
• Bar charts
• Recent submissions table

💾 Export Options

• PDF
• CSV & JSON

---

🧠 AI/ML Models

Uses DistilBERT, BART, and RoBERTa for sentiment, summarization, and Q&A.

Recognition Score = min(1.0, Base_Score + Tag_Boost + Length_Boost)

---

🛠️ Tech Stack

• Streamlit
• Transformers
• PyTorch
• Langdetect
• Pandas
• ReportLab
• pdfplumber

---

📦 Installation

Clone repo, create venv, install dependencies, run Streamlit app.

git clone https://github.com/yourusername/ai-feedback-analytics.git
cd ai-feedback-analytics
python -m venv venv
venv/Scripts/activate
pip install -r requirements.txt
streamlit run app.py

---

🚀 Usage

Supports text input, file upload, dashboard insights, Q&A, and export.

---

⚙️ Configuration

Modify districts, departments, competency tags, and model parameters in app.py.

---

📁 Project Structure

ai-feedback-analytics/
├── app.py
├── requirements.txt
├── README.md
├── Screenshots

---

📸 Screenshots

Feedback Form

Analysis Results

Dashboard

Export Data Page

Q&A Chat

---

🔌 API Reference

Includes load_models(), process_text(), extract_simple_entities(), and calculate_recognition_score().

---

📞 Contact

Email: 22cs01075@iitbbs.ac.in

---

➕ Additional Files

requirements.txt

streamlit>=1.28.0
transformers>=4.30.0
torch>=2.0.0
pandas>=2.0.0
langdetect>=1.0.9
reportlab>=4.0.0
pdfplumber>=0.10.0

.gitignore

__pycache__/
*.py[cod]
venv/
.streamlit/
transformers_cache/

Models Working

1. How the AI/ML Models Are Used

We are using three pre-trained "pipeline" models from Hugging Face Transformers.

Sentiment Analysis

• Model: distilbert-base-uncased-finetuned-sst-2-english
• Purpose: Reads the full feedback and decides whether the sentiment is positive or negative.
• Used for:
• Showing the sentiment emoji (😊 or 😞)
• Calculating the Recognition Score (positive → higher score)

Summarization

• Model: facebook/bart-large-cnn
• Purpose: Reads the complete text and generates a shorter paragraph capturing the key points.
• Used for:
• The Summary box in the "View Details" section
• Preview summaries in the Dashboard tab

Question-Answering (Q&A)

• Model: deepset/roberta-base-squad2
• Purpose: Powers the Q&A Chat tab by answering user questions based on all collected feedback.
• Inputs:
• A question from the user
• A combined context made from all processed feedback
• Process: Extracts the exact span of text from the context that answers the question.

---

2. How They Work (Simple Explanation)

Sentiment Analysis (DistilBERT)

A text classifier trained on thousands of reviews. It learned patterns like:

• "excellent", "appreciated", "helpful" → Positive
• "poor", "rude", "disappointed" → Negative

It reads your entire feedback and predicts which label fits best.

Summarization (BART)

A sequence-to-sequence model with two components:

• Encoder: Reads and understands your full text
• Decoder: Generates a brand-new shorter version, word by word

Works like a smart journalist who rewrites long paragraphs into concise summaries.

Question-Answering (RoBERTa)

An extractive model — it doesn’t invent answers. Instead, it:

• Looks at the question
• Searches the context
• Highlights the exact words that answer the question

🔮 Future Improvements

Your current AI system includes two advanced ML components (sentiment, summarization, Q&A) and several simple keyword-based components (districts, departments, competencies). The most impactful upgrades come from improving the latter.

1. Use Zero‑Shot Classification for Competency Tags

Problem: extract_competency_tags() only finds tags if exact keywords appear in text. It misses cases like:

"The officer was kind and understanding" → should match compassion, but keyword isn't present.

Solution: Use a Zero‑Shot Classification model.

• This model understands the meaning of text and labels, even without exact keywords.
• Provide feedback + your labels:
• bravery, professionalism, compassion, life saving, rapid response, etc.
• Model returns the best‑matching competencies with confidence scores.

Benefit: Much smarter, semantic tag extraction.

---

2. Use a Named Entity Recognition (NER) Model for Districts, Departments & Officer Names

Problem: extract_simple_entities() depends on fixed keyword lists.

• Cannot detect new department names.
• Cannot detect officer names (e.g., “Constable Das”).

Solution: Add a pre-trained NER pipeline.

• Identifies PERSON, ORG, and LOCATION entities.
• Extracts:
• Police officer names
• Station/department names
• Districts and localities

Benefit: More accurate and expandable entity extraction.

---

3. Implement RAG (Retrieval‑Augmented Generation) for Smart Q&A

Problem: Current Q&A loads all feedback text into a single context.

• Slow
• Breaks with large datasets (token limit)
• Not scalable

Solution: Use RAG architecture:

1. Store each feedback entry as a vector embedding.
2. Use a vector database (FAISS, Chroma, Pinecone).
3. When user asks a question:

• Retrieve top 3–5 most relevant feedback entries.

4. Feed only those into the Q&A model.

Benefit:

• Faster Q&A
• More accurate answers
• Scales to thousands or millions of entries

---