Text In Sight: Transformers Based Sentiment and Thematic Analysis System

TextInsight is a transformer driven NLP system that performs sentiment classification, thematic zero shot categorization, summarization, token-level explainability, and dataset level analytics. The project integrates multiple pre-trained transformer models into a unified pipeline and provides an interactive Streamlit dashboard for real-time analysis.

Overview

The system analyzes user-generated text such as reviews, feedback, comments, and short documents. Using modern transformer models, it can:

• Classify sentiment
• Identify themes using zero-shot inference
• Generate concise summaries
• Provide token-level explanations
• Support batch-level analytics and visualizations

Both single-text and CSV-based workflows are supported.

Key Features

1. Sentiment Analysis

Model: distilbert-base-uncased-finetuned-sst-2-english
Outputs:

• Sentiment label (Positive/Negative)
• Confidence score
• Token-level explanation using attention-based scoring

2. Thematic Classification (Zero-Shot)

Model: facebook/bart-large-mnli
Predicts themes without additional training.

Supported domains:

• Education
• Technology
• Healthcare
• Finance
• Entertainment
• AI
• User Experience

Outputs:

• Top theme label
• Theme confidence score
• Score breakdown for each candidate theme

3. Text Summarization

Model: sshleifer/distilbart-cnn-12-6
Generates summaries for:

• Entire datasets
• Theme-specific subsets

4. Token-Level Explainability

Uses transformer attention weights to compute token contributions toward sentiment.

5. Batch CSV Analysis

For each entry, the system:

• Predicts sentiment
• Assigns a theme
• Computes scores
• Generates summaries
• Produces WordClouds and analytics

6. Visualizations

Includes:

• Sentiment distribution (Pie chart)
• Theme distribution (Bar chart)
• WordClouds per theme

7. Model Evaluation

Provides:

• Accuracy
• Precision
• Recall
• F1-score
• Confusion matrix
• Classification report

Models Used

Sentiment Model

distilbert-base-uncased-finetuned-sst-2-english
Distilled BERT fine-tuned on SST-2.

Thematic Model (Zero-Shot)

facebook/bart-large-mnli
BART model trained on MNLI for zero-shot classification.

Summarization Model

sshleifer/distilbart-cnn-12-6
Distilled BART trained on CNN/DailyMail.

Tech Stack

• Python
• Streamlit
• Hugging Face Transformers
• PyTorch
• Plotly Express
• Matplotlib
• WordCloud
• Pandas & NumPy
• scikit-learn

Project Structure

TextInsight/
├── app.py
├── model_pipeline.py
├── model_evaluation.py
├── evaluation_data.csv
├── requirements.txt
└── README.md

Running the Application

Install dependencies:

pip install -r requirements.txt

Run the application:

streamlit run app.py

Running Evaluation

Prepare a CSV with:

• text
• true_sentiment
• true_theme

Then run:

python3 model_evaluation.py

This prints accuracy, precision, recall, F1-score, and confusion matrices.

Made by SHIVANSH TRIPATHI