LibSenti is an end-to-end, AI-powered application that leverages advanced machine learning and natural language processing (NLP) techniques to classify sentiment polarity — Positive, Neutral, or Negative — from student-submitted reviews of IIT and NIT libraries. It combines a robust backend sentiment analysis engine with a visually rich, interactive Streamlit application for real-time review exploration, data insights, and institutional benchmarking.

• ✅ Real-time Review Prediction (📝 Sentiment Predictor Tab)
  Users can input custom reviews to receive live sentiment predictions with confidence probabilities and visual feedback.

• ✅ Unigram WordCloud Visualization (🔤 Unigram WordClouds Tab)
  Generates wordclouds for individual institutions using most frequent single-word terms found in reviews.

• ✅ Bigram WordCloud Comparison (🔗 Bigram WordClouds Tab)
  Displays two-institution comparison of most common word pairs (bigrams) extracted from reviews.

• ✅ Sentiment Pie Chart Comparison (🔍 Pie Chart Comparison Tab)
  Side-by-side sentiment distribution pie charts for any two selected institutions, including precise percentage labels.

• ✅ IIT vs NIT Sentiment Analysis (📊 IIT vs NIT Chart Tab)
  Presents a consolidated sentiment comparison chart contrasting IITs and NITs at a glance.

• ✅ Library Experience Highlights (🌟 Library Experiences Tab)
  Displays standout user-submitted reviews—both best and worst experiences—curated by sentiment and length.

• Architecture: BERTForSequenceClassification
• Dataset: IIT & NIT library reviews (labeled Positive, Neutral, Negative)
• Frameworks: PyTorch, Transformers (HuggingFace)
• Accuracy: ~96% on test data
• Label Distribution Handling: Threshold-based for confident classification

LibSenti/ ├── assets/ │ ├── wordclouds/ # Wordcloud PNGs for each IIT/NIT │ └── iit_vs_nit_sentiment_comparison.png ├── saved_model/ # Trained BERT model and tokenizer ├── app.py # Main Streamlit application ├── train_model.py # Script to train the BERT model ├── cleaned_iit+nit_library_reviews.csv ├── sentiment_iit_library_reviews.csv └── README.md # This file

Class imbalance is addressed using weighted loss during training and probability thresholds during prediction.

• Base Model: bert-base-uncased
• Framework: Hugging Face Transformers + PyTorch
• Strategy: Fine-tuned with weighted cross-entropy loss for class imbalance
• Threshold logic for better handling of imbalanced classes
• Trained using Trainer API with evaluation metrics and early stopping

Run training script:

python train_model.py

This script handles:

• Preprocessing
• Tokenization
• Model fine-tuning
• Class weight balancing
• Model saving to ./saved_model/

Launch the Application:

streamlit run app.py 

• 📥 Review Classifier (📝 Sentiment Predictor Tab)
  Enter any library review and instantly receive a sentiment prediction (Positive, Neutral, Negative).

• 📈 Sentiment Probabilities (📝 Sentiment Predictor Tab)
  Visualize the confidence scores for each sentiment using interactive progress bars to assess prediction certainty.

• ☁️ WordCloud Comparator

  • 🔤 (Unigram WordClouds Tab): Select and compare two institutions to explore most frequent individual keywords.
  • 🔗 (Bigram WordClouds Tab): Compare most common two-word combinations to find phrase patterns in reviews.

• 📊 Sentiment Pie Chart Comparison (🔍 Pie Chart Comparison Tab)
  Instantly loads sentiment distribution charts for selected institutions side-by-side for intuitive visual analysis.

• 🧮 IIT vs NIT Overall Chart (📊 IIT vs NIT Chart Tab)
  A comparative sentiment distribution chart to analyze trends across all IITs vs NITs.

• 🌟 Library Experience Highlights (🌟 Library Experiences Tab)
  Shows handpicked positive and negative user reviews with institution tags and styled formatting.

• Add LIME/SHAP Explainability for BERT
  Integrate model interpretation techniques to explain why a review was labeled positive/negative.

• Include More Institutions
  Expand dataset to cover regional universities, IIITs, NLUs, and other public libraries for broader benchmarking.

• Review Metadata Integration
  Include attributes like review date, source, device, or student/staff tag for richer context and filtering.

• Clustering or Topic Modeling
  Apply LDA/BERT-topic to identify trending topics or issues discussed across institutions.

• Sentiment Timeline Analysis
  Show how sentiment for a specific institution evolves over time (e.g., semester-wise or pre/post renovation).

• User Feedback Module
  Allow users to correct or rate the model's prediction to improve performance and trust.

• Multilingual Support
  Add language detection and support for Hindi, Tamil, etc., using multilingual BERT (e.g., xlm-roberta-base).

Aman Srivastava [amansri345@gmail.com]