Skip to content

IMPLEMENTATION_DETAILS.md

Latest commit

 

History

History
73 lines (55 loc) · 3.91 KB

IMPLEMENTATION_DETAILS.md

File metadata and controls

73 lines (55 loc) · 3.91 KB

Technical Documentation: Sentiment Analysis Web App

This document provides a comprehensive overview of the technical architecture, algorithms, and implementation logic used in the Sentiment Analysis Web Application.

1. Technology Stack

Backend

  • FastAPI: A high-performance Python web framework used to serve the analysis endpoints.
  • Uvicorn: An ASGI web server implementation for Python, used to run the FastAPI application.
  • Scikit-learn: Used for the traditional Machine Learning pipeline (TF-IDF + Logistic Regression).
  • HuggingFace Transformers: Used to load and inference the BERT (Bidirectional Encoder Representations from Transformers) model.
  • PyTorch: The underlying deep learning framework for BERT inference.
  • Datasets (HuggingFace): Used in the training script to fetch real-world IMDB movie reviews.

Frontend

  • HTML5: Semantic structure of the application.
  • Vanilla CSS3: Implements a Dynamic Theme System (Light/Dark mode) using CSS Custom Properties (Variables) and data-attributes.
  • Vanilla JavaScript (ES6+): Handles asynchronous API calls (using fetch), DOM manipulation, and persistence of theme preferences in LocalStorage.

2. Machine Learning Algorithms & Models

The application employs two distinct approaches to sentiment analysis to provide comparative insights.

Approach A: TF-IDF + Logistic Regression

  • Vectorization (TF-IDF): Converts text into numerical vectors, emphasizing unique words and de-emphasizing common stop words.
  • Preprocessing Pipeline: Every input is stripped of HTML tags, URLs, and extra whitespace before analysis to ensure accuracy.
  • Neutrality Guard: If the model confidence is < 60%, the result is labeled as Neutral. This prevents the model from forcing a guess on non-opinionated text (like "Hello").

Approach B: BERT (Transformers) - Optimized

  • Model: distilbert-base-uncased-finetuned-sst-2-english. A transformer-based model that uses self-attention to understand context bidirectionally.
  • Optimized Word Importance: Since running BERT is computationally expensive, the importance calculation now:
    1. Filters out common stop-words (the, is, an).
    2. Limits analysis to the 15 most meaningful words in the sentence.
    3. Uses a perturbation-based "Leave-One-Out" strategy to measure confidence drops.

3. UI/UX Design System

Theme Management

The application features a Dual-Theme system:

  • Light Mode: A clean, minimalist "SaaS-style" aesthetic with high readability.
  • Dark Mode: A professional, deep-navy theme optimized for low-light environments.
  • Persistence: Your choice is saved in localStorage, so the app remembers your preference even after a refresh.

Visualizations

  • Horizontal Bar Charts: Dynamically generated bars that scale proportionally based on the most influential word in the set.
  • Mood Emojis: Dynamic badges that change based on sentiment (😊 Positive, 😠 Negative, 😐 Neutral).
  • Robust Error Handling: Each model card has its own error state. If one model fails to respond, the other will still display its results, ensuring a "graceful failure" experience.

4. API & Directory Structure

Endpoint Logic

  1. POST /predict/tfidf: Returns sentiment analysis via the Scikit-learn pipeline.
  2. POST /predict/bert: Returns high-accuracy sentiment via the Transformer model.

Directory Overview

SentimentAnalysis/
├── main.py                    # Backend server & Inference logic
├── train_tfidf_real.py        # Standalone script to train on IMDB data
├── index.html                 # Frontend structure & Theme toggle
├── style.css                  # Design system & CSS Variables
├── script.js                  # Async API logic & Dynamic UI
├── models/                    # Saved .pkl models and BERT checkpoints
└── README.md                  # Project overview and setup