System Overview

This project is an end-to-end sentiment analysis system designed to extract, preprocess, and classify online discussions using a multitask DeBERTa-based transformer model.

Architecture Diagram

FrontendComments ExtractionClassificationVisualisationBackendPreprocessingAPI CallsAPI CallsHugging FaceTrained Model

The diagram illustrates the interaction between the frontend, backend services, and the deployed Hugging Face model.

Multitask DeBERTa Model

At the core of this system is a fine-tuned DeBERTa transformer model hosted on Hugging Face. The model is trained to perform multiple natural language understanding tasks simultaneously, enabling comprehensive sentiment and content analysis from a single inference pass.

The model is publicly available on Hugging Face and can be accessed programmatically via inference APIs:

View model on Hugging Face

Model Architecture

Input LayerDeBERTaV3-base12 Encoder BlocksTask-SpecificLSTMsTask-SpecificAttentionTask-SpecificDenseFusion +NormalizationTask-SpecificClassifiersTask Outputs

The multitask DeBERTa architecture processes shared contextual embeddings through task-specific branches, followed by feature fusion and dedicated classification heads for each prediction objective.

End-to-End Workflow

  1. Comment Extraction: User-provided links (e.g., Reddit or YouTube) are parsed and relevant comments are retrieved.
  2. Preprocessing: Text data is cleaned, normalized, and structured before inference.
  3. Model Inference: The processed text is sent to the Hugging Face-hosted DeBERTa model via backend API calls.
  4. Classification: Predictions for sentiment, emotion, and hate speech are returned.
  5. Visualization: Results are aggregated and presented through interactive charts and tables in the frontend.