README.md

metadata

language:
  - as
license: mit
library_name: scikit-learn
tags:
  - sentiment-analysis
  - text-classification
  - assamese
  - svm
  - baseline
  - machine-learning
metrics:
  - accuracy
  - f1
  - precision
  - recall
  - roc_auc
model-index:
  - name: assamese-sentiment-svm
    results:
      - task:
          type: text-classification
          name: Sentiment Analysis
        dataset:
          name: Assamese Movie & Media Reviews
          type: assamese-reviews
        metrics:
          - name: Accuracy
            type: accuracy
            value: 0.850389
          - name: Macro F1
            type: f1
            value: 0.83656
          - name: Macro Precision
            type: precision
            value: 0.843188
          - name: Macro Recall
            type: recall
            value: 0.831493
          - name: ROC-AUC
            type: roc_auc
            value: 0.93135

Assamese Sentiment Analysis — SVM Baseline

This repository contains a classic Support Vector Machine (SVM) model optimized for binary sentiment analysis on Assamese textual reviews. It serves as a foundational machine learning baseline to benchmark and assess the performance improvements offered by custom deep learning architectures and pre-trained transformer language models.

---

Model Summary

• Model Architecture: Linear Support Vector Classification (LinearSVC wrapped in CalibratedClassifierCV)
• Task: Binary Sentiment Classification
• Language: Assamese (as)
• Developer: Avinabh Dutta
• Primary Framework: scikit-learn
• Top Performance: 85.04% Accuracy | 0.8366 Macro F1

---

Architecture & Configuration

The baseline pipeline relies on traditional machine learning text representation techniques:

• Feature Extraction: TF-IDF Vectorization (maps unstructured Assamese text sequences into numerical feature arrays based on corpus-wide statistical weighting).
• Classifier Engine: LinearSVC from scikit-learn using a Linear Kernel.
• Probability Calibration: The core SVM classifier is wrapped inside a CalibratedClassifierCV block to generate calibrated posterior probabilities, enabling accurate ROC-AUC continuous curve assessments.
• Persistence Format: The pipeline components are serialized and saved via joblib.

---

Performance & Evaluation Metrics

As a baseline, the model demonstrated solid foundational accuracy, completing the evaluation phase with the following exact metrics:

Metric	Score	Percentage
Accuracy	0.850389	85.04%
Macro F1-Score	0.836560	83.66%
Macro Precision	0.843188	84.32%
Macro Recall	0.831493	83.15%
ROC-AUC	0.931350	93.14%
Cohen's Kappa	0.673386	67.34%

---

Dataset Overview

The dataset consists of localized social media commentaries, media assessments, and audience reviews covering drama (নাটক) and cultural content in Assamese.

• Target Classes: Binary Classification (0: Negative, 1: Positive).

---

How to Use

Because this is a classic machine learning pipeline saved with joblib, you can load and run inference natively using Python and scikit-learn:

import joblib

# Load the saved SVM model pipeline (including the TF-IDF vectorizer)
model_pipeline = joblib.load("assamese_sentiment_svm.joblib")

# Example Assamese text
sample_text = ["চলচ্চিত্ৰখনৰ কাহিনী আৰু অভিনয় অতি সুন্দৰ হৈছে।"]

# Predict sentiment class and probabilities
prediction = model_pipeline.predict(sample_text)
probabilities = model_pipeline.predict_proba(sample_text)

print(f"Prediction: {prediction}")
print(f"Probabilities: {probabilities}")

---

Limitations & Comparative Context

This SVM baseline model provides essential comparative context for evaluating more advanced neural architectures. However, it suffers from several structural limitations inherent to classic machine learning techniques:

• Strict Bag-of-Words Reliance: Because the TF-IDF vectorizer maps tokens into static, frequency-based vector coordinates, the model is completely blind to word order, long-range semantic syntax, and spatial discourse transitions.
• Vulnerability to Morphological Shifts: The architecture struggles significantly with Assamese language variations, structural morphology shifts, and implicit sentiment alterations—nuances that deep learning layers and transformer attention mechanisms process natively.

---

Citation & Contact

If you use this model or refer to this work in your research, please link back to this Hugging Face repository or the GitHub project repository. If you use this model, please cite the associated preprint:

Dev, C., Baishya, B., Dey, N., Dutta, A., & Dutta, S. (2026). Sentiment Analysis of Assamese Reviews using Transfer Learning Models. Preprint. https://doi.org/10.2139/ssrn.6357982