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:

python

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

AvinabhDutta-Dev/assamese-sentiment-svm

Assamese Sentiment Analysis — SVM Baseline