Multidomain Sentiment Analysis for Marathi

A Technical Guide Using Transformer Models

Project Context

This project addresses the task of sentiment analysis for the Marathi language, which, despite being spoken by over 83 million people, is considered a low-resource language in the field of Natural Language Processing (NLP). The primary challenge is building a robust model that performs well across multiple domains (e.g., movie reviews, social media, news). Transformer-based models have shown great promise for such tasks, but their application requires specific strategies to overcome data scarcity and domain-specific linguistic nuances. This guide outlines the key challenges and a systematic approach to tackle them.

📉

Data Scarcity

High-quality, labeled datasets for Marathi are limited, especially across multiple domains, making it difficult to train models from scratch.

🔄

Domain Shift

Language use, vocabulary, and context can vary significantly between domains (e.g., formal news vs. informal tweets), causing models trained on one domain to perform poorly on another.

🗣️

Code-Mixing

Marathi text, especially on social media, often mixes Devanagari script with Latin script (English words), a phenomenon that models must be able to handle effectively.