/Sentiment-Analysis-of-Marathi-Tweets

A Comparitive study of Different Transformer Based Architecture.

Primary LanguageJupyter NotebookGNU General Public License v3.0GPL-3.0

Sentiment-Analysis-of-Marathi-Tweets

We have performed a comparitive study of different transformer based architectures, on top of deep learning models like fully connected feed forward neural netword, LSTM and BiLSTM for a downstream task of sentiment analysis on Marathi tweets. The tweets can be classified in three classes, positive, negtive and neutral. Details about dataset and evaluation are given below.

Dataset statistics

There are 15,864 tweets, that are divided into training set, test set and validation set, and are class balanced. The dataset statistics is shown below:

Split Positive Negative Neutral Total
Train 4038 4038 4038 12114
Test 750 750 750 2250
Validation 500 500 500 1500

Models

We have used Transformer based models for language modelling and feeded the generated embeddings in the deep learning models. The models used are:

  • Transformer based:

    • mBERT
    • IndicBERT
    • MuRIL (baseline)

  • Deep Learning models

    • Dense fully connected NN
    • LSTM
    • BiLSTM

  • Machine Learning models

    • Naive Bayes
    • Linear SVM
    • SVM Radial basis functions kernel
    • Random Forest