Text Analysis on Indonesian Languages Using NusaX

Project Overview

This project explores text analysis tasks using the NusaX dataset, a multilingual corpus for Indonesian and Javanese languages. The focus is on two key NLP tasks:

1. Machine Translation: Translating between Indonesian-English and Indonesian-Javanese using pre-trained models.
2. Sentiment Analysis: Implementing classical and advanced machine learning models for sentiment classification of Indonesian text.

---

Features

• Machine Translation Models:
  • Helsinki-NLP/opus-mt-id-en: Pre-trained model for Indonesian-English translation.
  • IndoBenchmark/IndoBART: Pre-trained model for Indonesian-Javanese translation.
• Sentiment Analysis Models:
  • Classical Models: SVM, MLP, and Word2Vec embeddings.
  • Pretrained Models: VADER and RoBERTa.

---

Requirements

Ensure you have the following libraries installed:

• scikit-learn
• torch
• transformers
• nltk
• gensim
• matplotlib

Installing Dependencies

Run this in a cell within the Jupyter Notebook to install all dependencies:

!pip install scikit-learn torch transformers nltk gensim matplotlib

Usage

1. Dataset Preparation

1. Download the NusaX dataset from its official repository.
2. Organize the dataset into:
   • Machine Translation: Pairwise datasets for Indonesian-English and Indonesian-Javanese.
   • Sentiment Analysis: Labeled datasets for sentiment classification in Indonesian.

2. Machine Translation

Run the machine translation cells in the notebook to:

• Perform translations using the following models:
  • Indonesian-English: Helsinki-NLP/opus-mt-id-en.
  • Indonesian-Javanese: IndoBenchmark/IndoBART.

3. Sentiment Analysis

Run the sentiment analysis cells in the notebook to:

• Train and evaluate the following models:
  • Classical Models: SVM, MLP.
  • Transformer-Based Models: RoBERTa and other pretrained models.

4. Visualizations

• Visualize confusion matrices and accuracy trends for sentiment analysis models.
• Evaluate BLEU scores for machine translation tasks.

---

Results

Machine Translation

Model	Translation Task	BLEU Score
Helsinki-NLP/opus-mt-id-en	Indonesian to English	32.4
IndoBenchmark/IndoBART	Indonesian to Javanese	27.8

Sentiment Analysis

Model	Accuracy	Precision (Macro Avg)	Recall (Macro Avg)	F1-Score (Macro Avg)
SVM	79%	0.77	0.77	0.77
MLP	77%	0.78	0.75	0.76
VADER	37.25%	0.16	0.33	0.19
Word2Vec + SVM	48%	0.43	0.39	0.39

---

References

1. Helsinki-NLP. opus-mt-id-en. Hugging Face. Retrieved from https://huggingface.co/Helsinki-NLP/opus-mt-id-en.
2. IndoBenchmark. IndoBART. Hugging Face. Retrieved from https://huggingface.co/indobenchmark/indobart.
3. IndoNLP. NusaX: Multilingual Parallel Corpus for Indonesian Local Languages. GitHub. Retrieved from https://github.com/IndoNLP/nusax.

---

Future Improvements

• Expand analysis to include additional local languages supported by NusaX.
• Fine-tune transformer models with a more diverse and larger dataset for improved sentiment classification.
• Optimize models for computational efficiency to enable broader deployment in resource-constrained environments.

---

Acknowledgments

This project was made possible through the use of pre-trained models and datasets available on Hugging Face and the NusaX repository.