Indonesian Sentiment Analysis

Project Overview

This project explores the potential of combining Graph Neural Networks (GNNs) and IndoBERT, a pre-trained language model for Indonesian, to analyze sentiment in social media data. By leveraging the strengths of both approaches, we aim to achieve a more nuanced and accurate understanding of sentiment expressions in Indonesian text compared to traditional methods. 

Methods

In this project, we employ late fusion technique to combine the embedding from the graph and  IndoBERT. We add parameter λ to give a weight for each embeddings. The method described in the figure below.

Result

As a result, our proposed method improves the performance of text classification task. With the right weight, it can be seen that combining representation from GNN and IndoBERT respectively can improve the accuracy.