A Novel Approach for Enhancing Vietnamese Sentiment Classification

Abstract

Sentiment analysis has become an essential and fundamental task in natural language processing that has many useful applications. There exist various methods for solving the sentiment classification problem using traditional machine learning models, deep neural networks, and transfer learning methods. It is still interesting for Vietnamese sentiment classification problems to understand the impact of different ensemble techniques and deep learning approaches for building the most suitable model. This work aims to study the Vietnamese sentiment classification on one public dataset used in the Vietnamese Sentiment Analysis Challenge 2019 and another large-scale dataset, namely “AISIA-Sent-002”, which was collected from the Vietnamese e-commerce websites by ourselves. We explore five distinct ensemble schemes, including the classic methods (Uniform Weighting and Linear Ensemble) and the feature importance-based advanced methods (Gating Network, Squeeze-Excitation Network, and Attention Network) for Vietnamese sentiment analysis. We do these ensemble techniques with five individual deep learning models: TextCNN, LSTM, GRU, LSTM + CNN, and GRU + CNN. Extensive experiments on two datasets show that the ensemble methods perform much better than any individual model and significantly outperform the competition’s winning solution with a large margin. Finally, we aim to publish our source codes to contribute to the current research community related to natural language processing.

C. V. Nguyen and K. H. Le—Equal contribution.