Enhancing Sentiment Analysis Effectiveness with LSTM Variants, and Stratified K-Fold on Imbalanced Dataset

Andriyanto, Rifki
Kusrini, Kusrini

Building of Informatics, Technology and Science
Vol 6 No 2 (2024): September 2024

Sentiment analysis on hotel reviews often faces the challenge of class imbalance, where positive reviews significantly outnumber negative or neutral ones. This study aims to improve the effectiveness of sentiment analysis models on imbalanced hotel reviews by examining combinations of word embedding methods (FastText, Word2Vec, Doc2Vec) and model architectures (LSTM, BiLSTM, BiLSTM-Attention). Class imbalance is addressed using SMOTE, and model evaluation is conducted using Stratified K Fold cross-validation. Results show that Doc2Vec consistently outperforms FastText and Word2Vec as a word embedding method, especially when combined with the BiLSTM-Attention architecture. The use of SMOTE and Stratified K Fold also proves effective in improving model performance on imbalanced datasets. This study concludes that the selection of appropriate word embedding methods and model architectures, along with the implementation of class imbalance techniques, is crucial in developing effective and robust sentiment analysis models for hotel reviews.