Ensemble Learning for Sentiment Analysis of Translation-Based Textual Data

Ensemble learning is a technique that combines several learners to generate a model characterized by more generalized predictions than the constituent learners. Despite the number of conducted studies about ensemble learning of sentiment analysis and the ones that studied the impact of translation on sentiment analysis, the studies that consider ensemble learning on translated text are limited. Here different techniques of ensemble learning such as bagging, boosting, and stacking were applied to classify an English dataset that was translated to modern standard Arabic, which in turn translated manually to Bahraini dialects. Interestingly, this study revealed the outperformance of stacking ensemble based on LSTM as base-learners and decision tree as meta-learner over the other ensemble techniques by achieving 99.52%, 99.25%, and 98.52% of mean accuracy in English, modern standard Arabic, and Bahraini dialects, respectively.