Modified EDA and Backtranslation Augmentation in Deep Learning Models for Indonesian Aspect-Based Sentiment Analysis

Abstract

In the process of developing a business, aspect-based sentiment analysis (ABSA) could help extract customers' opinions on different aspects of the business from online reviews. Researchers have found great prospective in deep learning approaches to solving ABSA tasks. Furthermore, studies have also explored the implementation of text augmentation, such as Easy Data Augmentation (EDA), to improve the deep learning models’ performance using only simple operations. However, when implementing EDA to ABSA, there will be high chances that the augmented sentences could lose important aspects or sentiment-related words (target words) critical for training. Corresponding to that, another study has made adjustments to EDA for English aspect-based sentiment data provided with the target words tag. However, the solution still needs additional modifications in the case of non-tagged data. Hence, in this work, we will focus on modifying EDA that integrates POS tagging and word similarity to not only understand the context of the words but also extract the target words directly from non-tagged sentences. Additionally, the modified EDA is combined with the backtranslation method, as the latter has also shown quite a significant contribution to the model’s performance in several research studies. The proposed method is then evaluated on a small Indonesian ABSA dataset using baseline deep learning models. Results show that the augmentation method could increase the model’s performance on a limited dataset problem. In general, the best performance for aspect classification is achieved by implementing the proposed method, which increases the macro-accuracy and F1, respectively, on Long Short-Term Memory (LSTM) and Bidirectional LSTM models compared to the original EDA. The proposed method also obtained the best performance for sentiment classification using a convolutional neural network, increasing the overall accuracy by 2.2% and F1 by 3.2%. Doi: 10.28991/ESJ-2023-07-01-018 Full Text: PDF