Aspect-Based Sentiment Analysis for Service Industry

In today’s digital age, customer feedback, particularly gathered from various sources like mobile application reviews, has emerged as a critical resource for service-providing organizations to gain valuable insights into their customers’ experiences. As the key objective of service-providing organizations is to facilitate their customers with better services, customer feedback or opinion is a vital resource for such organizations to improve and enhance their services for the betterment of their customers. Explicitly mentioned opinions have been widely studied in research, while a significant gap exists in addressing implicitly described views. Furthermore, most existing research focuses on product-oriented corpora, emphasizing specific product aspects and features. This article presents a novel study on performing end-to-end aspect-based sentiment analysis (ABSA) by extracting implicit opinion terms, categorizing them, and assigning polarity to each term from mobile app reviews in English. Through this study, we developed a domain-specific, service-oriented, and aspect-based annotated dataset and introduced a novel two-step hybrid approach. The first step involves extracting multiple opinion terms using a rule-based approach. The second step employs machine learning and deep learning algorithms to classify the extracted opinion terms into general aspect categories. This two-step approach effectively addresses the double-implicit problem commonly encountered in the previous work on implicit aspects and opinion mining. In addition to traditional machine learning and deep learning models, we fine-tuned BERT to carry out the ABSA task. This approach utilized a pipeline method, where each task’s output serves as the subsequent task’s input, ensuring a seamless flow of information and improved performance. This multi-step pipeline begins with the extracted opinion terms classification into aspect categories and ends with the assignment of sentiment polarity. Experiments with a hold-out test set for the first step (opinion term extraction using a rule-based approach) achieved an accuracy and precision score of 81.4% and an F1 score of 0.99 %, outperforming several baselines. Further experiments with a range of machine learning and deep learning algorithms for classifying extracted opinion terms into general aspect categories yielded accuracy scores ranging from 0.68% to 0.74 % and F1 scores ranging from 0.23% to 0.28%. Experiments on sentiment classification using various machine learning algorithms showed accuracy ranges from 0.58% to 0.68% and F1 scores from 0.47% to 0.49%. This two-step approach for implicit opinion term, aspect extraction, and classification outperforms many baseline systems. By leveraging BERT’s contextual understanding and fine-tuning it for our specific domain, we significantly improved the accuracy and robustness of our aspect-based sentiment analysis. This approach effectively captured both explicit and implicit opinions from mobile app reviews. Specifically, our method achieved an accuracy of 0.80% and an F1 score of 0.78% for aspect categorization, and an accuracy of 0.79% and an F1 score of 0.70% for sentiment classification, demonstrating substantial improvements over traditional methods.