Attention-based Sentence Extraction for Aspect-based Sentiment Analysis with Implicit Aspect Cases in Hotel Review Using Machine Learning Algorithm, Semantic Similarity, and BERT

Abstract

The development of the aspect-based sentiment analysis (ABSA) method to work on the case of implicit hotel reviews in depth has not been done much. The problem of extracting aspect and opinion words based on syntaxis and semantics is not only influenced by different of sentence structure types but can also be influenced by word sense disambiguation (WSD) level. So, it needs deep attention to solve these problems. For example, the review "You can't say its cheap because food is cheaper in Chinatown.", where "food is cheaper in Chinatown" is still widely extracted as target terms because there are explicit element of aspect and opinion. In fact, it requires in-depth attention to be able to extract and capture the implicit element "can't say its cheap" as a target term. However, there has been not many research that discusses the details of the ABSA process related to this case. Therefore, we propose an attention-based sentence extraction method for ABSA with implicit aspect cases in hotel review. The method purpose is to improve the ABSA accuracy for hotel reviews based on the cases that have not been solved. First, we develop a pre-processing method to the make the data ready to be processed. Then, we build a set rule-based algorithm to get the word types and the relationship of each word in the sentence. These rules function to identify and mark the candidates of aspect and opinion terms based on the review sentence structure types (simple, compound, complex, compound-complex) and to identify and mark the factors that influence the WSD level (conjunction, punctuation, contrast, intensification) in each sentence. The candidates result of aspect and opinion terms are used as input for the aspect categorization process. The aspect categorization process is carried out using machine learning algorithm, implicit aspect corpus, BERT embedding, and semantic similarity to obtain the aspect categories of each review. Furthermore, the ABSA process is carried out using the BERT sentiment analysis method. Finally, the evaluation process for aspect categorization and ABSA are done with the good result. The evaluation result of aspect categorization obtains 91.31% for accuracy, 91.81% for precision, 89.43% for recall, and 90.61% for f1-measure. Meanwhile, the evaluation result of ABSA obtains 98.10% for accuracy, 98.11% for precision, 96.98% for recall, and 97.54% for f1-measure.