Unsupervised Semantic Approach of Aspect-Based Sentiment Analysis for Large-Scale User Reviews

Aspect-based sentiment analysis (ABSA) has recently attracted increasing attention due to its extensive applications. Most of the existing ABSA methods been applied on small-sized labeled datasets. However, real datasets such as the Amazon and TripAdvisor contain a massive number of reviews. Thus, applying these methods on large-scale datasets may produce inefficient results. Furthermore, these existing methods extract huge number of aspects, most of which are not relevant to the domain of interest. But, on other hand, some of the infrequent relevant aspects are excluded during the extraction process. These limitations negatively affect the performance of the ABSA process. This article, therefore, aims to overcome such limitations by proposing an efficient approach that is suitable for real large-scale unlabeled datasets. The proposed approach is a combination of hybridizing a frequency-based approach (word level) and a syntactic-relation based approach (sentence level). It was enhanced further with a semantic similarity-based approach to extract aspects that are relevant to the domain, even terms (related to the aspects) are not frequently mentioned in the reviews. The extracted aspects according to the proposed approach are used to generate a total review sentiment score after estimating the weight and the rating of each extracted aspect mentioned in the review. The assignment of the weight of each extracted aspect is calculated based on a modified TF-IDF weighting scheme and the assignment of the aspect rating is calculated based on a domain-specific lexicon. Effectiveness of the extracted aspects is evaluated against two baselines available from existing literature: fixed aspect and extracted aspects. Evaluation was also performed by using a general lexicon and a domain-specific lexicon. Results in terms of F-measure and accuracy on Amazon and Yelp datasets show that the extracted aspects using the proposed approach with the domain-specific lexicon outperformed all the baselines.