Aspect Oriented Suggestion Extraction from Online Reviews

In the world of business, products need to evolve adjusting to the needs of the customer to ensure customer satisfaction. The abundance of opinionated text on the internet contains suggestions made by users that can be used to improve products. These suggestions are important to multiple stakeholders; as product improvements for businesses or as tips and advice to consumers. Until now, the extraction of suggestions has usually been defined as a problem of classifying sentences into suggestion and non-suggestion classes. No work has attempted to differentiate suggestions on the basis of intended receiver. Extracting these suggestions with respect to the aspects that reviewers are not satisfied with, could be used as a potential solution to improve products. To address these shortcomings, this study proposes a novel task decomposition called Aspect Oriented Suggestion Extraction to identify product improvements. It contains three main subtasks: suggestion classification, beneficiary classification and aspect extraction. The proposed approach proved to be very effective in determining suggestions from non-suggestions, achieving an F-score of 91% with BERT pretrained language model, outperforming state-of-the-art methods on suggestion mining. Experimental results on hotel reviews show the effectiveness of the techniques. The proposed framework, being the first of its nature, yields promising results.