Sentiment Analysis (SA) helps to automatically and meaningfully discover hotel customers’ satisfaction from their shared experiences and feelings on social media. Several studies have been conducted to improve the precision of SA in the hospitality industry, which vary in data preprocessing techniques, feature representation, sentiment classification levels, and models, and they use different datasets. Such variations are worthy of attention and monitoring. Despite the importance of SA in hospitality and tourism, review studies identifying gaps and suggesting future research directions are limited. This article introduces a systematic literature review to label and discuss state-of-the-art studies that deal with SA for hotel reviews.
Recently, the amount of content generated on online hospitality platforms has increased exponentially and has changed people's ways of life. Consumers often refer to online reviews before deciding which hotel to choose. These reviews provide firsthand information, essential to improving hotel services' quality. However, the massive amount of review data and its unstructured nature make it a difficult challenge. Indeed, many researchers were interested in exploring the field of sentiment analysis in the hotel industry. In particular, they have given more attention to aspect-based sentiment analysis, which categorizes opinions by aspect and identifies the sentiment related to each aspect. However, studies examining the Arabic language are limited compared to English. Our paper aims to explore aspect category detection as a sub-task of aspect-based sentiment analysis using Arabic reviews. We relied on the SemEval-2016 Arabic dataset for hotel reviews. As this data suffers from an imbalanced distribution, we propose an approach for multi-label data augmentation of the minority classes in this used dataset. Then, we propose a specific preprocessing for this Arabic reviews dataset. Our aspect category prediction approach is based on the classifier chains technique. In fact, unlike previous works that treat each label separately, we handle the dependencies between the various labels. Our findings show that our proposed approach achieves a good F1score that outperforms the pioneering related work approaches.