© 2021, Emerald Publishing Limited.Purpose: Machine learning algorithms are useful to effectively analyse, and therefore automatically classify online reviews. The purpose of this paper is to demonstrate a novel text-mining framework and its potential for use in the classification of unstructured hotel reviews. Design/methodology/approach: Well-known data mining methods (i.e. boosted decision trees (BDT), classification and regression trees (C&RT) and random forests (RF)) in conjunction with incorporating five-fold cross-validation are used to predict the star rating of the hotel reviews. To achieve this goal, extracted features are used to create a composite variable (CV) to deploy into machine learning algorithms as the main feature (variable) during the learning process. Findings: BDT outperformed the other alternatives in the exact accuracy rate (EAR) and multi-class accuracy rate (MCAR) by reaching the accuracy rates of 0.66 and 0.899, respectively. Moreover, phrases such as “clean”, “friendly”, “nice”, “perfect” and “love” are shown to be associated with four and five stars, whereas, phrases such as “horrible”, “never”, “terrible” and “worst” are shown to be associated with one and two-star hotels, as it would be the intuitive expectation. Originality/value: To the best of the knowledge, there is no study in the existent literature, which synthesizes the knowledge obtained from individual features and uses them to create a single composite variable that is powerful enough to predict the star rates of the user-generated reviews. This study believes that the proposed method also provides policymakers with a unique window in the thoughts and opinions of individual users, which may be used to augment the current decision-making process.