TripAdvisor Review Sentiment Mining

Document-level sentiment analysis assigns one overall polarity, positive, negative or neutral, to a whole review. Aspect-based sentiment analysis recognises that a single review can feel differently about different features, so it identifies the aspects discussed (room, service, location, value, cleanliness) and assigns sentiment to each. As Xiang and colleagues showed, decomposing hotel reviews this way reveals which specific facets drive satisfaction, information a single overall score hides, which is why aspect-level analysis is often preferred for actionable hospitality insight.

Pang and Lee review both. Lexicon-based methods score text by counting and weighting sentiment-bearing words from a dictionary; they need no training data and are transparent, but can miss context and domain-specific usage. Supervised machine-learning methods (Naive Bayes, SVM, or modern neural and transformer models) learn from labelled reviews, often using star ratings as labels, and usually achieve higher accuracy in-domain but require training data and are less interpretable. The choice depends on available labelled data, accuracy needs and the value placed on transparency.

Review text is full of the hard cases Pang and Lee highlight: negation ('not clean'), sarcasm and irony ('what a luxurious broom closet'), mixed sentiment within one review, comparatives, and domain-specific or multilingual language. Star ratings used as labels can also diverge from the text. These factors, plus reviewer self-selection and fake reviews, mean accuracy varies by dataset, so models should be validated against human-coded samples using precision, recall and F1 before their aggregate outputs are trusted.