Aspect-Based Review Mining

Ordinary document-level sentiment analysis assigns a single polarity, positive or negative, to a whole review, which loses crucial detail: a review can love one feature and hate another. Aspect-based review mining instead identifies the specific product features customers mention and determines sentiment toward each one separately, then aggregates these into a feature-by-feature summary. Hu and Liu introduced exactly this feature-level framing in 2004, extracting features, linking opinions to them, and summarizing per feature. The practical difference is decisive for marketing: overall sentiment tells you a product is liked, while aspect-based mining tells you the screen is loved and the battery is criticized, which is the actionable insight a product or brand team needs.

Hu and Liu's key observation is that the features customers care about are the ones they mention repeatedly, and that features are typically nouns or noun phrases. The pipeline part-of-speech tags the reviews, treats nouns and noun phrases as candidate features, and uses frequent-itemset (association-rule) mining to keep those that appear in enough reviews to exceed a support threshold, the frequent features. It then recovers additional, less frequent features by finding nouns that appear near known opinion words, on the logic that an adjective expressing an opinion is usually opining about some feature nearby. This combination captures both the obvious, often-discussed aspects and some rarer ones, though very rare or implicitly referenced features can still be missed.

They are useful but imperfect, and several factors distort them. Linguistically, sarcasm, comparatives, conditionals, and especially negation are hard, mishandling 'not good' flips the result, and domain-specific vocabulary can fool a generic sentiment lexicon. At the corpus level, reviews suffer selection bias, people with extreme experiences are more likely to post, and fake or incentivized reviews can skew the counts, so the proportion of positive to negative mentions may not reflect the true customer base. Hu and Liu's original lexicon-and-rule approach also makes errors that modern supervised and transformer-based methods reduce but do not eliminate. Best practice is to validate a sample of automated labels against human judgment, handle negation carefully, tune the lexicon to the domain, and interpret feature summaries as strong indicators rather than exact measurements.