📄 Extracting Product Features and Opinions from Reviews

Summary

The paper presents OPINE, an unsupervised system that mines product reviews to find the aspects buyers discuss and the opinions they express about them. Review mining is split into four subtasks: finding features, finding opinions about them, deciding polarity, and ranking opinion strength. The paper covers the first three. OPINE is built on the KnowItAll Web extraction framework. It parses reviews with MINIPAR and keeps frequent noun phrases as candidate features. Each candidate is scored by PMI against meronymy-style discriminator phrases, with hit counts taken from the Web as well as the review text. Opinion candidates are found through syntactic dependency rules near known features. Their semantic orientation is then resolved with relaxation labeling, which considers the word, the feature it describes, and the sentence. On Hu and Liu's datasets, OPINE beats the previous system's feature-extraction precision by 22% while losing only 3% recall. It also reports precision and recall for opinion phrase extraction and polarity in context. This matters because shoppers often care about features that star ratings and overall rankings hide.

Key points

• Data: 7 product classes with 1621 reviews in total, including the 5 public datasets from Hu and Liu (2004). Extra Hotels reviews came from tripadvisor.com and Scanners reviews from amazon.com.
• Feature extraction ablation:
  • The Feature Assessor alone raises Hu's precision by 6%.
  • Web PMI statistics add about 14.5% on average.
  • About one third of OPINE's precision gain comes from assessment on reviews and two thirds from Web statistics.
• Feature types: Parts make up 52% of extracted features, features of parts 24%, related concepts 9%, related concepts' features 8%, and properties 7%.
• Hotels and Scanners evaluation:
  • Precision was 89% on 450 extractions, with inter-annotator agreement of 86%.
  • Recall was 73% on 179 agreed features drawn from 800 annotated sentences.
• Opinion tasks (550 sentences with known features):
  • Opinion phrase extraction reached 79% precision and 76% recall.
  • Polarity detection reached 86% precision and 89% recall.
  • Baselines were PMI++ and Hu++. PMI++ had slightly higher recall but lower precision.

Interesting ideas

• I1. Web hit counts as a validity check. OPINE scores each candidate by PMI with class-specific cue phrases such as "scanner has", then feeds the scores to Naive Bayes as binary features. This is a cheap way to filter noisy extractions in any domain where small corpora give weak statistics.
• I2. Ablations that attribute the gain. The authors run the new assessor on the baseline's own candidates and also toggle the Web statistics on and off. This separates exactly where the improvement comes from, which is a design worth copying when a new system changes several components at once.
• I3. Polarity in layers of context. Orientation is resolved for a word alone, then for the word with its feature, then within the specific sentence. This handles adjectives like "big" or "small", whose polarity flips depending on what they describe. Evaluating each layer separately shows how much context actually contributes.
• I4. Dependencies instead of a word window. Linking opinions to features through parser dependencies, rather than taking the k nearest words, gives cleaner feature–opinion pairs. The approach could carry over to any language that has a decent dependency parser.
• I5. Cheap cues to separate parts from properties. WordNet's IS-A hierarchy combined with suffixes like "-ity" or "-iness" distinguishes the two. Low-cost morphological signals of this kind are easy to port to morphologically rich languages.