Semantic Polarity of Adjectival Predicates in Online Reviews

Web users produce more and more documents expressing opinions. Because these have become important resources for customers and manufacturers, many have focused on them. Opinions are often expressed through adjectives with positive or negative semantic values. In extracting information from users’ opinion in online reviews, exact recognition of the semantic polarity of adjectives is one of the most important requirements. Since adjectives have different semantic orientations according to contexts, it is not satisfying to extract opinion information without considering the semantic and lexical relations between the adjectives and the feature nouns appropriate to a given domain. In this paper, we present a classification of adjectives by polarity, and we analyze adjectives that are undetermined in the absence of contexts. Our research should be useful for accurately predicting semantic orientations of opinion sentences, and should be taken into account before relying on an automatic methods.

💡 Research Summary

The paper addresses a fundamental challenge in opinion mining: determining the semantic polarity of adjectival predicates in online reviews. While many sentiment‑analysis systems treat adjectives as fixed “positive” or “negative” cues, the authors argue that this approach ignores the context‑dependence of many adjectives, especially when they modify different product features. To investigate this, they collected five domain‑specific corpora—cosmetic products, hotels, hospitals, mobile phones, and movies—each containing roughly 23,000–27,000 tokens of Korean user reviews. Using the Geuljabi lexical analyzer, they extracted all adjectival predicates (about 300 distinct types per domain) and the nouns they modify.

The central contribution is a two‑tier classification of adjectives: (1) Absolute Polarity adjectives whose sentiment is stable across domains (e.g., “good”, “beautiful”, “dangerous”), and (2) Relative Polarity adjectives whose sentiment flips depending on the feature they describe (e.g., “long”, “big”, “strong”). Their corpus analysis shows that about 41 % of adjectives belong to the absolute class, while 59 % are relative. This distribution implies that any keyword‑based sentiment lexicon that assumes fixed polarity will inevitably misclassify a large portion of opinion sentences.

To illustrate the problem, the authors present a case study with the Korean adjective “khuta” (“big”). In hotel reviews, “khuta” can appear in factual statements (“There are big buildings around the hotel”) as well as evaluative ones (“The hotel is big and wonderful”). Extraction experiments reveal that using “khuta” as a sentiment cue yields only 79 % precision for hotels and 77 % for mobile phones, with roughly 20 % of retrieved sentences being neutral facts. This demonstrates that relative adjectives generate substantial noise if context is ignored.

Recognizing that the polarity of a relative adjective is determined by the noun it modifies, the authors construct detailed feature lists for each domain. For cosmetics, features include color, scent, chemical ingredients, effects, physical symptoms, price, and design. Hotels are broken down into facilities, supplies, service, cleanliness, food, and surroundings. Similar taxonomies are built for hospitals, mobile phones, and movies. By pairing adjectives with these domain‑specific feature nouns, the system can disambiguate whether an adjective conveys a positive or negative opinion.

Building on this analysis, the paper proposes an “Opinion‑Feature Dictionary” that integrates three layers of information: (i) a lexicon of absolute‑polarity adjectives with fixed sentiment scores, (ii) a mapping of relative‑polarity adjectives to the feature nouns that trigger a positive or negative interpretation, and (iii) the hierarchical feature taxonomy for each domain. Such a resource enables sentiment‑analysis pipelines to move beyond simple bag‑of‑words and to perform context‑aware polarity assignment, thereby improving both precision and recall.

The authors conclude that a nuanced treatment of adjectival polarity—distinguishing absolute from relative adjectives and explicitly modeling adjective‑feature relations—is essential for accurate opinion mining. They suggest future work to expand the approach to additional domains, to refine the feature taxonomies, and to explore neural models that can learn adjective‑feature interactions automatically. Overall, the study provides a solid linguistic foundation for more reliable sentiment analysis in Korean and highlights methodological considerations that are applicable to other languages as well.