📄 Beauty Before Age? Applying Subjectivity to Automatic English Adjective Ordering

Summary

English speakers strongly prefer certain orders for stacked adjectives before a noun ("long metal fork" sounds right, "metal long fork" does not). Earlier automatic systems learned these orders from observed adjective sequences, so they suffered from sparse data and domain shifts. Hill instead tests the linguistic claim that more speaker-dependent (subjective) modifiers sit farther from the noun. He takes the 197 most frequent BNC adjectives and computes six corpus-derived proxies for subjectivity from 24M words of the written BNC: modifiability by intensifiers, comparative/superlative use, predicative use after copulas, polarity, existence of an -ly adverb, and nominality. Gold orders come from Google n-gram bigram counts, using the 12,000 pairs with the clearest preference. A logistic regression over the two adjectives' feature profiles orders unseen pairs with 73.0% accuracy under 5-fold cross-validation. The result shows that a small set of general semantic features, with no direct evidence about the pairs themselves, carries real predictive power. It also suggests that combining such features with direct-evidence methods could make ordering systems more robust.

Key points

• Data: 19,306 unordered pairs were built from 197 BNC adjectives (other, only and very were removed). The 12,000 pairs with the largest proportional frequency gap in the Google n-gram Corpus were kept and balanced so that half were preferred orders and half dispreferred.
• Model: logistic regression on concatenated feature vectors, with an 80:20 split and 5-fold cross-validation. Overall test accuracy was 73.0%.
• Accuracy depends on preference strength. It reaches 86.2% on the 3,000 pairs with the strongest preference and 70.1% on the 3,000 weakest pairs.
• Individual features:
  • Every feature except COMPARABILITY is a significant predictor in the joint model.
  • PREDICATIVITY is the best single feature (68.6% alone).
  • NOMINALITY has a negative coefficient, meaning more noun-like modifiers sit closer to the head.
• Sensitivity checks:
  • Extracting features from only BNC Section A (7M words) lowers accuracy to 71.4%.
  • Adding the direct-evidence feature LEFTTENDENCY raises it to 76.3%.
  • For reference, a reported n-gram baseline reaches 82.3% on Brown, but the setups are not directly comparable.

Interesting ideas

• I1. Each word gets a fixed profile of graded corpus statistics, and the classifier orders a pair by comparing the two profiles, so it generalizes to pairs never seen together.
• I2. Abstract semantic properties are approximated with cheap syntactic tests: following an intensifier signals gradability, following a copula signals predicative use, and having an -ly adverb signals evaluativeness.
• I3. Gold labels are taken only from pairs with a clear frequency contrast, and accuracy is reported separately for strong and weak preference cases.
• I4. A deliberately feature-poor model fits even its own training data poorly, and the paper treats this as evidence that the feature set is incomplete rather than as a tuning problem.
• I5. According to the paper, these subjectivity features might also help determine the internal bracketing of nominal groups, for example telling apart the structure of "wild animal hunt" and that of "wild birthday party".

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (back): Generation could order stacked modifiers in noun phrases from per-adjective profiles stored in the lexicon, with each choice explainable through named features.
• I2 (world): Compiled adjective knowledge could hold gradability and predicativity scores counted over Mova's own UD-parsed corpora (copula and advmod relations instead of bigrams), giving graded properties for states and measures.
• I3 (learn): Silver data could keep only items where the evidence is lopsided, and evaluation could be split by confidence band to check that marginal cases do not hide failures.
• I4 (learn): When a rule or feature set underperforms on training items, Mova could flag a missing feature instead of adding weights, keeping models small and interpretable.
• I5 (tree): The parser could use adjective subjectivity or nominality scores to decide whether a modifier attaches to the head noun or to an inner nominal modifier inside compound noun phrases.