What’s scarier? Big bad wolf or bad big wolf? The three little pigs and little red riding hood and her grandma probably wouldn’t be bothered with the distinction. But there’s a natural order to language in which “bad big wolf” doesn’t sound quite right. 

The English language has various grammatical rules that speakers follow to communicate effectively. When referring to the past, we attach the suffix “-d” to the end of a verb. The rule is easy to follow and applies to a large number of verbs. Here are some examples:

• save –> saved
• start –> started
• hunt –> hunted
• play –> played 

When an adjective comes before a noun, prenominal adjectives, the ordering rules are not as straightforward. For example, English language speakers prefer a beautiful red car instead of a red beautiful car. Why do we have that preference? 

One grammarian, Sweet, proposed two related principles of prenominal adjective order: 

1. the definiteness of denotation  
2. the closeness of adjective/noun in meaning 

According to these principles, an adjective furthest from the noun denotes that noun’s most independent property. That’s why we prefer a beautiful red car because the car’s beauty is subjective, but its redness is a matter of fact. 

Sweet’s principles have received empirical support, but recent attempts to explain prenominal adjective order have been more specific about the semantic categories to which adjectives belong. Dixon proposed we order prenominal adjectives this way: 

value ->dimension->physical property->speed->human propensity->age->color->noun 

If a dog had the following characteristics, 

• old
• heavy
• wonderful
• slow
• brown
• loving
• large

We should prefer the wonderful, large, heavy, slow, loving, old, brown dog to any other order. 

Dixon’s category order is generally consistent with Sweet’s simpler principles, both of which describe how we order adjectives. But, why do we follow that order?

Chris Westbury (pictured below) introduced a new theory of prenominal adjective order in a paper published in Psychonomic Society journal, Psychonomic Bulletin & Review.

According to this theory, we order adjectives based on likely need. For example, we more likely need the word soft than the word tillable. Soft has broad, general application, and the tillable is much more specific. We, therefore, prefer soft tillable soil to tillable soft soil. According to Westbury, this is a psychologically plausible explanation of prenominal adjective order because it builds on a broad range of previous theoretical and empirical work in psychology.

Westbury tested this theory by first gathering 18,015 distinct adjective-adjective-noun triplets from the Corpus of Contemporary American English (COCA). A simple three-layer neural network then transformed each adjective into a mathematical vector of elements. Essentially, each vector was a mathematical representation of an adjective. Adjectives such as good, cool, excellent, superb, and great each have their own vectors. Because these adjectives have similar meanings, their vectors are correlated. If we take the average of these vectors, then that average vector would best define the concept of goodness. The figure below shows the correlation relationship between 282 adjectives. The adjectives related to meaning form clusters and the large gray text near the clusters indicate their meaning. 

Westbury measured likely need by calculating a category-defining vector (CDV) and measuring the distance between an adjective’s vector and the CDV. To compute the CDV, Westbury took the 1,000 most common adjectives from Google’s multi-billion-word news corpus, created a vector for each word, and then averaged those vectors together. If an adjective’s vector is similar to the CDV, then that adjective has a high likely need. If an adjective’s vector is dissimilar to the CDV, then that adjective has a low likely need.

With the distance-to-the-CDV measure for the 18 thousand adjective-adjective-noun triplets in hand, Westbury built a classifier and trained it using cross-validation on one-third of the data (6,005 distinct adjective-adjective-noun triplets). The classifier was then tested on two-thirds of the data (the remaining 12,010 triplets). Crucially, Westbury flipped the adjective-adjective ordering in half of the test data while keeping the other half as is (i.e., the observed ordering found in the COCA). The classifier accurately detected whether an adjective pair was flipped or not 72% of the time. This result suggests that the distance-to-the-CDV measure, a proxy for likely need, is predictive of prenominal adjective order. 

Westbury then conducted an experiment where participants judged whether prenominal adjectives ordered based on their distance-to-the-CDV were correct or incorrect. Every participant made decisions about 250 phrases. In each trial, one phrase, which ordered the adjectives by likely need, appeared on top (e.g., big fat dog) while the phrase with the reverse adjective ordering appeared below (e.g., fat big dog). 

The predictors in the linear mixed-effects model included the likely need estimates from the classifier discussed above and a random intercept for participants (because each participant made multiple responses). When the adjective ordering elicited a high likely need estimate, the model predicted participants to have a strong preference for that order (see the figure below). Similarly, when the adjective ordering elicited a low likely need estimate (i.e., when the adjectives were not ordered according to likely need), the model predicted participants to have a weak preference for that order. Thus, people typically preferred phrases with prenominal adjectives ordered by likely need.

The likely need explanation of prenominal adjective ordering is an appealing and interesting theory that builds on a broad range of previous theoretical and empirical work in psychology. For that reason, it gives us insight into how we communicate with one another. 

Psychonomic Society article featured in this post:

Westbury, C. (2020). Prenominal adjective order is such a fat big deal because adjectives are ordered by likely need. Psychonomic Bulletin & Review, 1-17. doi: 10.3758/s13423-020-01769-w