The red dress is cute: why subjective adjectives are more often predicative

2.1 (Noun-like) reference versus (verb-like) predication

It is often argued that the prototypical function of nouns is to refer, while the prototypical function of verbs is to predicate – to describe some event or state of a noun’s referent (Croft 1990; Hopper and Thompson 1985; Thompson 1989). From this standpoint, as previewed above, attributive adjectives aid in reference by modifying a noun, while predicative adjectives predicate like verbs.

These parallels to nouns versus verbs are argued to lead to differences the temporal duration of the property described by attributive versus predicative adjectives. Because nouns tend to describe temporally stable referents, attributive adjectives are claimed to often describe lasting properties based on their affinity to nouns; conversely, many verbs describe dynamic changes, so predicative adjectives are claimed to describe transient states based on their similarity to verbs (Bolinger 1967; Ferris 1993; Givón 1984; Hollmann 2021; Nelson 1976; Quirk et al. 1972; Saylor 2000; Thompson 1989). In the literature on the attributive/predicative distinction, such temporal dimensions constitute the most widely discussed explanatory factor.

At the level of adjective types, temporal properties are sometimes invoked to explain why certain adjectives beginning with the a- prefix (asleep, awake, afraid) strongly prefer to appear in predicative position: such adjectives describe transient properties, which favor predication rather than reference (Bolinger 1967; Schlüter 2008). In the child language literature (Blackwell 1998; Nelson 1976; Saylor 2000), it is found that adjectives such as big, little, tiny, new, old, good, bad, nice, and favorite – describing lasting properties – tend to be attributive in speech to and by children, while hungry, clean, dirty, and sorry – describing transient states – tend to be predicative. At the level of tokens, even the same adjective type is claimed to evoke different temporal contours when it is used as attributive versus predicative; for Wierzbicka (1986), her red cheeks suggests that someone’s cheeks are characteristically red, whereas her cheeks were red connotes a transient state.

Aiming to distill type-level distributional hypotheses, one might therefore predict that adjectives describing lasting properties should tend to be attributive, while those describing transient properties should tend to be predicative. In fact, however, introspective data reveals striking counterexamples to this claim. In the literature on generic sentences (Carlson 1977; Kratzer 1995), researchers distinguish stage-level predicates such as available (3a), which describe transient states, versus individual-level predicates such as altruistic (3b), which describe lasting characteristics. Crucially, predicative adjectives can describe individual-level properties (3b) as well as stage-level ones, showing that predicative adjectives do not always describe transient properties.

Conversely, attributive adjectives can also describe stage-level properties (4), showing that attributive adjectives do not always describe lasting states.

Moreover, while her cheeks were red may describe a transient state (Wierzbicka 1986), this inference depends as much on the subject noun cheeks (a person’s cheeks can change color suddenly) as on the predicative position of the adjective red. With other subject nouns, red is taken to describe a durable characteristic (5).

Example (5) also illustrates that the inferred temporal duration of a property can vary widely across tokens of the adjective type describing it. Along the same lines, Maienborn (2004) notes that blonde is individual-level when it describes a person’s stable natural hair color, but stage-level if it describes transient hair dye, showing that the distinction between stage-level and individual-level predicates cannot be drawn at the level of adjective types.

Although attributive versus predicative adjectives are often said to be distinguished by temporal duration, this paper does not attempt to test that prediction in corpus data because no data are available to label the temporal duration of adjectives in attributive versus predicative contexts.^[1] I turn instead to other dimensions of meaning that lend themselves better to type-level corpus exploration.

2.2 Discourse structure

As previewed above, attributive and predicative adjectives are distinguished by their contribution to the discourse (Bybee and Thompson 2022; Kaiser and Wang 2021; Thompson 1989). Predicative adjectives directly engage the Question Under Discussion (Roberts 2012 [1998]) by offering the sentence’s “main news” (also known as its foregrounded, at-issue content; Abbott 2000; Potts 2005; Simons et al. 2010), as in (6). In contrast (Kaiser and Wang 2021), the information offered by attributive adjectives is secondary (7), offering at most one part of a larger answer to the QUD.

In attributive position, an adjective may appear in a definite or indefinite noun phrase, typically dictated by whether its referent – and therefore also, perhaps, the properties attributed to it – is framed as discourse-old or discourse-new (Heim 1982; but see, e.g., Coppock and Beaver 2015 for an analysis in terms of uniqueness rather than familiarity). Assuming that definite determiners carry presuppositions (Strawson 1950), an attributive adjective in a definite noun phrase is backgrounded as part of the presupposition that the noun it modifies has a familiar referent. Indefinite noun phrases do not presuppose familiarity, but attributive adjectives within indefinite noun phrases are also more backgrounded than predicative adjectives because they do not themselves constitute the main point of the sentence.

Table 1 lays out this mapping between the status of information in discourse and the realization of adjectives presenting that information; one cell is empty because it would be redundant (Büring 2003; Stalnaker 1979) for discourse-old information to be presented as “main news.”

Of course, the definite determiner can be used for new information (Abbott 2000; Poesio and Vieira 1998), but signals that this information is to be treated (accommodated; Karttunen 1974; Lewis 1979; von Fintel 2008) as old. Usually, a speaker will ask hearers to accommodate information that is uncontroversial and unsurprising (my sister rather than my boa constrictor; Abbott 2000), about topics for which the hearer will defer to the speaker’s epistemic authority (Chemla 2008; Schlenker 2012): having a sister is unsurprising and people typically know whether they have one, so if a speaker uses an expression presupposing that they have a sister, the sister should be uncontroversially added to the Common Ground.

Since the attributive/predicative distinction depends on the discourse context, the distribution of attributive versus predicative adjectives varies across genres or topics with systematically different Common Grounds. Counting the total number of attributive versus predicative adjective tokens in conversational corpora, Englebretson (1997) argues that attributive adjectives are used when introducing new referents, explaining why over 60 % of adjective tokens are attributive in a bank meeting in which many different loan applicants are discussed in sequence. In contrast, predicative adjectives are used to predicate information about known referents, which is why over half of adjective tokens are predicative in a conversation between a couple discussing people they both already know. The amount of shared referents in the Common Ground, he says, dictates the distribution of attributive versus predicative adjectives.

The rate of attributive versus predicative adjective tokens also varies across the spoken versus written medium (Biber and Gray 2011; Biber et al. 1999, 2010; Bybee and Thompson 2022). Attributive adjectives are far more common in writing than in speech, perhaps because writing is overall more information-dense. Writing and revision allows the author to add more information to a sentence than they could formulate in real time. Attributive adjectives have been gaining ground diachronically (Biber and Gray 2011; Biber et al. 2010), perhaps as society has become more informationally complex or as word-processing tools have made it easier to revise writing. When the Common Ground between a writer and their many hypothetical readers is unclear, writers may exploit that uncertainty by requiring readers to accommodate presuppositions, efficiently packing multiple propositions into a single sentence (Abbott 2000).

In sum, we might expect an adjective to favor attributive position if it describes information that is discourse-old or uncontroversial and thus easily accommodated. An adjective should favor predicative position in three overlapping cases: if the adjective describes information that is discourse-new and not easily accommodated; information that answers the Question Under Discussion (of course, informative answers are inherently discourse-new); or information that is potentially controversial, so that the speaker should offer it up for discussion rather than presuming that hearers will accommodate it.

To make distributional predictions about adjective types, one must elaborate how an adjective’s typical discourse contributions are grounded in meaning. For example, the claim explored above – that adjectives describing time-stable properties tend to be attributive, while those describing transient properties tend to be predicative – links information structure to meaning, in that we might expect time-stable properties to be discourse-old (and thus perhaps uncontroversial), while transient properties may be discourse-new. Similarly, Bolinger (1967) grounds an adjective’s discourse contribution in its meaning when he observes that the dented bell is a more felicitous description than the rung bell, because dents are part of the visual Common Ground and thus more helpful for reference. Predictions about the visual Common Ground are explored below, but first I introduce further dimensions of meaning that shape an adjective’s contribution to discourse.

2.3 Restrictive versus non-restrictive uses of attributive adjectives

Attributive adjectives (cute/red dress) modify nouns in two different ways.^[2] Prototypically, attributive adjectives “restrict” the denotation of the noun by carving out a proper subset of its referents: on its restrictive interpretation, the red dress distinguishes it from other (potential or actual; Martin 2014) dresses of other colors. In contrast, in the “non-restrictive” case (my lovely husband), the modified noun denotes the same referent as the unmodified version, without implying any non-lovely husbands.

Non-restrictive modifiers are not informative for reference, so they are claimed to be felicitous only when the modifier is relevant in some other way to the discourse (Leffel 2014; Martin 2014; Schlenker 2005): in (8a), the loveliness of my husband explains why I will bring him to dinner, while it is much harder to imagine why his height is worth mentioning. Tall can be nonrestrictive in a context where height is relevant (8b).

Because restrictive and non-restrictive interpretations are defined only for attributive adjectives, this distinction does not immediately yield predictions about which adjectives should tend to be attributive versus predicative. Such predictions only emerge, I argue, when this distinction is synthesized with insights about subjective adjectives, to which I turn next.

2.4 Subjectivity

Subjective adjectives such as cute, fun, and tasty (Lasersohn 2005) describe the speaker’s personal opinion and therefore allow for “faultless disagreement” (Kölbel 2004): the sense that neither party in (9) is wrong. With objective adjectives such as red, on the other hand, one of the two parties would arguably be factually incorrect.

As Vardomskaya (2018) explains, predicates are subjective when (because) there is no social consensus about what evidence would establish their truth. To use Vardomskaya’s examples, even the same predicate can be subjective in one context (Obama won the debate, meaning he performed well in the presidential debate) but objective in another (Sharon won the debate, meaning she has been declared the official winner of the World Debating Championship). Cute is thus more subjective than red because there is no consensus about what constitutes evidence for cuteness, while the optical basis for color is more widely accepted. Subjective adjectives also often carry emotional valence, emphasized in psychology (Osgood et al. 1957) and natural language processing (Pang and Lee 2008), so the study of subjective adjectives also advances the endeavor to bring affective meaning into semantics (e.g., Potts 2007) and corpus linguistics (e.g., Stubbs 1995).

For vague dimensional adjectives such as expensive (Kennedy 1999, 2007), the standard for what (cost) “counts as” expensive is subjective too, yielding faultless disagreement (10).

But Kennedy (2013) observes that vague dimensional adjectives such as expensive allow faultless disagreement in the positive form only, whereas personal taste predicates such as cute allow such disagreement in the comparative form too. In (11) but not (12), we sense that one party is factually wrong.

Therefore, while cute and expensive are both subjective, cute is even more so (Kaiser and Wang 2021). To capture these facts, expensive is taken to describe a subjective threshold along the objective scale of cost (Kennedy 1999, 2007), while cute is relativized to a judge parameter: cute according to j (Lasersohn 2005; Stephenson 2007; though see Barker 2013; Pearson 2013 for alternatives). Red is an objective, sortal predicate of individuals.

Along with semantics, subjectivity affects syntax and pragmatics. Scontras et al. (2017, 2019) and Hahn et al. (2018) use subjectivity to explain the ordering of multi-adjective attributive strings: why the cute red dress sounds better than ?the red cute dress. A referent is more efficiently identified, Scontras et al. (2019) argue, if the noun first composes with objective predicates (red) that reliably narrow down its denotation. Subjective adjectives (cute) are less useful for resolving reference because people may disagree about what they describe, so they appear further from the noun.

A person’s opinions are private until they verbalize them, so subjective adjectives typically describe discourse-new information and thus, when attributive, tend to appear in indefinite noun phrases in corpora (Vartiainen 2013). Vartiainen compares twenty frequent adjectives that he deems to be subjective (such as significant, interesting, useful, terrible, pleasant) to twenty more that he deems objective (industrial, military, professional, western, physical) in the British National Corpus (Leech et al. 2011), finding that the subjective ones occur more frequently in the indefinite part-of-speech-tagged string a(n) ADJ N-Sg while the objective ones favor the definite the ADJ N-Sg. Of course, the referent of an indefinite noun phrase is typically discourse-new along with the information described by the attributive adjective, so Vartianen’s results also show that subjective adjectives are more often used to elaborate new referents than to pick out old ones. This interpretation echoes the aforementioned claim that subjective adjectives are less useful for distinguishing among a set of familiar referents because they describe information that is private to the speaker. This paper aims to expand on Vartiainen’s finding.

Returning to the distinction between restrictive and non-restrictive attributive adjectives, subjective adjectives seem to favor the non-restrictive interpretation (Martin 2014; Umbach 2006). Perhaps cute and lovely are not effective for restricting a set of potential referents because, as discussed in the context of adjective ordering, people may disagree about what they apply to. Since non-restrictive adjectives are only felicitous when they are relevant for some reason, perhaps subjective adjectives are also suited to a non-restrictive interpretation because the speaker’s affective stance is easily accepted as relevant to whatever they are saying (Schlenker 2005). In Schlenker’s example (13), the objective adjective blond violates Grice’s quantity maxim (Grice 1989 [1975]) because it does not narrow the already-unique referent for president, but the subjective adjective stupid is informative because it contributes the speaker’s attitude.

Subjectivity also has consequences for the attributive/predicative distinction. Exploring how media consumers distinguish fact from opinion, Kaiser and Wang (2021) show that sentences containing subjective adjectives are rated by experimental participants as more subjective when the adjective appears as the predicate of the sentence (the orchestra was amazing), versus as an attributive (the amazing orchestra) or in a relative clause (the orchestra, which was amazing, …). This result is not surprising from the perspective of information structure: the whole sentence is rated as more subjective when the subjective adjective is framed syntactically as its central contribution.

Seeking to identify subjective adjectives for automatic sentiment analysis, Wiegand et al. (2013) show that adjectives occurring frequently in predicative position (brilliant) are more subjective (more likely to be listed in the sentiment lexicons of Taboada et al. 2011; Wilson et al. 2005) than those favoring attributive position (financial), but their paper does not explore the theoretical underpinnings or consequences of this insight.

On the one hand, we have established that predicative position is used for discourse-new information, and subjective adjectives are often discourse-new because a speaker’s personal opinions are inherently private (Vartiainen 2013). Moreover, we have seen that subjective adjectives are less useful for reference, because people may disagree on what they apply to, meaning that they are ill-suited to serve as restrictive attributives. These factors might contribute to the finding of Wiegand et al. (2013) that subjective adjectives favor predicative position.

On the other hand, new information can appear in attributive contexts as well as predication, distinguished by whether that information is framed as the main point versus secondary. So if subjective adjectives tend to appear in predicative position, one might need to explain why they are often framed as the main point.

That question in turn depends on the theory of subjective adjectives in discourse, discussed by Barker (2013), Malamud and Stephenson (2014), Pearson (2013), Stephenson (2007) among others. If Alice asserts (14a) or (14b) and others accept it as true, do they accept that Alice finds the dress cute or do they endorse it as cute from their standpoint also?

If cute is given a judge parameter (cute according to j), then the question is whether this judge parameter is fixed to the speaker or whether it is extended to other interlocutors if they accept (14a)–(14b) as Common Ground.

If (14a) just states that the dress is cute according to Alice, then this opinion should be uncontroversial because no one can challenge Alice on her own aesthetic tastes (Pearson 2013). Perhaps, then, this information would be equally suited to attributive position (14b), whether definite or indefinite, because we have seen that speakers often ask hearers to accommodate uncontroversial information rather than asserting it. This point is highlighted by Schlenker (2005) in a discussion of expressives (Potts 2005) – words such as damn and bastard that express a speaker’s emotional attitude – among which Schlenker includes subjective adjectives such as stupid. He says (Schlenker 2005: 392) that such words “invite accommodation” because “the speaker should know what his attitude towards an object is, and thus […] the addressee would be foolish to challenge such information.” If subjective adjectives invite accommodation, then perhaps they should be easily backgrounded as non-restrictive attributives.

On the other hand, if (14a)–(14b) ask hearers to agree that they also find the dress cute, then this opinion is potentially controversial (Gunglogson 2008), because hearers may disagree. For Stephenson (2007), (14a) asserts that Alice finds the dress cute; but if the assertion is accepted, then it becomes Common Ground that the dress is cute according to all interlocutors. This analysis explains not just why disagreements are faultless, but why disagreements arise in the first place: an assertion using a subjective adjective asks hearers to agree with the speaker’s opinion. From this standpoint, subjective adjectives are suited to predicative position because speakers should feel some pressure to propose controversial opinions for discussion as in (14a), rather than presuming that hearers should accommodate them as in (14b).

That pressure is presaged by Acton and Potts (2014) in a study of the 2008 United States Vice Presidential candidate Sarah Palin. Acton and Potts (2014) show that Palin frequently used definite and demonstrative noun phrases packed with subjective content, such as (15), thus presupposing a shared emotional outlook with her audience. They suggest that Palin’s speech drew vociferous hatred from her critics in part because critics felt manipulated by determiners carrying presuppositions that they rejected. Palin illustrates that speakers who background subjective opinions may annoy their audience with false presuppositions of agreement, suggesting that speakers may face some social pressure to foreground such information.

In sum, the type-level distinction between subjective and objective adjectives shapes their contribution to discourse, with consequences for their distribution as attributive versus predicative.

2.5 Predictions

This literature can be distilled into a series of corpus predictions.

When used in attributive contexts, subjective adjectives should favor indefinite Noun Phrases (16a) because subjective opinions are typically discourse-new (Vartiainen 2013) and not useful for distinguishing among familiar referents. Moreover, subjective adjectives should be more often predicative than objective adjectives (16b) (Wiegand et al. 2013). This finding would be consistent with claims that subjective opinions are discourse-new (Vartiainen 2013), in that predicative position is used for new information; and that they are not useful for reference (Scontras et al. 2017, 2019). It would also be consistent with the idea that subjective statements are controversial (Stephenson 2007) rather than self-fulfilling (Schlenker 2005), so that speakers may feel pressure to foreground such opinions rather than presuming that hearers will accommodate them. To preview, both of these predictions are manifested in the corpus studies reported below.

The next set of predictions concerns visual (color) adjectives such as red in two different discourse contexts: image captions for seeing people versus image descriptions for people with vision impairments (Kreiss et al. 2022). Just as subjective opinions are said to be discourse-new in general, visual adjectives are discourse-new in descriptions for people with vision impairments. We therefore predict that attributive visual adjectives should favor indefinite NPs in descriptions where they are discourse-new, more so than in captions where the visual context is Common Ground (17a). To foreshadow, this prediction is indeed manifested.

Our final prediction (17b) emerges from the idea that visual adjectives in descriptions work like subjective adjectives in being discourse-new and not useful for reference. If those factors are taken to explain why subjective adjectives are more often predicative (16b), then we should similarly expect visual adjectives to be more often predicative in descriptions versus captions (17b). On the other hand, if we assume that subjective adjectives tend to be predicative because they are controversial, then we no longer predict visual descriptions to pattern like subjective adjectives, because visual descriptions are often not controversial.

To preview, (17b) is not manifested. These results are consistent with an analysis whereby visual descriptions are different from subjective adjectives in terms of the epistemic authority of the speaker/writer. The writer holds full authority over the visual scene that they’re describing to a low-vision reader, so they are free to background visual information; but they do not have authority over whether readers/hearers will agree with their subjective opinions, so they feel some pressure to foreground such opinions for discussion.

These predictions synthesize insights from the literature about how the lexical semantics of an adjective type shapes the discourse contribution of its tokens, which in turn affects the syntactic realization of those tokens in attributive versus predicative position.

3.1 CoCA

This corpus study analyzes 64.7 million words from the Corpus of Contemporary American English between 2008 and 2012, using the genres of academic, newspapers, magazines, and spoken – chosen to keep the data to a manageable size.

Using the SpaCy dependency parser (Honnibal and Johnson 2015), the text was split into clauses. For each clause containing any words that CoCA tags as adjectives,^[4] attributive tokens were identified if they modified a noun using SpaCy’s amod dependency, or, to capture multi-adjective strings (cute red dress), if they were directly adjacent to another adjective that did so.^[5] The determiner of the larger noun phrase (if any) was also recorded. Predicative tokens were identified if they used the oprd, acomp, or attr dependencies. A token was excluded if the adjective appeared as the sentence’s root (Nice!), or if it modified a noun that was itself a predicate, such as that’s a great idea (Bybee and Thompson 2022), because such structures are both attributive and predicative at the same time. I also excluded appositives such as Taormina, the last Byzantine base in Sicily, was … (Doron 1992), and prepositional complements to verbs such as regard, view, take, see, look, feel, and sound (that seems like a simple thing), because these, too, blur the line between attributive and predicative uses. Overall, 219,547 tokens were excluded on these grounds.

Looking only at adjectives with human subjectivity ratings in the data of Futrell et al. (2020) and Dyer et al. (2023), the data comprise 1,983,532 adjective tokens (86 % of them attributive, 14 % predicative) of 690 adjective types. As expected, the percentage of predicative tokens is lowest in the academic genre (9 %), and highest in the spoken genre (24 %).

This token-level information was collated into a type-level dataframe which recorded, for each adjective, its percentage of attributive versus predicative tokens and the percentage of definite (the) versus indefinite (a, an, some) determiners on the noun phrases in which it occurs as an attributive modifier. For the purpose of labeling definiteness, I excluded bare nouns and only considered those with the determiners the, a/an, and some. For each adjective, I also gathered its per-million-word frequency as an adjective.

Using these data, I ran a linear regression in R (R Core Team 2024) to test the hypothesis (16a) from Vartiainen (2013) that subjective adjectives favor indefinite NPs when they are attributive. The independent variable is the adjective’s subjectivity rating in the annotations from Futrell et al. (2020) and Dyer et al. (2023); the dependent variable is the percentage of indefinite versus definite determiners on the noun phrases in which an adjective occurs as an attributive modifier. As predicted (Figure 1), more subjective adjectives are associated with a greater percentage of indefinite determiners when used as attributives – a substantial and highly significant effect (β = 0.44 on a 0-to-1 scale, t(684) = 10.95, p < 0.001, adjusted R ² = 0.15). (These results are replicated in a model that includes the adjective’s log-transformed per-million-word count as another additive predictor).

Figure 1:

Percentage of indefinite versus definite determiners on the noun phrases in which an adjective occurs as an attributive modifier (left), and percentage of predicative versus attributive tokens of adjectives (right), as a function of the adjective’s subjectivity rating, in CoCA from 2008 to 2012.

To illustrate, 77 % of attributive tokens of cute appear in indefinite noun phrases (18), compared to 40 % for red (19). These examples show both adjectives in both definite and indefinite noun phrases.

Next, I tested the hypothesis (16b) from Wiegand et al. (2013) that subjective adjectives are more often predicative. The independent variable is again the adjective’s subjectivity rating; the dependent variable is the percentage of predicative versus attributive tokens of the adjective. As predicted (Figure 1), more subjective adjectives are more often predicative – a large and highly significant effect (β = 0.53 on a 0-to-1 scale, t(688) = 15.02, p < 0.001, adjusted R ² = 0.25). (As above, these results are replicated if the log-transformed per-million-word count is included as an additive predictor).

For example, 61 % of tokens of cute are predicative (20), versus only 4 % for *red* (21). These examples show both attributive and predicative tokens for both adjectives.

3.2 Art catalog

The United States National Gallery of Art is a free museum in Washington, D.C. with an explicit “Open Data” policy welcoming people to use their web collection for research. Some works in this collection are associated with an “overview” (which I call a “caption”), a few paragraphs discussing the significance and historical context of the work (22); some also offer an “image description” for people with vision impairments (23). These descriptions follow published guidelines^[6] advising that they should aim to be specific and objective in communicating the elements of the visual scene.

These two types of text are parallel to the image captions and descriptions from Wikipedia compiled into the Concadia dataset of Kreiss et al. (2022), and I take inspiration from them in comparing language across paired texts discussing the same image with different pragmatic goals. But Concadia favors short sub-sentential phrases (e.g., grocery store photo of several bunches of bananas), so the National Gallery catalog was chosen because its full paragraphs offer more chances to find predicative adjectives.

I used Python and the Beautiful Soup html parser to scrape the catalog listings for sixty thousand art objects in the National Gallery catalog. In total, 338 Public Domain images possess both a caption/overview and a description, yielding a corpus of 182,909 total words (55 % from descriptions, 45 % from captions). This text was split into sentences using NLTK’s sentence tokenizer (Loper and Bird 2002), and then into clauses using the SpaCy dependency parser (Honnibal and Johnson 2015). The dependency parser was used to identify attributive versus predicative tokens of adjectives, again excluding attributive adjectives within predicative noun phrases (that’s a great idea; Bybee and Thompson 2022) as well as appositives and prepositional complements to verbs such as see, regard, sound, seem, and feel. These data were arranged into a type-level dataframe recording, for each adjective (in captions and in descriptions), its percentage of attributive versus predicative tokens; and the percentage of definite versus indefinite determiners on the noun phrases in which it occurs as an attributive modifier.

Captions and descriptions are compared separately in the study of visual adjectives below, but I look only at captions in this exploration of subjectivity because descriptions rarely contain subjective adjectives. Among the top twelve most subjective adjectives in the corpus (handsome, beautiful, happy, fantastic, charming, romantic, creative, best, special, amazing, pleasant, good), all of their tokens appear in captions rather than descriptions. In captions, focusing on adjectives that appear at least twice, there are 2,861 tokens of 268 adjective types for which subjectivity ratings are available, of which 90 % are attributive, 10 % predicative.

First, I tested the hypothesis (16a) from Vartiainen (2013) that subjective adjectives favor indefinite NPs when they are attributive. The independent variable is the adjective’s subjectivity rating from Futrell et al. (2020) and Dyer et al. (2023); the dependent variable is the percentage of indefinite versus definite determiners on the noun phrase in which an adjective occurs as an attributive modifier. As predicted (Figure 2), when used in attributive contexts, more subjective adjectives are associated with a greater percentage of indefinite determiners – a subtle but significant effect (β = 0.29 on a 0-to-1 scale, t(231) = 2.03, p < 0.05, adjusted R ² = 0.013).

Figure 2:

Percentage of indefinite versus definite determiners on the noun phrases in which an adjective occurs as an attributive modifier (left), and percentage of predicative versus attributive tokens of adjectives (right), as a function of the adjective’s subjectivity rating, in 81,692 words of National Gallery captions.

The colloquial word cute does not appear in the National Gallery data, so I use successful as an example of a subjective adjective, with a subjectivity rating of 0.73. For successful, 66 % of attributive tokens occur in indefinite noun phrases (24), compared to 56 % for red (25).

Turning to (16b), as predicted (Figure 2), more subjective adjectives are more often predicative (β = 0.16 on a 0-to-1 scale, t(266) = 2.75, p < 0.01, adjusted R ² = 0.024). The effect is smaller than the one found above for CoCA, perhaps because predicative tokens are so rare in the National Gallery, but still meaningful.

To illustrate, 20 % of tokens of successful are predicative (26), versus 0 % of tokens for red. (26) shows examples of successful in both predicative and attributive positions; above, (25) shows red in attributive position.

In sum, both predictions (16a)–(16b) are consistent with the data, both in CoCA and in the National Gallery of Art.