1 Introduction

Imagine the following scenario: a kindergarten group has to move to another room. The children are helping their teachers to pack up their toys in boxes, among them all kinds of balls including bouncy balls, basketballs, soccer balls, ping-pong balls, and tennis balls, as well as hopper balls, and many other toys such as marbles. In this context, one of the teachers may say Look at this one; this one is big, pointing to one of the balls. Then, pointing to a different ball, she may say Look at this one; this one is small. In order to evaluate whether these utterances are appropriate descriptions of the objects the teacher is pointing to, the child has to understand what her teacher uses as a basis of comparison for the adjectives. Put differently, the child needs to know the comparison class. Big and small are gradable adjectives and their interpretation crucially relies on knowledge of the comparison class. That is, judging an entity as big or small in a given context requires determining the threshold for big or small, which depends on the class of entities it belongs to and how it compares to the entities in this class (e.g., Cresswell 1976). Utterances such as This one is small/big differ from utterances like This one is small/big for a ball in that the latter indicate the comparison class overtly, via the for-PP. In the former case, the child needs to determine the comparison class herself: is the object big for a ball, big for toys, big for fitting in a box, etc.? Cues to the comparison class may be non-linguistic or linguistic in nature. Non-linguistic cues concern the visual context (e.g., presence of other toys in our scenario), most relevant for the current study, but they may also concern the discourse context (e.g., the current topic of the conversation like packing boxes) and world knowledge (e.g., the typical size of balls). Linguistic cues include the modified head noun, for instance balls in big balls. The child can employ the category label ball to restrict the comparison class to the set of balls. Accordingly, in order to fulfill a request such as Please hand me the big balls, the child’s task is to decide which of the balls in a specific situation are big and which ones are not. In the scenario sketched above, the child may opt for the soccer balls and basketballs and decide against the tennis balls, bouncy balls, and ping-pong balls. If the teacher in the same scenario says Please hand me the big toys, the child may choose the hopper balls and the basketballs and ignore the other balls, along with other small toys like marbles.

Adjectives such as big and small are referred to as relative gradable (henceforth RGA) and are distinguished from absolute gradable adjectives such as clean and dirty (henceforth AGA) (Kennedy & McNally 2005). The two classes of gradable adjectives differ regarding the role of the comparison class in that the comparison class seems to affect the threshold much less for AGAs than for RGAs. The requests Please hand me the clean balls and Please hand me the clean toys, for instance, will most likely result in picking out the same set of balls, whereas the requests Please hand me the big balls and Please hand me the big toys would not.

Capitalizing on these properties of RGAs and AGAs, acquisition research has asked whether children are sensitive to the distinction between RGAs and AGAs. Most experimental studies have focused on the role of non-linguistic cues for the interpretation of RGAs and have examined which contextual information children use for calculating the threshold. Here, we extend this line of research and investigate the role of linguistic cues for children’s interpretation of RGAs, comparing them to non-linguistic cues and to the role of both types of cues for AGAs. More specifically, we investigate whether the threshold shifts as a function of changes in the taxonomic category encoded by the modified head noun (basic-level vs. superordinate-level, henceforth called ‘taxonomic label’), and we compare the effect of this linguistic change with the effect of a non-linguistic change involving the object categories present in the visual context. German-speaking preschoolers and adults participated in a picture-choice task targeting the comprehension of the RGAs groß ‘big’ and klein ‘small’ and the AGAs sauber ‘clean’ and dreckig ‘dirty’ in different linguistic and visual contexts. This experimental design allows us to uncover the role of taxonomic labels relative to the role of visual contextual cues in determining the comparison class and in calculating the threshold. It also allows us to seek further support for the finding that children distinguish RGAs and AGAs. By comparing child and adult data, we explore whether children and adults differ in the cognitive representation of the comparison class (henceforth abbreviated as CC).

The paper is structured as follows: Section 2 summarizes the theoretical analysis of RGAs and AGAs and spells out the acquisition task for the child. Section 3 reviews previous research on children’s interpretation of RGAs and AGAs. Our study is presented in Section 4, the findings are discussed in Section 5. Section 6 offers a conclusion.

2 Gradable adjectives

Gradable adjectives are commonly analyzed within degree-based frameworks (e.g., Cresswell 1976; von Stechow 1984; Kennedy 2007).¹ According to this analysis, gradable adjectives denote relations between individuals and degrees on a scale associated with some dimension of measurement such as height (type <d,<e,t>>); this is illustrated in (1) for big. Degrees are understood as representations of measurement; a scale is a set of degrees that is totally ordered regarding a dimension such as height. In (1), big is a measure function that maps its argument onto the height scale (type <e,d>).²

To evaluate whether the property denoted by a gradable adjective is true of an individual, the individual must be related to a degree that exceeds a certain threshold to a sufficiently salient extent (Kennedy 2007). The unmarked positive form is assumed to combine with a null degree morpheme (pos) that binds the degree variable and introduces the threshold. Kennedy’s (2007) implementation of pos is given in (2):

There is agreement that context typically affects the threshold for RGAs but not for AGAs. Both adjective classes are discussed in turn in Section 2.1 and 2.2. In Section 2.3, we sketch the child’s acquisition task, given the theoretical analysis of gradable adjectives.

3 Previous research on children’s interpretation of gradable adjectives

Empirical data is crucial for understanding the nature of the CC: experimental studies can inform us about which contextual cues are actually used to calculate the threshold and can inform theoretical accounts that to date do not spell out how exactly the context provides the threshold. In Section 3.1 and 3.2, we summarize previous research on gradable adjectives addressing non-linguistic cues and linguistic cues, respectively. RGA studies are discussed regarding the effects of different contextual cues for children’s interpretations; AGA studies are discussed regarding the question of whether children and adults distinguish them from RGAs. Note that adults are asked directly about which CC they have in mind in a specific context (e.g., Tessler & Goodman 2022), while children’s interpretation of gradable adjectives is inferred from their responses to questions or from their object choices.

4 The present study

4.3 Methods

4.3.1 Materials

Four easy-to-depict adjectives were tested: groß ‘big’, klein ‘small’, sauber ‘clean’, and dreckig ‘dirty’. We included a positive RGA (big) and its negative counterpart (small) to control for the role of adjective polarity (see B&S; Pagliarini et al. 2022). A maximum AGA (clean) and its minimum counterpart (dirty) were included to examine whether interpretation patterns differ for minimum and maximum AGAs (see Syrett et al. 2010).

Participants saw sets of picture cards; each picture card had the same size (14 x 14 cm) and depicted one object from the category ‘toys’.⁵ The object category ‘toys’ was chosen for all adjectives: as toys often come in different sizes, corresponding to different extents of smallness or bigness, we could create natural transitions between objects. In addition, toys can clearly be depicted as dirty or clean. Importantly, the toys we used as well as their names were likely to be familiar to preschool children.

Each object was of a different color, except for the soccer balls, and showed the property described by the respective adjective to a different extent. All test prompts had the form Please hand me the ADJ_Plural N_Plural, with neutral, non-contrastive intonation.

As shown in Figure 1 for RGAs and in Figure 2 for AGAs, each adjective occurred in three conditions, with two items per condition. In all three conditions, participants received linguistic information – via the taxonomic label in the test prompt – and non-linguistic information – via the objects in the visual array – regarding the possible CC. What differed across conditions was the combination of the specific test prompt and the specific visual array. By comparing participants’ thresholds across conditions, we could detect which contextual changes were most prominent in inducing a threshold shift: changes in the taxonomic label (keeping the visual array constant), changes in the visual array (keeping the taxonomic label constant), or changes of both, the taxonomic label and the visual array.

Figure 1: Example items for the RGA big. Conditions: a. basic_baseline, b. basic_expansion, c. toys_expansion. The numbers were not present in the experiment; they are added for easier reference to the objects in the remainder of the paper.

Figure 2: Example items for the AGA dirty. Conditions: a. basic_baseline, b. basic_expansion, c. toys_expansion. The numbers were not present in the experiment; they are added for easier reference to the objects in the remainder of the paper.

In Condition 1 (basic_baseline⁶, Figure 1a/2a), the visual setup consisted of a set of eight toys from one of four basic-level categories: water balloons or hopper balls, all varying in size, and buckets or balls, all varying in dirtiness. We refer to them with numbers 5 to 12, with object 5 being the smallest/dirtiest object, and object 12 being the biggest/cleanest object. The four additional smaller/dirtier objects in basic_expansion and toys_expansion are labeled 1 to 4; the four additional bigger/clean objects 13 to 16. In the basic_baseline test prompts, the adjectives modified a basic-level head noun (Wasserbomben ‘water ballons’, Hüpfbälle ‘hopper balls’, Teddys ‘teddy bears’, Bälle ‘balls’).

In Condition 2 (basic_expansion, Figure 1b/2b), children saw 12 toys from two different basic-level categories: the eight original objects of the basic_baseline condition and four toys from a different basic-level category (soccer balls, buckets, or dolls). The additional objects expanded the object array either at the upper end of the scale, i.e., by big (Figure 1) or clean objects, or at the lower end, i.e., by small or dirty (Figure 2) objects. By adding additional objects from a second basic-level category of toys, the distribution of the adjectival property changed compared to the basic_baseline condition. The noun in the test prompt was a basic-level noun, identical to the basic_baseline condition.

In Condition 3 (toys_expansion, Figure 1c/2c), the visual setup was identical to the basic_expansion condition, but in the test prompts, the adjectives modified the superordinate-level noun toys.

In short, basic_baseline differed from toys_expansion regarding the visual setup and the taxonomic category encoded by the head noun. Basic_baseline differed from basic_expansion only regarding the visual setup, and basic_expansion differed from toys_expansion only regarding the taxonomic category encoded by the head noun.

The visual setups and the nouns used were the same for big and small and for clean and dirty, respectively. This design allowed us to compare children’s responses for adjectives of different polarity in the same visual context and in the same child. Table 2 provides the complete list of test items.

Table 2: Full list of test items.

Condition	Item	Noun in test prompt	Visual setup
basic_baseline	big-1/small-1	water balloons	8 water balloons of different size
	big-2/small-2	hopper balls	8 hopper balls of different size
	clean-1/dirty-1	teddies	8 teddies of different dirtiness
	clean-2/dirty-2	balls	8 balls of different dirtiness
basic_expansion	big-1/small-1	water balloons	8 water balloons + 4 bigger soccer balls
	big-2/small-2	hopper balls	8 hopper balls + 4 smaller buckets
	clean-1/dirty-1	teddies	8 teddies + 4 clean buckets
	clean-2/dirty-2	balls	8 balls + 4 dirtier buckets
toys_expansion	big-1/small-1	toys	8 water balloons + 4 bigger soccer balls
	big-2/small-2	toys	8 hopper balls + 4 smaller buckets
	clean-1/dirty-1	toys	8 teddies + 4 clean buckets
	clean-2/dirty-2	toys	8 balls + 4 dirtier buckets

• Note. There are two items per condition for each adjective, referred to by –1 and –2.

Ten filler trials were added in-between test trials to minimize potential influence from the prior test item. Each filler trial consisted of eight picture cards (see Figure 3). Similar to the test trials, the picture cards displayed toys. In each trial, objects from two different basic-level categories were shown (buckets, dices, Lego® bricks, soccer balls, books) which differed in shape (round or square) and color (red or blue). The fillers were similar to the test trials, but had different adjectives (red, blue, round, square) and were uttered by the child (see Section 4.3.2 for the procedure).

Figure 3: Example filler trial.

The test phase was preceded by a practice phase, where the children were introduced to the objects and their names and to the task. The experimenter first showed pictures of single exemplars of the test objects and asked the child to label them. Then, the child saw all exemplars together and was asked And what are all these? to prompt the superordinate-level noun toys. Next, the children received three practice trials, where they saw four objects from different basic-level categories and heard a test prompt, which did not contain any adjective (e.g., Please hand me the dolls). Following B&S, we opted for plural definite DPs in the practice phase and in the test trials. The definite plural DP presupposes the existence of objects and signals reference to the maximal element of a given set (Link 1983), which can be plural individuals, or in singleton sets, atomic individuals. We assured participants that they could select as many or as few objects as matched the description, and that choosing one object was a licit choice. Participants also received feedback in the practice phase, if they noticed that only one object showed the property: the experimenter explained that the puppet’s request was the same independent of the number of objects matching the description. If, in the practice phase, the child forgot to select one or some required objects, the experimenter gave feedback as well.

4.3.2 Procedure

The study was carried out in two sessions (session 1: big and clean; session 2: small and dirty). Visual stimuli and order of presentation were the same across sessions to allow for a comparison of the antonymous adjectives. To minimize possible influence from choices in session 1 on performance in session 2, the two sessions were about twelve days apart.

To make the task engaging for the child, we implemented the experiment as a game: a puppet played by the experimenter and the child took turns selecting picture cards. The children sat next to the experimenter on the floor or at a table large enough to show all picture cards at the same time. At the beginning, the experimenter explained that the puppet wanted to play a game. The puppet and the child each received a die, with arbitrary patterns for the puppet and with a square, a circle, a red dot, and a blue dot for the child. In the test trials, the puppet made a request and the child had to select the matching picture cards; in the filler trials, the roles were reversed. At the beginning of each trial, the experimenter distributed the picture cards. Importantly, the picture cards were presented in random order (see Figure 4). This mode of presentation does not suggest a specific relation between the objects to the participants. Instead, participants need to establish their own ordering, silently, verbally, or by moving the cards, before they can respond to requests such as Please hand me the big water balloons. We believe that this is a close proxy to real-world scenarios, since relationships between objects in reality are rarely as transparent as they appear in an ordered series of objects.

Figure 4: Example presentation of visual stimuli.

On the puppet’s turns, the puppet rolled her die without letting the child see the outcome and made her request, which corresponded to the test trials. In the child’s turns, the child rolled her die. When the die showed blue, for example, the child had to ask the puppet to hand her the blue toys. The puppet sometimes chose the correct and sometimes the incorrect items. The participant’s (and the puppet’s) task was to select those objects that in their opinion matched the request; the experimenter emphasized that the person selecting the right picture cards more often would win the game.

4.3.3 Data analysis and expected responses

4.3.3.1 Data analysis

Numbers are used to refer to the objects. Recall that the eight objects in basic_baseline are labeled 5 to 12, with object 5 being the smallest/dirtiest object, and object 12 being the biggest/cleanest object. The four additional smaller/dirtier objects in basic_expansion and toys_expansion are labeled 1 to 4; the four additional bigger/clean objects 13 to 16.

Participants’ responses were analyzed in three steps. First, for each trial we analyzed which objects were selected. In a next step we determined for each trial the participants’ individual ‘cutoff point’ (henceforth: cutoff), which marks the rank of the transition between entities for which an adjectival property is true and for which it is not. For big this was the smallest object considered as being big (together with all bigger objects), for small it was the biggest object considered as being small (together with all smaller objects). For clean it was the dirtiest object considered as being clean (together with all cleaner objects), and for dirty it was the cleanest object considered as being dirty (together with all dirtier objects). To illustrate, if a child selected object 10, 11, and 12 when asked for big water balloons in the scenario of Figure 1, the cutoff is 10, marking the rank of transition between big and non-big objects. If participants missed an object within a series after the cutoff, this response was classified as unanalyzable.⁷ Proportion of choices for each object in an array and the cutoff were taken to express participants’ threshold.

In the third step of the analysis, we compared the cutoffs of each condition in a pairwise manner (basic_baseline – toys_expansion, basic_baseline – basic_expansion, basic_expansion – toys_expansion) for each item (big-1, big-2 etc.; see Table 2) and analyzed whether the cutoffs shifted. Unlike B&S, who compared mean cutoffs, our within-subjects design allowed us to identify individual shifts in the cutoff. This way we could uncover which of the contextual changes (visual setup + taxonomic label, taxonomic label, or visual setup) more often caused participants to shift their thresholds. Consequently, we can answer the question of whether the CC is preferably determined by linguistic or by non-linguistic visual cues.

Based on the third step of the analysis, a dependent variable cutoff shift was created with values classified as yes, no, other, or undefined. We illustrate each type with example cutoffs for item big-1 and the contextual change ‘visual setup + taxonomic label’, i.e. a pairwise comparison of the conditions basic_baseline and toys_expansion (see Figure 1a and 1c): a cutoff shift was present, if the cutoff was shifted in the expected direction given the distribution of size in the two conditions, e.g., when the cutoff is 9 in basic_baseline and 11 in toys_expansion (the cutoff in toys_expansion should be higher because bigger objects were added).⁸ No cutoff shift occurred when the cutoffs were identical across conditions, e.g., 9 in basic_baseline and in toys_expansion. Other responses were those that showed a shift, but given the distribution of size in the two conditions, in the unexpected direction, e.g., cutoff 9 in basic_baseline and cutoff 8 in toys_expansion. A shift was coded as undefined when the cutoff for one of the conditions in the comparison was unanalyzable (see above).

4.3.3.2 Expected responses

To address question (Q1), we considered participants’ cutoffs in the basic_baseline trials. Under the assumption that AGAs describe the maximal or minimal value of a property, we expected the threshold for clean to be the maximal lack of dirtiness (= object 12) and the threshold for dirty the minimal existence of dirtiness (= object 11). For the RGAs big and small, we expected thresholds around the center of the object series.

To answer research question (Q2), we analyzed to what extent the cutoff shifted according to a contextual change (visual setup, taxonomic label, visual setup + taxonomic label). Following the assumption that the interpretation of AGAs is not context-dependent (modulo cases of potential deviance from a minimal or maximal threshold, see Section 2.2), we expected the same cutoff across all three conditions for AGAs. The cutoffs of RGAs should shift between conditions, as their interpretation is context-dependent.

As for our main research question (Q3), we compared the frequency of cutoff shifts for each type of contextual change (visual setup + taxonomic label, taxonomic label, visual setup). For RGAs, we expect a shift of the cutoff when both, visual setup and taxonomic label, change, i.e., we contrast basic_baseline and toys_expansion. That is, participants’ object choices should differ in basic_baseline and in toys_expansion: for example, participants may select water balloons 10, 11, 12 in basic_baseline (Figure 1a) and only water balloon 12 together with all the bigger soccer balls in toys_expansion (Figure 1c). We expect the same number of cutoff shifts when only the taxonomic category of the noun changes (i.e., contrasting basic_expansion, Figure 1b, and toys_expansion, Figure 1c), and we predict significantly fewer cutoff shifts when only the visual setup changes (i.e., contrasting basic_baseline and basic_expansion), under the assumption that comparison classes are preferably based on linguistic cues. In the latter case, participants should select the same objects, e.g. water balloons, in basic_baseline (Figure 1a) and basic_expansion (Figure 1b), if they determine the CC based on the taxonomic label of the noun and ignore the visual presence of the bigger soccer balls. In case the cutoff is shifted, it should be higher when objects are added at the upper end of the object array, and it should be lower when objects are added at the lower end of the object array.

For AGAs, we do not expect participants to shift the threshold: for instance, participants should consider the same balls dirty across conditions (Figure 2a–c) and should select the additional dirty buckets in toys_expansion (Figure 2c).

To answer (Q4), we analyzed the adult data in a parallel way. Given that adults are proficient speakers with rich linguistic and non-linguistic knowledge, it may be that they use the different cues to the CC more flexibly than children do.

4.4 Results

The analysis of the thresholds in the basic_baseline condition is reported in Section 4.4.1, addressing (Q1). The analyses of the cutoff shifts, for children and for adults, are reported in Section 4.4.2, addressing (Q2), (Q3) and (Q4).

4.4.1 Thresholds in basic_baseline

We first analyzed children’s individual object choices. Figure 5 displays how often children selected the objects 5 to 12 in the AGA trials (for the numbering see Figure 2a). In the two clean trials, the object without dirty spots (object 12) was selected in almost all cases (with one exception by a 3-year-old). Choices for objects with some dirty spots are attributable to a few individual children: four children (from all age groups) sometimes also treated object 11, with one spot of dirt, as being clean; and six children (aged 3 and 4) selected all objects in one or both of the clean trials. In the two dirty trials, the objects with at least one dirty spot (object 5–11) were selected very frequently. Two children (aged 3 and 5) did not select all objects with dirt; one child in each age group selected all objects as dirty, including object 12, which had no dirty spots.

Figure 5: Children’s object choices for clean and dirty in basic_baseline (5 = dirtiest object, 12 = cleanest object).

Figure 6 displays how often children selected the objects 5 to 12 in the RGA trials (for the numbering see Figure 1a). In the two big trials, the biggest object was selected in all cases except one (by a 3-year-old). The proportion of choices increased around the center of the scale. In the two small trials, the smallest object was selected in all cases except one (by a 3-year-old); the proportion of choices decreased around the center of the scale.

Figure 6: Children’s object choices for big and small in basic_baseline (5 = smallest object, 12 = biggest object).

We now turn to the analysis of the cutoffs in basic_baseline, which is based on 86 observations per adjective (two trials per child). Table 3 lists for each of the adjectives the median cutoff together with the cutoff that occurred most often (mode) and the range.

Table 3: Children’s cutoff: median, mode, minimum, and maximum per adjective.

Adjective	Median	Mode	Minimum	Maximum
Big	9	9	5	12
Small	7	6	5	10
Clean	12	12	5	12
Dirty	11	11	5	12

We analyzed whether children’s behavior differed depending on the age group. Notably, the cutoffs did not differ between the three-, four- and five-year-olds (Kruskal-Wallis test; clean: Χ²(2) = 4.018, p = .134; dirty: Χ²(2) = 2.878, p = .237; big: Χ²(2) = 4.709, p = .095, small: Χ² (2) = 1.411, p = .494). Within each age group, the cutoffs for clean and dirty differed significantly (Wilcoxon test; 3-year-olds: z = –3.638, p < .001, r = 0.6, n = 18; 4-year-olds: z = –3.170, p = .002, r = 0.4, n = 29; 5-year-olds: z = –5.568, p < .001, r = 0.7, n = 33), with strong effect sizes for the 3- and 5-year olds and a moderate effect size for the 4-year-olds. The cutoffs for big and small differed significantly within each age group as well (Wilcoxon test; 3-year-olds: z = –3.638, p < .001, r = 0.6, n = 18; 4-year-olds: z = –3.170, p = .002, r = 0.4, n = 29; 5-year-olds: z = –5.568, p < .001, r = 0.7, n = 33), with strong effect sizes for the 3- and 5-year olds and a moderate effect size for the 4-year-olds.

Overall, the proportion of object choices and the resulting cutoff points in the basic_baseline condition indicate a threshold around the center of the scale for big and small, a maximal threshold for clean, and a minimal threshold for dirty. These results are very similar to the adult data; a detailed description is provided as Supplementary Material.

4.4.2 Shift of cutoff points

For our main analysis, we examined whether children’s and adults’ cutoffs were sensitive to changes in the context. Comparing the cutoffs across the three conditions (basic_baseline, basic_expansion, toys_expansion) in a pairwise manner, we investigated which of the contextual changes (visual setup + taxonomic label, taxonomic label, visual setup) trigger a shift of the cutoff. We first report the results for the children and then the results for the adults.

4.4.2.1 Children

For each item and each adjective, we analyzed three contextual changes (visual setup + taxonomic label, taxonomic label, visual setup). In total, there are 1,032 observations, 516 for AGAs (clean, dirty) and 516 for RGAs (big, small). Recall that shifts of the cutoff are expected for RGAs, but not for AGAs. As far as AGAs are concerned, the cutoff was shifted in only 5.8% (30/516) of the cases. In 79.1% (408/516) of the cases the cutoff was not shifted; 13.3% (69/516) of the responses were classified as undefined, because the cutoff in one of the conditions could not be determined (see footnote 7); finally, 1.7% (9/516) of the responses were other shifts. As for RGAs, the cutoff was shifted in 44.8% (231/516) of the cases. In 23.6% (122/516) of the cases the cutoff was not shifted; 15.1% (78/516) of the responses were shifts in the unexpected direction, and 16.5% (85/516) of the cases were classified as undefined. Undefined responses were excluded from further analyses.

Recall that for RGAs we expected different shifting patterns depending on the type of contextual change. A shift of the threshold should occur when both the taxonomic category encoded by the noun and the properties of the visual setup changed. Under the assumption that the CC is preferably determined based on linguistic cues and that the threshold is calculated accordingly, a shift of the cutoff should occur when only the taxonomic category encoded by the noun changed, but less often when only the visual setup changed. Therefore, we inspected the shifts for RGAs for each type of contextual change. The results are shown in Figure 7; the width of the violin plots varies by the density of data points in a specific region. The frequency of shifts varied with type of contextual change. The violins for two of the three conditions (visual setup + taxonomic label, taxonomic label) for both big and small resemble each other and differ from the violins for the visual setup condition. These latter show wider parts in the middle and in the lower part, indicating that the majority of children rarely shifted the threshold; they did so in about half of the cases or almost never (big: mean = 33.3%, SD = 36.8%; small: 27.3%, SD = 29.8%). The former violins show wider parts in the middle and in the upper part, indicating that children shifted the threshold more frequently, in about half of the cases or nearly always. Children shifted the threshold for the RGAs big and small most frequently, when both the taxonomic label and the visual setup changed (big: mean = 64.9%, SD = 37.0%; small: mean = 70.0%, SD = 30.2%). Descriptively, the threshold shifted slightly less often when only the taxonomic label changed (big: mean = 58.6%, SD = 37.3%; small: mean = 64.0%, SD = 30.7%).

Figure 7: Frequency of cutoff shifts for the relative gradable adjectives by adjective and type of contextual change in the child group. The black line indicates the mean.

The shifting data for RGAs were analyzed with generalized mixed effects logistic regression (1 = shift-yes, 0 = shift-no or shift-other); we used the glmer function in the lme4 package (Bates et al. 2015) in R (R Core Team 2022). The shifting data were fitted into a model using maximum-likelihood estimation (Laplace Approximation). Contextual change (visual setup + taxonomic label, taxonomic label, visual setup) was entered as a fixed effect, using treatment coding and visual setup + taxonomic label as the reference level, with a random intercept for Participant.⁹ The model revealed a highly significant difference between visual setup + taxonomic label and visual setup (β = –1.4904, Z = –5.774, p < .001), but not between visual setup + taxonomic label and taxonomic label (β = –0.3598, Z = –1.456, p = .145). Further inspection via pairwise comparisons with Tukey adjustment (emmeans package, Lenth 2022), showed a significant difference between visual setup + taxonomic label and visual setup (p < .001), and between visual setup and taxonomic label (p < .001), but not between visual setup + taxonomic label and taxonomic label (p = .312). Put differently, changes in both visual setup and taxonomic label, and changes in the taxonomic category of the noun trigger a shift of children’s threshold for RGAs more often than changes in the visual setup.

A second model was built adding Adjective (big, small) as fixed effect. This model did not improve the fit of the data compared to the simpler model (AIC = 563.35 vs. 564.96), and the difference in fit from the model without Adjective as a fixed effect was not significant (Χ²(1) = 0.3962, p = .529). That is, children’s shifts of the cutoff were not influenced by the polarity of the RGA.

4.4.2.2 Adults

Parallel to the child group, we analyzed for each item and each adjective the contextual changes (visual setup + taxonomic label, visual setup). In total, there are 624 observations by 26 adults, 312 for AGAs (clean, dirty) and 312 for RGAs (big, small). In 99.4% (310/312) of the AGA cases, the threshold was not shifted; a shift occurred only once (0.3%), and one additional shift (0.3%) was classified as other. The cutoff for clean was consistently the object without dirt (object 12). The cutoff of dirty was consistently object 11, i.e., adults selected all objects with at least one spot of dirt, and the object without dirt was never selected.

Overall, the cutoff of RGAs was shifted more often than the cutoff of AGAs, namely in 50.6% (158/312) of the cases. In 35.6% of the cases (111/312), the cutoff did not shift; 11.2% of the cases (35/312) were shifts in the unexpected direction, and 2.6% of the cases (8/312) were classified as undefined. Undefined responses were excluded from further analyses.

In a next step, we examined the adult shifting pattern more closely to address the question of whether adults and children are sensitive to the same contextual changes (Q4), i.e., whether the shift of the threshold in the adult group is triggered by the same contextual changes as in the child group. The results, displayed in Figure 8, are largely similar to the data for the child group. The frequency of shifts varied with type of contextual change; the violins for two of the three conditions (visual setup + taxonomic label, taxonomic label) for both big and small resemble each other, compared to the violins for the visual setup condition. Adults shifted the threshold for the RGAs big and small most frequently, when both the taxonomic category encoded by the noun and the visual setup changed (big: mean = 67.3%, SD = 31.4%; small: mean = 60.0%, SD = 28.9%). Descriptively, the threshold shifted marginally less often when only the taxonomic category encoded by the noun changed (big: mean = 58.0%, SD = 31.2%; small: mean = 58.3%, SD = 31.9%) and much less often when only the visual setup changed (big: mean = 44.0%, SD = 33.3%; small: 22.9%, SD = 29.4%).

Figure 8: Frequency of cutoff shifts for the relative gradable adjectives by adjective and type of contextual change in the adult group. The black line indicates the mean.

A model similar to the child group was run for the adult group, with Contextual change as a fixed effect, and a random intercept for Participant. The model revealed that the difference between visual setup + taxonomic label and visual setup was highly significant (β = –1.261, Z = –4.219, p < .001), but the difference between visual setup + taxonomic label and taxonomic label was not significant (β = –0.2248, Z = –0.773, p = .439). Further inspection via pairwise comparisons with Tukey adjustment (emmeans package, Lenth 2022), showed a significant difference between visual setup + taxonomic label and visual setup (p < .001), and between visual setup and taxonomic label (p = .001), but not between visual setup + taxonomic label and taxonomic label (p = .719). Similar to the child group, changes in both visual setup and taxonomic label, and changes in the taxonomic category of the noun trigger a shift of adults’ threshold for RGAs more often than changes in the visual setup.

A second model was built adding Adjective (big, small) as a fixed effect. This model did not improve the fit of the data compared to the simpler model (AIC = 407.37 vs. 406.63), and the difference in fit from the model without Adjective as a fixed effect was not significant (Χ²(1) = 2.7408, p = .098). Adults’ cutoff shifts were unaffected by RGA polarity in parallel with children.

5 Discussion

The current study tested children’s and adults’ interpretation of absolute gradable and relative gradable adjectives (AGA and RGA) to examine how listeners determine the comparison class (CC) and calculate the threshold. We created a picture-choice task that varied the taxonomic category encoded by the modified head noun in the test prompt (basic-level, superordinate-level) and/or the composition of the visual setup (baseline, expansion). Our first finding addresses the question of whether children have different thresholds for AGAs and RGAs (Q1). Our analysis of children’s responses in the basic_baseline condition revealed that already the three-year-olds exhibited different thresholds for AGAs and RGAs. The threshold for RGAs was located around the center of the object series. As for AGAs, children’s threshold for the maximum AGA clean was oriented towards the maximal lack of dirtiness and the threshold for the minimum AGA dirty towards the minimal existence of dirtiness. This pattern is in line with findings from previous studies using Scalar Judgement Tasks and Sorting Tasks with physical objects (see Syrett et al. 2006; Foppolo & Panzeri 2013; Gotowski & Syrett, 2020). Addressing the question of for which type of gradable adjectives (RGAs and/or AGAs) changes in the context result in shifts of the threshold (Q2), we found that children shifted the threshold for RGAs when the context changed, whereas their thresholds for AGAs were not affected by contextual changes. This pattern indicates that by age three children’s threshold for RGAs, but not for AGAs, is context-dependent, strengthening the earlier observation that beginning age 3, children are sensitive to the differences between RGAs and AGAs. The main conclusion here is that the ability to distinguish between RGAs and AGAs was unaffected by the specific task (Picture-choice vs. Scalar Judgement/Sorting) and by the mode of presentation (pictures of objects vs. physical objects), that is, by preschool age the distinction between RGAs and AGAs is firmly established.

Our main goal was to investigate which contextual changes trigger a shift of the threshold (Q3). More specifically, we examined whether children use linguistic cues that indicate the taxonomic category (basic-level, superordinate-level), i.e. the modified head noun, for determining the CC and the threshold in the presence of non-linguistic visual cues, i.e. the number of basic-level categories and the distribution of the specific property displayed in the visual setup. Our data provides novel evidence that children preferably use the information provided by the taxonomic category encoded by the modified head noun to calculate the threshold for RGAs. More precisely, children exhibited different thresholds across trials when the taxonomic category encoded by the modified noun changed (e.g, water balloons vs. toys), and they exhibited the same thresholds when the adjective modified the same basic-level noun. Notably, objects in the visual setup not denoted by the noun affected children’s thresholds much less.

Overall, our data reveal that the taxonomic category encoded by the modified head noun plays a more prominent role in determining the CC of RGAs than the composition of the visual setup. By relying on familiar objects that allowed for clear taxonomic categories, this novel finding corroborates Barner & Snedeker’s (2008) results, which were based on fantasy objects and fantasy nouns. Regarding question (Q4) of whether children are sensitive to the same contextual changes as adults, we found that the preference for the linguistic cue observed in adults holds for children as well, as shown by their sensitivity to the same contextual changes.

Our findings for RGAs allow us to explore two further questions regarding the role of the modified head noun in the creation of comparison classes, which we discuss in turn below: Why is the linguistic cue via the modified noun preferred to the non-linguistic cue via the visual setup in determining the CC? Why is there a certain amount of variability in terms of the contextual changes that caused the threshold to shift or not to shift?

The first issue relates to the overarching question of the interaction between language and conceptual organization. We propose that linguistic cues may be privileged over non-linguistic cues for three reasons. First, language can influence cognitive processing during mental computations, because it guides attention to aspects of the perceptual environment that are relevant for communication; other not verbalized aspects are deemphasized (see the review in Ünal & Papafragou 2016). Accordingly, the modified head noun may provide a more salient cue than the visual setup, because it articulates the features of the context relevant for determining the CC. Second, it has been found that linguistic labels serve to emphasize categories, especially at the more abstract superordinate level (Waxman & Markow 1995). Assuming that linguistic labels direct listeners’ attention to similarities between objects, linguistic information can facilitate the establishment of object categories in a way that non-linguistic information cannot. After all, perceptual similarity is less obvious at the superordinate than at the basic level. Toys, for example, come in very different shapes, materials, composition, colors, and sizes, whereas teddy bears do not. Accordingly, nouns like toy, just like teddy bear, provide a shared label that initiates the search for commonalities among sets of objects that are otherwise very different. Linguistic labels have a further important advantage over non-linguistic cues: by labeling a conceptual category, children understand that objects that share a label also share their underlying unobservable properties, e.g., toys share the property of being artifacts that are used to play with (Waxman & Markow, 1995; Fairchild et al. 2018). Put differently, words, in particular nouns, invite the formation of categories. We propose that this property of nouns helps us to create comparison classes. These comparison classes are a prerequisite for further internal categorization, for example when evaluating properties expressed by adjectives.

Let us now turn to the open issue of why there was some variability regarding whether child and adult participants did or did not shift the threshold. On the one hand, the threshold shifted very frequently when the taxonomic category encoded by the noun or both taxonomic label and visual setup changed, but in 43% of the cases it did not. On the other hand, the threshold rarely shifted when only the visual setup changed, but in some cases (28%) it did shift.¹⁰ We propose that both of these findings are related to a specific semantic property of RGAs: vagueness. Vagueness can be understood as uncertainty regarding the threshold, even if the CC is established. Typically, vagueness is described by three characteristics: fuzzy boundaries, the existence of borderline cases, and susceptibility to the Sorites Paradox (e.g., Solt 2015; Burnett 2016, for an overview). The notion of fuzzy boundaries captures the observation that there seem to be no sharp boundaries between entities satisfying a vague adjective (e.g., being big) and entities that do not satisfy it (e.g., not being big). Think of a big car: if we make this car smaller by shrinking its size millimeter by millimeter, it will be impossible to determine the exact size at which this big car turned into a car that no longer counts as being big. Borderline cases are entities for which it is difficult or impossible to judge whether they satisfy the property denoted by the gradable adjective. Consider different types of cars: whereas a truck can be clearly identified as a big car and a Smart car as not big, the judgement for a Honda Civic, for instance, is more difficult. For this borderline case, speakers are inclined to judge it as being neither big nor not big or as both big and not big (Solt & Gotzner 2010; Alxatib & Pelletier 2011; Égré & Zehr 2018). Finally, vague adjectives give rise to the Sorites Paradox,¹¹ which follows from the lack of sharp boundaries and insensitivity or indifference to small changes in the relevant property (Morzycki 2015; Solt 2015).

We suggest that the relaxed dependency of the threshold on the CC and the resulting flexibility of the threshold can be accounted for in terms of vagueness as follows. Flexible thresholds are a reflex of the uncertainty regarding the threshold due to fuzzy boundaries and the existence of borderline cases. In our experiment the size differences between the single objects were rather small (8–10 mm) so that the objects were very similar regarding the property encoded by the adjectives (big, small). It is possible that participants were sometimes indifferent to small differences regarding size and assumed no precise boundary between big and not-big and between small and not-small objects. Consequently, this boundary may but not need shift from trial to trial, i.e. between CCs. Likewise, variance may exist regarding the objects in the borderline area: a borderline case may or may not be selected across trials. Note that this explanation may also account for the infrequent shifts in the unexpected direction.

How children discover in which situations the threshold for AGAs is susceptible to contextual changes, we leave for future research. We acknowledge that in our experimental task there was little need for tolerating imprecision, i.e., to deviate from the minimal or maximal threshold. At this point, we leave open whether children may opt for non-minimal or non-maximal thresholds for AGAs when reasoning about imprecision is called for (e.g., with specific CCs).

In summary, the current study provides novel empirical evidence that the taxonomic category encoded by the modified head noun is a privileged cue for determining the CC and for calculating the threshold. The creation of CCs accesses cognitive representations of object categories and hierarchies that can be articulated via the modified head noun. If this linguistic information is not available, non-linguistic visual contextual information is considered.

The present study examined the taxonomic category encoded by the noun as a linguistic cue and the visual context as a non-linguistic cue, by systematically varying both factors. More research is needed to better understand to what extent our findings are related to the objects’ organization into natural taxonomic hierarchies (i.e. subset-superset relations). As pointed out by one reviewer, this can be tested by extending our design to taxonomically unrelated objects (e.g., water balloons and watermelons). To develop a comprehensive picture of the role of non-linguistic information in the creation of comparison classes, future studies should manipulate other sources of non-linguistic information, such as world knowledge and discourse context. Our study used non-propositional requests, such as Please hand me the big water balloons, as did Barner & Snedeker (2008). More research needs to be done to see if what we found extends to propositions, such as This water balloon is big, which make an assertion about one entity at a time. Finally, findings by Tessler et al. (2020) for adults suggest that adjective position mediates how the noun contributes to inferring the comparison class: when the adjective was in predicative position, adult listeners were less likely to use the noun as a cue to the comparison class than when the adjective was in attributive position (This is a big water balloon). Whether the same applies to children remains to be seen.¹²

6 Conclusion

Mastering the semantics of gradable adjectives is a complex acquisition task. Our study contributes to the question of how children and adults interpret relative and absolute gradable adjectives. We assessed to what extent German-speaking three- to five-year olds and adults use the taxonomic category encoded by the modified head noun (basic-level, superordinate-level) and visual contextual information to determine the comparison class and to calculate the threshold for two relative gradable (big, small) and two absolute gradable (clean, dirty) adjectives. Using the method of picture-choice, we presented children with an unordered display of existing objects aiming at a close proxy for real-life scenarios. We found that, in this experimental setup, children and adults interpret relative, but not absolute gradable adjectives as context-dependent, thus providing further evidence that central aspects of adjective semantics are mastered by age three. Furthermore, we showed that children are able to establish an ordering among multiple objects also in unordered arrays, organized in basic-level and superordinate-level categories, and to evaluate for which objects a gradable adjective is true. Most importantly, our study is the first to demonstrate for existing objects that, from age three, linguistic information provided by the taxonomic label is privileged over non-linguistic visual information in creating the comparison class of relative gradable adjectives. This finding adds to the notion that language is a powerful tool to modulate our mental computations.

Data availability

Data files, R script, additional experiment materials and results: https://osf.io/hc2p3/?view_only=b42dd5c131d84f38ab93c2c4456021cb.

Ethics and consent

This study was approved by the Research Ethics Board of DIPF Leibniz Institute for Research and Information in Education (#DIPF_EK_2021_30). Informed consent was obtained from each adult participant and the legal guardian of each child participant. All research data has been anonymized.

Funding information

This research was supported by a grant from the German Research Foundation (DFG, GRK 2016/1 “Nominal Modification”, PI: Esther Rinke). This study was also part of the project Vari, supported by the IDeA – Center for Research on Individual Development and Adaptive Education of Children at Risk, Frankfurt am Main, Germany.

Competing interests

The authors have no competing interest to declare.