Determining the cognitive biases behind a viral linguistic universal: the order of multiple adjectives

Across typologically diverse languages, adjectives tend to appear in a preferred order (often popularized as a universal hierarchy), e.g., subjective comment > color > material, yielding sequences like “beautiful blue porcelain earrings.” Traditional accounts either posit a syntactic, innately specified hierarchy (SOT) or a multifactorial cognitive origin (mCOT) where factors such as subjectivity/objectivity, inherentness, noun-specific frequency/collocability, and phonological weight guide preferences. Despite extensive descriptive work, the cognitive underpinnings—why these preferences arise—remain unclear. This study aims to connect the observed phenotypic behavior to cognitive needs and environmental triggers, posing two research questions: (I) What cognitive needs are subserved by adjective ordering preferences (AOPs)? (II) Is there interspeaker variation in the preferences, and if so, how can this coexist with claims of a strong linguistic universal? The work tests parts of the commonly cited hierarchy and evaluates whether violations are truly marked/ungrammatical or instead acceptable, and whether preferences vary across speaker groups.

Two main strands of explanation exist. The syntactic origin theory (SOT) posits an innate, universal DP-internal hierarchy with designated positions for adjective classes, embedded in a broader spine (Demonstratives > Numerals > Adjectives > Nouns). The multifactorial cognitive origin theory (mCOT) treats ordering as preferences resulting from factors including: degree of subjectivity vs. object-orientation, inherentness/absoluteness, noun-specific frequency and collocability/idiomaticity, and phonological weight/length. Empirical work shows robust preferences but also flexibility, with distance effects among adjacent classes (e.g., shape and color more interchangeable). Prior experiments and corpus studies report that violating orders can be acceptable and used in spontaneous speech. Recent calls emphasize the need to identify the cognitive basis for these factors and to consider environmental/statistical learning influences.

Design: Timed acceptability judgment task measuring (i) acceptability on a 3-point Likert scale (‘correct’, ‘neither correct nor wrong’, ‘wrong’) and (ii) reaction times (RTs). The presence of processing cost for marked orders would be evidenced by longer RTs relative to unmarked orders. Stimuli and conditions: All items were Adjective-Adjective-Noun strings with Spelke object nouns (e.g., “I bought a square black table”). Three conditions were tested, each with two orders (congruent vs. incongruent relative to the commonly cited hierarchy) and three test structures per order, totaling 18 test structures:

1. size adjective – nationality adjective
2. shape adjective – color adjective
3. subjective comment adjective – material adjective Congruent orders followed the standard hierarchy (e.g., size before nationality); incongruent orders reversed it. The task was implemented in Ibex Farm. Participants: N=139 bilingual adults who are native speakers of Greek and proficient in at least one Germanic language (mainly Norwegian, Swedish, Danish, English, or German). The original sample was 167; 28 were excluded based on pre-registered criteria: extremely fast responses (RT < 600 ms), incomplete task, non-native Greek, speech-pathology treatment, or neurological disorders (the latter three by self-report). All provided informed consent; ethical approval by the Norwegian Center of Research Data (55775/3/LH). Testing language: Greek. Participants had resided outside Greece for at least 4 years at testing (mean 11.8 years, SD 9.8), mainly in Scandinavia, the UK, or Germany. Procedure: The task was completed online. RTs were log10-transformed to normalize distributions; a 3 SD outlier filter was applied (removed 15/1251 congruent trials, 13/1251 incongruent). Analyses used jamovi 1.8 (R 4.0). Acceptability judgments were modeled via generalized linear models (multinomial). RTs were analyzed via generalized linear models treating RT as continuous. Judgments and RTs were also analyzed jointly to link decision type with processing time. Comparative analysis: Results were compared to a monolingual Greek group from prior work (n=140) using the same task to assess effects of language group/developmental trajectory. Accuracy for incongruent items was analyzed as logistic (target for incongruent = ‘wrong’). Data and materials are available at the provided repository link.

• Data points: RTs n=2474 (after outlier removal); judgments n=2502. Acceptability:
• Both congruent and incongruent sentences were predominantly judged ‘correct’. Overall, order had a significant effect on judgments (χ²=100.38, p<0.001); condition did not (χ²=2.22, p=0.694); order×condition interaction was significant (χ²=36, p<0.001).
• Within-condition effects of order: significant for size–nationality (χ²=56.6, p<0.001) and subjective comment–material (χ²=99.2, p<0.001), not significant for shape–color (χ²=2.39, p=0.302), consistent with a distance effect. Reaction times:
• No significant difference in processing times between congruent and incongruent orders (effect of order: χ²=2.837, p=0.092). Condition and order×condition effects were not significant (χ²=0.111, p=0.946; χ²=2.052, p=0.358).
• Judgment type strongly affected RT: ‘correct’ judgments were associated with the shortest decision times; the effect of judgment on RTs was significant (χ²=339, p<0.001). Notably, for accepted-as-correct trials, congruent orders showed slightly longer RTs than incongruent. Group comparison (bilinguals vs. monolinguals):
• Overall effect of language group on acceptability was not significant (χ²=0.075, p=0.963); the interaction group×order was marginally significant (χ²=6.459, p=0.040). Given ceiling ‘correct’ on congruent items, a focused analysis examined accuracy on incongruent items (target judgment = ‘wrong’).
• For incongruent items, accuracy differed by group (χ²=7.48, p=0.006), with bilinguals more likely to provide the target rejection; condition was significant (χ²=43.40, p<0.001); group×condition was not (χ²=4.53, p=0.104). Post-hoc tests (Bonferroni) supported main effects (e.g., shape–color vs. subjective comment–material p<0.001; size–nationality vs. shape–color p=0.011; size–nationality vs. subjective comment–material p=0.003; group p=0.014). Interpretation:
• Violating orders were highly acceptable and did not incur extra processing cost, contradicting the notion that they are ungrammatical or marked in processing terms. Preferences exist but are not rigid; shape–color shows near interchangeability.

Findings indicate that adjective ordering is governed by preferences rather than rigid universal restrictions. The absence of additional processing cost for hierarchy-violating orders suggests they are not marked at the parser level. The study proposes that AOPs arise from a synergistic interplay of cognitive biases and perceptual/usage factors: (1) Ambiguity Intolerance promotes orders that facilitate disambiguation, allowing flexible prioritization of properties (e.g., shape vs. color) depending on context; (2) Novel Information Bias leads less expected, more subjective or generic descriptors to occupy salient edge positions in Adj–Adj–N sequences, while highly collocational, object-oriented adjectives appear closer to the noun; (3) The Principle of Least Effort/Zipf’s law links frequency, genericity, and brevity, reinterpreting the phonological weight effect as an epiphenomenon of meaning-driven frequency. Nationality/origin adjectives are treated as sociopragmatic conventions expressing idiosyncratic relations rather than inherent properties, explaining their proximity to the noun. Shape and color may be close in perceptual processing and expectation, yielding interchangeability; visual syntax and production availability can modulate order. Interspeaker variation between bilinguals and monolinguals likely reflects sensitivity to the statistical distribution of input and exposure to prescriptive rules in L2 contexts. While the results do not adjudicate decisively between SOT and mCOT, they challenge claims of rigidity central to strong syntactic hierarchy accounts, aligning more with multifactorial cognitive explanations.

The study shows that violations of the commonly cited adjective order are highly acceptable and not slower to process, supporting the view that AOPs are preferences, not categorical constraints. It attributes these preferences to the combined effects of Ambiguity Intolerance, Novel Information Bias, and the Principle of Least Effort (Zipf’s law), with linguistic realization further shaped by statistical properties of the input, yielding interspeaker variation even within the same language. Comparing bilingual and monolingual Greek speakers revealed differences consistent with statistical learning and exposure to prescriptive rules. Overall, the work integrates cognitive biases and usage statistics to explain the origin and variability of adjective ordering preferences.

• The experimental design cannot adjudicate between a purely syntactic hierarchy account (SOT) and multifactorial cognitive accounts (mCOT); it was not intended to test that theoretical distinction directly.
• The phonological weight/length factor was controlled in the materials; thus, the present experiment is not directly informative about weight effects beyond theoretical reinterpretation.
• Stimuli were presented out of context, which clarifies default acceptability but may underrepresent context-driven ordering strategies.
• Participant sample comprised Greek L1 speakers (bilinguals compared to a monolingual Greek sample from prior work); generalizability across languages and broader populations was not tested within this study.