Language as shaped by the environment: linguistic construal in a collaborative spatial task

The paper investigates why languages vary in how they conceptualize and describe the world, focusing on whether such variation is motivated by identifiable environmental factors rather than purely stochastic cultural drift or universal innate biases. Prior work shows cross-linguistic diversity in conceptual domains (e.g., color terms; spatial frames of reference) and correlational links between linguistic structures and social-ecological variables. The authors hypothesize that linguistic conventions are contingent on environmental affordances: the structure of the shared, non-linguistic environment can motivate different conceptual construals that become routinized in interaction. Using an adapted Maze Game, they test whether different maze topologies (irregular, stratified, regular) elicit distinct linguistic strategies for communicating locations (FIGURAL, LINE, MATRIX, PATH), and whether these strategies stabilize over time and generalize to a neutral environment. Two main hypotheses guide the study: H1, that maze topology selects for specific strategies (with explicit predictions for each environment); and H2, that once established, strategies persist even in a neutral final maze, indicating conventionalization.

The authors review debates between nativist accounts predicting universal linguistic structures and relativist, cultural-evolutionary views emphasizing learned conventions and diversity. They survey evidence that languages adapt to their ecological niches: links between social variables and morphological complexity; phonetic variability tied to physiology and environment; and associations between physical environments and lexical or phonological inventories. Spatial language is highlighted as an example where geocentric vs egocentric frames correlate with local topography and sociocultural context. However, most evidence is correlational, with potential confounds (e.g., education, contact, subsistence) and limited causal inference. This motivates experimental approaches to uncover mechanisms by which environmental affordances shape conceptual construals and linguistic conventions during interaction.

Design: An adapted version of Garrod and Anderson’s Maze Game was implemented via a written chat system (Dialogue Experimental Toolkit). Dyads saw the same 7×7 grid-based maze on each trial but had different start/goal positions. Movement paths were blocked by gates; each participant’s gates could be opened only by switches that only their partner could operate (and vice versa), requiring dialogue to coordinate and convey switch locations. Each dyad completed 12 mazes in sequence: the first 11 were condition-specific; the 12th was neutral across conditions. Conditions (environmental affordances): Three maze topologies were created by systematically varying a 7×7 grid: (1) Irregular—featuring salient geometric/figurative shapes and protrusions, affording figural landmarking; (2) Stratified—featuring prominent horizontal displacements suggestive of rows/lines; (3) Regular—dense, grid-like distributions without salient landmarks, affording abstract coordinate or path descriptions. Participants: 66 university students (33 dyads; 24 male/42 female; mean age 23, SD 3) at Aarhus University, randomly paired and unfamiliar with each other. Three additional pairs were excluded for non-compliance or inability to solve the task. Informed consent was obtained per local ethics regulations. Procedure: Dyads communicated via text chat to identify and operate switches to open gates and reach goals. The task required establishing shared reference systems for positions. The first 11 mazes operationalized H1 (environmental selection of strategies); the final neutral maze operationalized H2 (persistence/generalization of established strategies). Coding of strategies: Utterances were coded into four referential strategies following prior Maze Game typologies: FIGURAL (reference to salient local shapes/landmarks), LINE (row/line-based descriptions), MATRIX (abstract grid coordinates: rows×columns), PATH (procedural path from a reference point: step sequences). Statistical analysis: Bayesian multilevel multinomial regression modeled the probability of each strategy as a function of condition (for H1) and, separately, using only the final trial (for H2). Random effects were included by dyad, interlocutor, and maze to regularize variability. Regularizing priors: normal centered at chance (25% use per strategy; log-odds ≈ −1) for strategy occurrence; half-normal (mean 0, SD 0.1) for varying effects; LKJ η=5 for correlation structures. Model quality was assessed via prior and posterior predictive checks, Rhat < 1.01, and effective sample sizes > 200. Temporal dynamics were tested with additional models including time (maze index) as linear and as monotonic; out-of-sample performance was compared using stacking weights based on PSIS-LOO. To examine transitions and attractors among strategies, discrete-time Markov chains were constructed per dyad and averaged within condition via bootstrap (n=100), yielding transition matrices and attraction strengths. Analyses were run in R (3.6.1), RStudio, tidyverse, brms (2.9.0), Stan (2.19), and MarkovChain (0.6.9.16).

H1 (environment selects strategies):

• FIGURAL: Strong evidence that irregular mazes increased FIGURAL descriptions vs stratified (Δlog-odds=1.75, 95% CI 1.14–2.36, ER>1000) and vs regular (2.91, 2.42–3.39, ER>1000).
• LINE: Strong evidence that stratified mazes increased LINE descriptions vs irregular (2.22, 1.55–2.88, ER>1000) and vs regular (3.13, 2.5–3.78, ER>1000).
• MATRIX: Only partial/negative evidence for higher MATRIX use in regular mazes: regular vs irregular (0.04, −0.92 to 0.78, ER=0.87; ER01=1.69), regular vs stratified (0.93, −1.65 to −0.23, ER=0.02; ER01=0.2). Although infrequent, MATRIX tended to be stable once adopted.
• PATH: Strong evidence that regular mazes increased PATH descriptions vs irregular (3.86, 2.71–5.02, ER>1000) and vs stratified (5.22, 3.9–6.56, ER>1000). Temporal dynamics (first 11 mazes):
• FIGURAL increased over time in irregular condition (0.19, 0.03–0.34, ER=36.74); FIGURAL increased more than MATRIX (0.40, 0–0.83, ER=19.1).
• LINE increased over time in stratified condition (0.92, 0.58–1.28, ER>1000), more than FIGURAL (0.58, 0.26–0.92, ER>1000) and PATH (2.19, 0.85–3.67, ER>1000).
• PATH increased over time in regular condition (0.72, 0.17–1.33, ER=306.69); PATH > FIGURAL over time (1.12, 0.32–2.00, ER=124); PATH > LINE (0.78, −0.07 to 1.72, ER=27). H2 (persistence in neutral final maze):
• FIGURAL: Irregular > regular (0.91, −0.01 to 1.83, ER=18.23); irregular > stratified showed only partial evidence (0.34, −0.56 to 1.22, ER=2.84; ER01=1.11).
• LINE: Stratified > irregular (0.95, 0.15–1.76, ER=35.36) and stratified > regular (1.06, 0.27–1.85, ER=89.91).
• MATRIX: No clear selection by condition (regular > irregular: 0.25, −0.58 to 1.07, ER=2.24; ER01=1.24; regular > stratified: 1.38, −0.28 to 3.07, ER=1.87; ER01=1.43).
• PATH: Regular > irregular (0.20, −0.61 to 0.98, ER=3.88); regular > stratified (0.83, −0.84 to 2.48, ER=10.4). Markov-chain dynamics (first 11 mazes):
• Strategy-specific attractors aligned with environments: FIGURAL (irregular), LINE (stratified), PATH (regular).
• MATRIX, while relatively rare, was highly stable when adopted; in the regular condition, probability of remaining in MATRIX on the next trial ≈ 0.97, versus ≈ 0.05 for remaining in FIGURAL. In stratified mazes, there was a ≈16% tendency to transition from MATRIX to LINE, suggesting LINE may be a stronger abstract attractor in that environment. Overall: Abundant evidence supports H1p1, H1p2, H1p4; partial evidence for H1p3. For H2, strong support for persistence of LINE (H2p2) and PATH (H2p4), partial for FIGURAL (H2p1), and evidence against MATRIX persistence advantage (H2p3).

Findings demonstrate that subtle environmental affordances in a shared task space can shape the emergence and stabilization of linguistic conventions. Different maze topologies biased dyads toward different conceptual construals and corresponding description strategies: irregular layouts promoted figural landmarking; stratified layouts elicited row-based schemes; and regular layouts favored procedural path descriptions. Over repeated interaction, these strategies became entrenched and, in many cases, persisted even when environmental affordances were neutralized in a final trial, indicating proto-conventionalization. MATRIX descriptions, although less frequently discovered, were highly stable once adopted, aligning with prior work showing community-level diffusion can promote abstract coordinate systems. The results address the causal question by experimentally manipulating environmental structure and observing downstream linguistic behavior, supporting the view that language use adapts to contextual niches and that local interactional dynamics can seed conventionalization. This helps bridge micro-level mechanisms (interaction, alignment, conceptual pacts) with macro-level patterns of linguistic diversity.

The study shows experimentally that environmental affordances can motivate the emergence and stabilization of distinct linguistic strategies during collaborative problem solving. By manipulating maze topology, the authors causally demonstrate selection of FIGURAL, LINE, and PATH strategies and their consolidation over time, with partial generalization to a neutral context. Although MATRIX strategies were less frequently adopted, they proved highly stable when discovered. These mechanisms provide a plausible pathway from local interactional dynamics to broader cross-linguistic variation. Future research should examine diffusion in community settings (partner swapping) to assess how strategies spread, test additional environmental structures and tasks for generality, integrate computational models with experimental data, and complement laboratory findings with fieldwork to enhance ecological validity and understand interactions among environmental, social, and historical factors.

The laboratory setting and abstract task limit ecological validity; real-world linguistic adaptation involves richer social, cultural, and environmental complexities. Sample size and distribution across conditions may constrain generalizability. The written-chat modality may differ from spoken interaction. The study did not include a community diffusion manipulation (e.g., rotating partners), which could affect which abstract strategies dominate. Relations between environment and linguistic construal are probabilistic, not deterministic; multiple attractors can coexist, and adoption trajectories vary across dyads.