In a study published in Frontiers in Language Sciences, researchers Jesse Holmes and Virve-Anneli Vihman from the University of Tartu demonstrate that listeners rely heavily on information structure to interpret ambiguous sentences in artificial languages. The study, which employed an artificial language learning paradigm, investigated how native speakers of Estonian assigned grammatical roles in sentences lacking explicit marking. The findings indicate a strong effect of givenness, where previously mentioned information was consistently interpreted as the subject, alongside a smaller, less certain effect of animacy. These results suggest that typicality biases, particularly those rooted in discourse context, play a key role in shaping how humans process differential argument marking systems.
Differential argument marking, or DAM, is a linguistic phenomenon where nouns are marked for their grammatical role only in specific contexts. This manifests as differential object marking, where only certain direct objects receive grammatical markers, or differential subject marking, where only specific subjects are marked. In many natural languages, such as Spanish or Fore, these markings are conditioned by semantic properties like animacy or discourse properties like definiteness. For instance, in Spanish, animate objects are often marked, whereas inanimate ones are not. The study aimed to understand whether these cross-linguistic patterns arise from cognitive biases that listeners use to resolve ambiguity when explicit markers are absent.
Previous research on DAM has primarily focused on production, examining how speakers choose to mark arguments. However, the comprehension side of this process has been less explored. The current study addressed this gap by testing how learners interpret ambiguous sentences where word order is flexible and no grammatical markers are present. The researchers hypothesized that if typicality biases exist, participants would assign roles based on expectations: given information would be interpreted as the subject, and new information as the object. Additionally, they predicted that information structure would take precedence over semantic properties like animacy in determining these interpretations.
The experiment involved 67 native speakers of Estonian, divided into two groups: one learning an object-marking language and the other a subject-marking language. The artificial languages contained eight nouns, four animate and four inanimate, along with three transitive verbs. Word order was flexible, with the verb always appearing at the end, but the subject and object could appear in either order. Marking was optional and appeared in only 50% of transitive sentences, uncorrelated with word order, animacy, or information structure. This design ensured that any biases observed in interpretation were not learned from the input but reflected pre-existing cognitive tendencies.
Participants underwent a three-day training regimen. On each day, they learned the nouns and sentence structures through interactive tasks with feedback. The critical test occurred on the third day, where participants interpreted unmarked sentences without receiving any feedback. In these trials, each sentence was compatible with two possible events, differing only in which referent was the agent. Participants had to choose the event they believed best matched the sentence. This forced-choice design allowed the researchers to measure how participants resolved ambiguity in the absence of explicit grammatical cues.
The results revealed a robust givenness effect. The Bayesian mixed-effects logistic regression models showed that when the first argument in a sentence was given information, participants were significantly more likely to interpret it as the subject. Conversely, when the first argument was new information, they were more likely to interpret it as the object. This effect was large and statistically robust, with the 95% credible interval for the givenness order parameter entirely excluding zero. This suggests that listeners strongly expect given information to align with the subject role, a tendency that likely reflects the general cognitive bias to anchor transitive events around the agent.
In contrast, the effect of animacy was smaller and less certain. While the trends aligned with typicality predictions—animate-inanimate order favoring subject interpretations and inanimate-animate order favoring object interpretations—the credible intervals for these parameters included zero. This indicates that while animacy may play a role, its influence is weaker and more variable than that of information structure. The researchers suggest that this difference may stem from the fact that givenness is a more abstract, discourse-based property, whereas animacy is a concrete, perceptual feature that participants might consciously adjust for during training.
Notably, there were no significant differences between the object-marking and subject-marking groups. Both groups exhibited similar patterns of interpretation, with no evidence that the type of marking learned affected how they resolved ambiguity in unmarked sentences. This finding challenges the idea that comprehension biases are specific to object-marking languages. Instead, it suggests that these biases are general properties of comprehenders, independent of the specific marking system they have learned. This has implications for understanding the cross-linguistic asymmetry between DOM and DSM, suggesting that the disparity may arise from factors other than comprehension biases, such as production tendencies or the frequency of atypical agents and patients.
The study also highlights the importance of information structure in the emergence of DAM systems. The strong givenness effect suggests that languages may develop marking systems to overcome listener expectations, particularly when information structure is atypical. For example, if a given object is marked, it may be because the listener would otherwise expect it to be the subject. This communicative account aligns with previous findings that DAM is a tool for efficiently communicating unexpected or atypical situations. The researchers propose that future studies should test this directly through communication games, examining how speakers adjust their production in response to listener comprehension biases.
Limitations of the study include the restriction to native speakers of Estonian. While this provides a valuable contrast to previous studies that relied exclusively on English speakers, it also introduces potential confounds. Estonian has a flexible word order and a complex object-marking system, which may have influenced the results. The researchers acknowledge that the observed effects of givenness and the bias for subject-first interpretations could be residuals of Estonian language biases. However, they argue that the lack of a difference between the DOM and DSM groups supports the idea that these effects are not solely driven by native-language marking biases. Future research should replicate these findings with speakers of a more diverse set of languages to assess the generalizability of the results.
The study’s design deliberately excluded a production phase to prevent participants from adopting their own grammatical conventions, which could have influenced how they interpreted subsequent sentences. In previous artificial language experiments, such as those by Fedzechkina et al. (2012) and Smith and Culbertson (2025), production was the primary outcome measure, and comprehension often served as a training check or a means to provide feedback. By removing production and feedback entirely, Holmes and Vihman isolated the listener’s default expectations. This methodological shift was crucial because feedback on ambiguous sentences would have taught a specific role assignment, thereby masking the very biases the study aimed to measure.
The artificial languages used in the experiment were constructed to ensure that marking was uncorrelated with word order, animacy, or information structure. Each language contained eight nouns, four animate and four inanimate, and three transitive verbs. The verb always appeared at the end of the sentence, while the subject and object nouns appeared in free variation, with 50% of sentences in SOV order and 50% in OSV order. The optional argument-marking suffix appeared in only 50% of transitive sentences and was distributed equally across all word orders, animacy types, and information structure conditions. This random distribution meant that participants could not rely on statistical regularities in the input to predict roles; any biases observed in their interpretations must have originated from pre-existing cognitive tendencies rather than learned associations.
The information structure manipulation was implemented through a specific discourse context. Before each transitive sentence, participants were presented with an existential construction, such as “there is X,” which introduced one noun into the discourse, making it given. This was followed by an interrogative construction that placed the other noun, the new argument, in focus. This setup created two distinct contexts for each transitive event: one where the agent was given and the patient was new, and another where the patient was given and the agent was new. The existential construction allowed the researchers to manipulate givenness without introducing other asymmetries between the arguments, while the interrogative construction motivated the presentation of two candidate events at test, as both were congruent answers to the question asked.
The Bayesian mixed-effects logistic regression models provided detailed estimates of the uncertainty surrounding the observed effects. For the unmarked sentence comprehension, the model’s intercept had a posterior mean of -0.75, with a 95% credible interval of [-1.37, -0.15], indicating a consistent overall bias against OSV interpretations. This suggests that participants generally favored SOV interpretations of unmarked sentences. The givenness order parameter had a posterior mean of -1.17 log-odds, with a 95% credible interval of [-1.96, -0.41], confirming that new-given order resulted in more OSV interpretations than given-new order. In contrast, the animacy order parameters for animate-inanimate and inanimate-animate orders had posterior means of -0.22 and 0.28 log-odds, respectively, with credible intervals that included zero. This statistical distinction underscores the robustness of the givenness effect compared to the weaker, less certain animacy effect.
The lack of difference between the object-marking and subject-marking groups has significant implications for theories of cross-linguistic asymmetry. Previous research, such as Fauconnier (2011), found that differential object marking is more common than differential subject marking, and that systems marking exclusively inanimate subjects are unattested. If comprehension biases were specific to object-marking languages, one might expect a difference between the two groups in this study. The absence of such a difference suggests that the cross-linguistic disparity between DOM and DSM may arise from factors other than comprehension biases, such as production tendencies, the relative frequency of atypical agents and patients, or the sources from which markers are recruited. This finding challenges the idea that the type of marking system learned modulates the listener’s reliance on typicality biases for disambiguation.
The researchers also considered the possibility that the observed biases are rooted in an agent bias, a cross-linguistic tendency to construe transitive events around the agent as the default anchor of clause structure. This bias often manifests as a tendency for agents to precede patients and for agents to be processed more easily in such configurations. The overall bias toward SOV interpretations in the study is consistent with this agent bias. However, the strong effect of givenness suggests that information structure may play a more dominant role in shaping comprehension than animacy or agency alone. The researchers propose that givenness might be a more general, omnipresent phenomenon, with agents appearing more frequently as given in discourse than as animate, and patients more frequently as new than as inanimate. This would explain the stronger typicality bias based on givenness observed in the study.
Subject of Research: Comprehension of differential argument marking in artificial languages
Article Title: Typicality biases in interpreting unmarked sentences: an artificial language learning experiment on differential argument marking
Article References: Holmes, J., & Vihman, V.-A. (2026). Typicality biases in interpreting unmarked sentences: an artificial language learning experiment on differential argument marking. Frontiers in Language Sciences, 5, Article 1862016. https://doi.org/10.3389/flang.2026.1862016
Image Credits: AI Generated
DOI: 10.3389/flang.2026.1862016
Keywords: differential argument marking, artificial language learning, information structure, givenness, animacy, psycholinguistics, word order, grammatical role assignment, listener expectations, Estonian, Typicality, biases
Cite Scienmag News
Courtney Benton. (November 10, 2026). Listeners Rely on Information Structure to Resolve Ambiguous Grammar. Scienmag. https://scienmag.com/listeners-rely-on-information-structure-to-resolve-ambiguous-grammar/
Courtney Benton. "Listeners Rely on Information Structure to Resolve Ambiguous Grammar." Scienmag, 10 November 2026, https://scienmag.com/listeners-rely-on-information-structure-to-resolve-ambiguous-grammar/. Accessed 22 November 2026.
Courtney Benton. "Listeners Rely on Information Structure to Resolve Ambiguous Grammar." Scienmag. November 10, 2026. https://scienmag.com/listeners-rely-on-information-structure-to-resolve-ambiguous-grammar/








