Inflectional morphology refers to the mapping from grammatical information to surface forms, which are typically realized as morphemes. This mapping often exhibits fusion, where several abstract features are expressed in a single morpheme that cannot be decomposed into meaningful parts. Here, we discuss crosslinguistic generalizations of morphological fusion. We argue that fusion reflects principles of efficient processing, as formalized by the memory–surprisal tradeoff (Hahn, Degen, & Futrell 2021), which is based on information-theoretic models of language processing from psycholinguistics. We first show that the existence of fusion itself can, in some situations, lead communicative codes to be more efficient under our processing model. Particularly, we reveal via simulation that the fusion of highly correlated features is more efficient for processing, whereas agglutination is more efficient when features are less correlated. We next discuss crosslinguistic patterns of fusion in real languages. First, we analyze well-known generalizations about features that are commonly fused across languages (e.g. tense, aspect, and mood), as well as a typological pattern regarding suppletion. In both cases, we find that the universals we study tend to reflect a tendency toward more efficient structure under our model of language processing. Finally, we use paradigm and frequency data from four languages to study informational fusion, a gradable measure of fusion defined in Rathi et al. 2021. We find that informational fusion is higher when features are highly correlated, which suggests that gradable fusion is also influenced by optimization for the memory–surprisal tradeoff.

This article is a crosslinguistic investigation of the hypothesis that the average information rate conveyed during speech communication results from a trade-off between average information density and speech rate. The study, based on seven languages, shows a negative correlation between density and rate, indicating the existence of several encoding strategies. However, these strategies do not necessarily lead to a constant information rate. These results are further investigated in relation to the notion of syllabic complexity.

The world's languages exhibit striking diversity. At the same time, recurring linguistic patterns suggest the possibility that this diversity is shaped by features of human cognition. One well-studied example is word order in complex noun phrases (like these two red vases). While many orders of these elements are possible, a subset appear to be preferred. It has been argued that this ordering reflects a single underlying representation of noun phrase structure, from which preferred orders are straightforwardly derived (e.g. Cinque 2005). Building on previous experimental evidence using artificial language learning (Culbertson & Adger 2014), we show that these preferred orders arise not only in existing languages, but also in improvised sequences of gestures produced by English speakers. We then use corpus data from a wide range of languages to argue that the hypothesized underlying structure of the noun phrase might be learnable from statistical features relating objects and their properties conceptually. Using an information-theoretic measure of strength of association, we find that adjectival properties (e.g. red) are on average more closely related to the objects they modify (e.g. wine) than numerosities are (e.g. two), which are in turn more closely related to the objects they modify than demonstratives are (e.g. this). It is exactly those orders which transparently reflect this—by placing adjectives closest to the noun, and demonstratives farthest away—that are more common across languages and preferred in our silent gesture experiments. These results suggest that our experience with objects in the world, combined with a preference for transparent mappings from conceptual structure to linear order, can explain constraints on noun phrase order.

Autoregressive language models generate text by predicting the next word from the preceding context. The regularities internalized from specific training data make this mechanism a useful proxy for historically situated readerly expectations, reflecting what earlier linguistic communities would find probable or meaningful. In this article, I pre-train a GPT model (223M parameters) on a broad corpus of Chinese texts (FineWeb Edu Chinese V2.1) and fine-tune it on the collected writings of Mao Zedong (1893–1976) to simulate the evolving linguistic landscape of post-1949 China. Identifying token sequences with the sharpest drops in perplexity – a measure of the model’s surprise – reveals the core phraseology of “Maospeak,” the militant language style that developed from Mao’s writings and pronouncements. A comparative analysis of modern Chinese fiction demonstrates how literature becomes unfamiliar to the fine-tuned model, generating perplexity spikes of increasing magnitude. The findings suggest a mechanism of attentional control: whereas propaganda backgrounds meaning through repetition (cognitive overfitting), literature foregrounds it through deviation (non-anomalous surprise). By visualizing token sequences as perplexity landscapes with peaks and valleys, the article reconceives style as a probabilistic phenomenon and showcases the potential of “cognitive stylometry” for literary theory and close reading .

What is the optimal level of questionnaire detail required to measure bilingual language experience? This empirical evaluation compares alternative measures of language exposure of varying cost (i.e., questionnaire detail) in terms of their performance as predictors of oral language outcomes. The alternative measures were derived from Q-BEx questionnaire data collected from a diverse sample of 121 heritage bilinguals (5–9 years of age) growing up in France, the Netherlands and the UK. Outcome data consisted of morphosyntax and vocabulary measures (in the societal language) and parental estimates of oral proficiency (in the heritage language). Statistical modelling exploited information-theoretic and cross-validation approaches to identify the optimal language exposure measure. Optimal cost–benefit was achieved with cumulative exposure (for the societal language) and current exposure in the home (for the heritage language). The greatest level of questionnaire detail did not yield more reliable predictors of language outcomes.

Language models can produce fluent, grammatical text. Nonetheless, some maintain that language models don’t really learn language and also that, even if they did, that would not be informative for the study of human learning and processing. On the other side, there have been claims that the success of LMs obviates the need for studying linguistic theory and structure. We argue that both extremes are wrong. LMs can contribute to fundamental questions about linguistic structure, language processing, and learning. They force us to rethink arguments and ways of thinking that have been foundational in linguistics. While they do not replace linguistic structure and theory, they serve as model systems and working proofs of concept for gradient, usage-based approaches to language. We offer an optimistic take on the relationship between language models and linguistics.

Fish swimming together in schools interact via multiple sensory pathways, including vision, acoustics and hydrodynamics, to coordinate their movements. Disentangling the specific role of each sensory pathway is an open and important question. Here, we propose an information-theoretic approach to dissect interactions between swimming fish based on their movement and the flow velocity at selected measurement points in the environment. We test the approach in a controlled mechanical system constituted by an actively pitching airfoil and a compliant flag that simulates the behaviour of two fish swimming in line. The system consists of two distinct types of interactions – hydrodynamic and electromechanical. By using transfer entropy of the measured time series, we unveil a strong causal influence of the airfoil pitching on the flag undulation with an accurate estimate of the time delay between the two. By conditioning the computation on the flow-speed information, recorded by laser Doppler velocimetry, we discover a significant reduction in transfer entropy, correctly implying the presence of a hydrodynamic pathway of interaction. Similarly, the electromechanical pathway of interaction is identified accurately when present. The study supports the potential use of information-theoretic methods to decipher the existence of different pathways of interaction between schooling fish.

Convergence of the expectation-maximization (EM) algorithm to a global optimum of the marginal log likelihood function for unconstrained latent variable models with categorical indicators is presented. The sufficient conditions under which global convergence of the EM algorithm is attainable are provided in an information-theoretic context by interpreting the EM algorithm as alternating minimization of the Kullback–Leibler divergence between two convex sets. It is shown that these conditions are satisfied by an unconstrained latent class model, yielding an optimal bound against which more highly constrained models may be compared.

We develop and demonstrate a computationally cheap framework to identify optimal experiments for Bayesian inference of physics-based models. We develop the metrics (i) to identify optimal experiments to infer the unknown parameters of a physics-based model, (ii) to identify optimal sensor placements for parameter inference, and (iii) to identify optimal experiments to perform Bayesian model selection. We demonstrate the framework on thermoacoustic instability, which is an industrially relevant problem in aerospace propulsion, where experiments can be prohibitively expensive. By using an existing densely sampled dataset, we identify the most informative experiments and use them to train the physics-based model. The remaining data are used for validation. We show that, although approximate, the proposed framework can significantly reduce the number of experiments required to perform the three inference tasks we have studied. For example, we show that for task (i), we can achieve an acceptable model fit using just 2.5% of the data that were originally collected.