Results mirror findings by Siew and Vitevitch and demonstrate that preferential attachment is the main network growth algorithm driving lexical learning at early second-language proficiency stages, while inverse preferential attachment prevails at more advanced proficiency stages. The present study was designed as a replication of Siew and Vitevitch's (2020a) study "An investigation of network growth principles in the phonological language network" with data of English-as-a-second-language learners. Their findings were confirmed for various languages, they fit with assumptions of cognitive efficiency in lexical memory and 2 retrieval, and thus are intriguing for second language research as well. They identified a unique developmental trajectory in network growth, with high-density neighborhoods becoming enriched through growth at early acquisition stages (the "preferential attachment" mechanism) but low-density neighborhoods gaining new neighbors at advanced acquisition stages (termed "inverse preferential attachment"). We demonstrate this flexibility through solving inverse problems arising in the analysis of ordinary and partial nonlinear differential equations and, in addition, to a black-box computer model generating spatiotemporal dynamics across a network.Ī recent study by Siew and Vitevitch (2020a) investigated word form lexica and their growth in children acquiring English and Dutch as first languages from a network perspective. Our method is general, extensible, and capable of learning a wide range of dynamical systems with potential model misspecification.
Through reduced-rank Gaussian processes and a conjugate model specification, our methodology is applicable to large-scale calibration and inverse problems. This enables efficient modeling of the dynamics over space and time. Assuming the dynamical system is controlled by a finite collection of inputs, Gaussian process regression learns the effect of these parameters through a number of training runs, driving the stochastic innovations of the spatiotemporal state-space component.
The joint melding makes use of both Gaussian and non-Gaussian state-space methods as well as Gaussian process regression. Calibration is achieved by melding information from observed data with simulated computer experiments from the mechanistic system. We develop an approach for fully Bayesian learning and calibration of spatiotemporal dynamical mechanistic models based on noisy observations. Results vary by fortis/lenis articulation, with changes in lenis vot shortening and fortis vot lengthening being linked to different types of neighborhoods in two different generations of speakers. Inferences about which phonological neighborhood characteristics are most conducive to sound change are drawn. The present study investigates how two changes in voice onset time (vot) in Austrian German onset plosives have appeared in certain types of phonological neighborhoods. What has remained underrepresented in the literature to date is the question of how phonological or phonetic changes are accommodated by phonological neighborhoods, or put differently, what the implications of language processing are for language change. The significance of the phonological neighborhood on lexical processing has been documented by decades of studies in the field, and it has become clear that the phonolog-ical connectivity of the mental lexicon is a crucial facilitator for word learning in both the production and perception domains.
Cognitive network science and computer simulations may provide insight to a wide range of speech, language, hearing, and cognitive disorders. We found that the structure of the network “protects” word retrieval despite decreases in processing efficiency words that are relatively easy to retrieve with efficient transmission of priming remain relatively easy to retrieve with less efficient transmission of priming. In Study 3, we demonstrated another way to model developmental or acquired disorders by manipulating how efficiently activation spread through the network. The performance of the model was positively correlated with naming accuracy by people with aphasia (PWA) on the Philadelphia Naming Test (PNT) across four types of aphasia. In Study 2, computer simulations examined the retrieval of a set of words.
Various measures showed the network remained well-connected (i.e., it is resilient to damage) until ~90% of the connections were removed. In this study of a phonological network of English words we asked: how does damage alter the network structure (Study 1)? How does the damaged structure influence lexical processing (Study 2)? How does the structure of the intact network “protect” processing with a less efficient algorithm (Study 3)? In Study 1, connections in the network were randomly removed to increasingly damage the network. A central tenet of network science states that the structure of the network influences processing.
0 kommentar(er)
