A learning perspective on the emergence of abstractions: The curious case of phone(me)s

In this study we propose an operationalization of the concept of emergence which plays a crucial role in usage-based theories of language. The abstractions linguists operate with are assumed to emerge through a process of generalization over the data language users are exposed to. Here, we use two types of computational learning algorithms that differ in how they formalize and execute generalization and, consequently, abstraction, to probe whether a type of language knowledge that resembles linguistic abstractions could emerge from exposure to raw data only. More specifically, we investigated whether a phone, undisputedly the simplest of all linguistic abstractions, could emerge from exposure to speech sounds using two computational learning processes: Memory-Based Learning (MBL) and Error-Correction Learning (ECL). Both models were presented with a significant amount of pre-processed speech produced by one speaker. We assessed (a) the consistency or stability of what these simple models learn and (b) their ability to approximate abstract categories. Both types of models fare differently regarding these tests. We show that only ECL models can learn abstractions and that at least part of the phone inventory and its grouping into traditional types can be reliably identified from the input.