Professor Janet Pierrehumbert presentation at language conference in Germany

Professor Janet Pierrehumbert presentation at language conference in Germany

The Centre's Professor of Language Modelling Janet Pierrehumbert was invited to talk at the 'Abstraction, Diversity and Speech Dynamics' conference from 3-5 May, hosted by the Institute of Phonetics and Speech Processing, Ludwig-Maximilians-Universität, Germany.

The conference is part of the University's Research Focus "How Words Emerge and Dissolve". The aim of the conference is to advance discussion on how multiple layers of semiotic and signal aspects of speech are connected and how those connections may be manifested differently in the world's languages and cultures, by bringing together scientists from various disciplines engaged in research on areas such as memory and its relationship to abstraction, feedback and feedforward control systems, and modelling the association between discrete categories and continuous speech dynamics.

Professor Pierrehumbert's talk is entitled 'Learning socio-indexical features of words'. She discussed how people who speak the same language differ in what words they use and also in their pronunciation patterns. Some of these differences become conventionalised as socio-indexical, permitting listeners to make social inferences about the speaker while also grasping the content of what the speaker is saying. This situation points to cognitive processes whereby people can learn social associations for words and pronunciation patterns, and propensities to imitate others engender groups sharing linguistic behaviours. But the propensities to imitate others must be limited, as otherwise the lexical choices and speech patterns would become completely homogenised within any given speech community.

In the talk, she reviewed a series of experimental studies about the power and limitations of socio-indexical learning. These studies indicate that: — Indexical associations for words, morphological patterns, and allophonic patterns are all quite learnable. — Categorical patterns can be learned with a remarkably small number of trials. Learning such patterns is much faster than learning gradient phonetic patterns. — Only a subset of the available statistical associations are learned. Contextual relevance shapes learning by affecting what examples are remembered and how generalisations are formed. -- "Contextual relevance" derives both from widely shared assumptions about how language works, and from individual social identities.