As far as we know, no single kind of cue carries sufficient information to enable a language to be successfully learnt, so some kind of cue integration seems essential. This talk uses computational models to study how a diverse range of information sources can be exploited in word learning. I describe a non-parametric Bayesian framework called Adaptor Grammars, which can express computational models that exploit information ranging from stress cues through to discourse and contextual cues for learning words. We use these models to compare two different approaches a learner could use to acquire a language. A staged learner learns different aspects of a language independently of each other, while a joint learner learns them simultaneously. A joint learner can take advantage of synergistic dependencies between linguistic components in ways that a staged learner cannot. By comparing “minimal pairs” of models we show that there are interactions between the non-local context, syllable structure and the lexicon that a joint learner could synergistically exploit. This suggests that it would be advantageous for a language learner to integrate different kinds of cues according to Bayesian principles. We end with a discussion of the broader implications of a non-parametric Bayesian approach, and survey other applications of these techniques.
Mark Johnson is a Professor of Language Science (CORE) in the Department of Computing at Macquarie University, and is Director of the Macquarie Centre for Language Sciences. He received a BSc (Hons) in 1979 from the University of Sydney, an MA in 1984 from the University of California, San Diego and a PhD in 1987 from Stanford University. He held a postdoctoral fellowship at MIT from 1987 until 1988, and has been a visiting researcher at the University of Stuttgart, the Xerox Research Centre in Grenoble, CSAIL at MIT and the Natural Language group at Microsoft Research. He has worked on a wide range of topics in computational linguistics, and is mainly known for his work on syntactic parsing and its applications to text and speech processing. Recently he has developed non-parametric Bayesian models of human language acquisition. He was President of the Association for Computational Linguistics in 2003 and will be President of ACL’s SIGDAT (the organisation that runs EMNLP) in 2015, and was a professor from 1989 until 2009 in the Departments of Cognitive and Linguistic Sciences and Computer Science at Brown University.