Modeling Infant Word Segmentation

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 29–38, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics Modeling Infant Word Segmentation Constantine Lignos Department of Computer and Information Science University of Pennsylvania lignos@cis.upenn.edu Abstract While many computational models have been created to explore how children might learn to segment words, the focus has largely been on achieving higher levels of performance and exploring cues suggested by artificial learning experiments. We propose a broader focus that includes designing models that display prop- erties of infants’ performance as they begin to segment words. We develop an efficient bootstrapping online learner with this focus in mind, and evaluate it on child-directed speech. In addition to attaining a high level of perfor- mance, this model predicts the error patterns seen in infants learning to segment words. 1 Introduction The last fifteen years have seen an increased inter- est in the problem of how infants learn to segment a continuous stream of speech into words. Much of this work has been inspired by experiments with in- fants focusing on what capabilities infants have and which cues they attend to. While experimental work provides insight into the types of cues infants may be using, computational modeling of the task provides a unique opportunity to test proposed cues on rep- resentative data and validate potential approaches to using them. While there are many potential approaches to the problem, a desirable solution to the problem should demonstrate acceptable performance in a simula- tion of the task, rely on cues in the input that an infant learner is able to detect at the relevant age, and exhibit learning patterns similar to those of in- fant learners. Most work in computational model- ing of language acquisition has primarily focused on achieving acceptable performance using a sin- gle cue, transitional probabilities, but little effort has been made in that work to try to connect these learn- ing solutions to the actual learning patterns observed in children outside of performance on short artificial language learning experiments. In this work we present a simple, easily extended algorithm for unsupervised word segmentation that, in addition to achieving a high level of performance in the task, correlates with the developmental pat- terns observed in infants. We discuss the connec- tions between the design and behavior of our algo- rithm and the cognitive capabilities of infants at the age at which they appear to begin segmenting words. We also discuss how our technique can easily be ex- tended to accept additional cues to word segmenta- tion beyond those implemented in our learner. 2 Related Work As this paper examines the intersection of infants’ capabilities and computational modeling, we discuss work in both domains, beginning with experimental approaches to understanding how infants may per- form the task of word segmentation. 2.1 Infant Word Segmentation A potential account of how infants learn to iden- tify words in fluent speech is that they learn words in isolation and then use those words to …

Content truncated to 500 words