How Ideal Are We? Incorporating Human Limitations into Bayesian Models of Word Segmentation

1. Introduction Word segmentation is one of the first problems infants must solve during language acquisition, where words must be identified in fluent speech. A number of weak cues to word boundaries are present in fluent speech, and there is evidence that infants are able to use many of these, including phonotactics However, with the exception of the last cue, all these cues are language-dependent, in the sense that the infant must know what some of the words of the language are in order to make use of the cue. For example, in order to know what the common stress pattern is for words of her native language, an infant has to know some words already. Since the point of word segmentation is to identify words in the first place, this seems to present a chicken-and-egg problem. Statistical learning has generated a lot of interest because it may be a way out of this problem, by providing an initial language-independent way to identify some words. Since infants appear to use statistical cues earlier than other kinds of cues (Thiessen & Saffran, 2003), statistical learning strategies could indeed be providing an initial bootstrapping for word segmentation. Statistical learning is often associated with transitional probability (Saffran et al., 1996), which has been shown to perform poorly on realistic child-directed speech (calculated over syllables: Gambell & Yang (2006); calculated over phonemes: Brent (1999)). However, a promising alternative approach is Bayesian learning. Researchers have recently shown that Bayesian model predictions are consistent with human behavior in various cognitive domains, including language acquisition (e. (henceforth GGJ) found that Bayesian learners performed very well on the word segmentation problem when given realistic child-directed speech samples, especially when compared to transitional probability learners. One critique of GGJ's model is that it is an " ideal learner " or " rational "