Search

In this study, results of computational simulations on Englishchild-directed speech are presented to uncover what distribu-tional properties of words make it easier to group them intolexical categories. This analysis provides evidence that wordsare easier to categorize when (i) they are hard to predict giventhe contexts they occur in; (ii) they occur in few different con-texts; and (iii) their contextual distributions have a low entropy,meaning that they tend to occur more often in one of the con-texts they occur in. This profile fits that of content words, espe-cially nouns and verbs, which is consistent with developmentalevidence showing that children learning English start by form-ing a noun and a verb category. These results further charac-terize the role of distributional information in lexical categoryacquisition and confirm that it is a robust, reliable, and devel-opmentally plausible source to learn lexical categories.

Previous studies have suggested that children possess cognitiverepresentations of multi-word units (MWUs) and that MWUscan facilitate the acquisition of smaller units contained withinthem. We propose that the formation of MWU representationsprecedes and facilitates the formation of single-word represen-tations in children. Using different computational methods,we extract MWUs from two large corpora of English child-directed speech. In subsequent regression analyses, we use ageof first production of individual words as the dependent and thenumber of MWUs within which each word appears as an in-dependent variable. We find that early-learned words appearwithin many MWUs – an effect which is neither reducible tofrequency or other common co-variates, nor to the number ofcontext words contained in the MWUs. Our findings supportaccounts wherein children acquire linguistic patterns of vary-ing sizes, moving gradually from the discovery of MWUs tothe acquisition of small-grained linguistic representations.

We test the predictions of different computational models ofcross-situational word learning that have been proposed in theliterature by comparing their behavior to that of young childrenand adults in the word learning task conducted by Ramscar,Dye, and Klein (2013). Our experimental results show that aHebbian learner and a model that relies on hypothesis testingfail to account for the behavioral data obtained from both pop-ulations. Ruling out such accounts might help reducing thesearch space and better focus on the most relevant aspects ofthe problem, in order to disentangle the mechanisms used dur-ing language acquisition to map words and referents in a highlynoisy environment.

Do children acquire rules for main stress assignment or do they learn stress in an exemplar-based way? In the language acquisition literature, the former approach has been advocated without exception: although they hear most words produced with their appropriate stress pattern, children are taken to extract rules and do not store stress patterns lexically. The evidence for a rule-based approach is investigated and it will be argued that in the literature this approach is preferred due to an extremely simplified interpretation of exemplar-based models. W e will report experiments showing that Instance-Based Learning, an exemplar-based model, makes the same kinds of stress related errors in production that children make: (i) the amount of production errors is related to metrical markedness, and (ii) stress shifts and errors with respect to the segmental and syllabic structure of words typically take the form of a regularization of stress patterns. InstanceBased Learning belongs to a class of Lazy Learning algorithms. In these algorithms, no explicit abstractions in the form of decision trees or rules are derived; abstraction is driven by similarity during performance. Our results indicate that at least for this domain, this kind of lazy learning is a valid alternative to rule-based learning. Moreover the results plead for a reanalysis of language acquisition data in terms of exemplar-based models.

This paper investigates the computationalgrounding of learning theories developedwithin a metrical phonology approach tostress assignment. In current researchthe Principles and Parameters approachto learning stress is pervasive. W e pointout some inherent problems associated withthis approach in leaxning the stress sys-tem of Dutch. T h e paper focuses on twospecific aspects of the learning task: w eempirically investigate the eflFect of inputencodings on learnability, and w e exam-ine the possibility of a data-oriented ap-proach as an alternative to the Principlesand Parameters approach. W e show thata data-oriented similarity-based machinelearning technique (Instance-Based Learn-ing), working on phonemic input encodingsis able to learn metrical phonology abstrac-tions based on concepts like syllable weight,and that in addition, it is able to extractgeneralizations which cannot be expressedwithin a metrical framework.