Subjects read texts describing pairs of individuals sentence by sentence in a self-paced reading time task and then answered questions about the individuals and recalled them. The task is designed to explore how people represent the binding of attributes to individuals. The reading time data show that the construction of representations is organised around what is known about the currently referenced individual. The more that is known the more slowly subjects read. This slowing is not due to articulatory rehearsal. A regression model of reading times describes the partitioning of working memory resources across the semantic structures being processed. The construction processes yield redundant representations consisting of sets of feature values encoding aspects of the information in the text and contributing independently to memory performance. Modelling in terms of feature representations enables prediction of the patterns of error in recall. The reading time and recall error models are interpreted as showing how processes of recruiting associations occupy increasing time as more properties are known of an individual.

We present a new account of the fine-grained structure of semantic categories derived from neuropsychological, behavioral, and developmental data. The account places theoretical emphasis on the functions of the referents of concepts. We claim (i) that the distinctiveness of functional features correlated with perceptual features varies across semantic domains; and (ii) that category structure emerges from the complex interaction of these variables. The representational assumptions that follow from these claims make strong predictions about what types of semantic information are preserved in patients showing category-specific deficits following brain damage. These claims are illustrated with a connectionist simulation which, when damaged, shows patterns of preservation of distinctive and shared functional and perceptual information which varies across semantic domains. The data model both dissociations between knowledge for artifacts and for living things and recent neuropsychological evidence concerning the robustness of functional information in the representation of concepts.

Abstract The speech signal is typically continuous; only a minority of word boundaries are marked by any recognizable acoustic cue such as a pause. The continuous nature of speech poses a problem for the adult speaker of the language, in that processing the signal requires a complete parse into words yet any string of more than a few segments is locally multiply ambiguous: given only a phonetic transcription, most words contain other words, in the way that curtain contains cur, or floor contains or. Segmentation strategies are available to the adult listener that are not given to the infant, as the former possesses both a lexicon containing the phonological specification of the words of the language, and a knowledge of the syntax and semantics of the language. For instance, the adult listener may recognize a word before its acoustic offset and hence may be able to predict the end of the current word and the start of the next; indeed, this strategy featured explicitly in one early model of word recognition (Cole and Jakimik 1980 ). The adult listener may be able to recruit syntactic knowledge to predict and identify closed-class words (the short grammatical, or function words) (Shillcock and Bard 1993), and hence identify their boundaries too. The infant, faced with the speech sounds of an unknown language, is unable to draw on such knowledge, yet over the first two years of life individual words are isolated in comprehension, stored and begin to be deployed in production. This chapter is concerned with the nature of the information that the infant might exploit to obtain a foothold on the segmentation problem.

We present a new account of category structure derived from neuropsychological and developmental data. The account places theoretical emphasis on functional information. We claim i) the distinctiveness of functional features correlated with perceptual features varies across semantic domains. ii) the perceptual features representing specific functional mechanisms are strongly correlated with their function. The representational assumptions which follow from these claims make strong predictions about what types of semantic information is preserved in patients showing category-specific deficits following brain damage. We present a connectionist simulation which, when damaged, shows patterns of preservation of distinctive and shared functional and perceptual information varying across semantic domains. The data model both classic dissociations between knowledge for artefacts and for living things and recent neuropsychological evidence concerning the robustness of functional information.

Speech is continuous, and isolating meaningful chunks for lexical access is a nontrivial problem. In this paper we use neural network models and more conventional statistics to study the use of sequential phonological probabilities in the segmentation of an idealized phonological transcription of the London-Lund Corpus; these speech data are representative of genuine conversational English. We demonstrate, first, that the distribution of phonetic segments in English is an important cue to segmentation, and, second, that the distributional information is such that it might allow the infant, beginning with only a sensitivity to the statistics of subsegmental primitives, to bootstrap into a series of increasingly sophisticated segmentation competences, ending with an adult competence. We discuss the relation between the behavior of the models and existing psycholinguistic studies of speech segmentation. In particular, we confirm the utility of the Metrical Segmentation Strategy (Cutler & Norris, 1988) and demonstrate a route by which this utility might be recognized by the infant, without requiring the prior specification of categories like "syllable" or "strong syllable."

This paper describes a connectionist model of the interaction between long and short term memory. The proposed model addresses both issues of old memory utilization and novel memory consolidation. It comprises two different auto-associator network components; one for short term memory (STM) and one for long term memory (LTM). The STM network component uses a Hebbian type fast learning algorithm, while the LTM component uses the relatively slow mean field theory algorithm (Peterson and Hartman 1989).

The paper discusses the extent to which autoassociators with hidden units trained using mean field theory suffer from the sequential learning problem, the excessive forgetting of old information when trained on new information. This problem has been shown to be serious under certain circumstances for back propagation networks. The paper demonstrates that the problem can occur for mean field autoassociators but is made less serious by increasing the similarity within a training block.

Stenning, Shepherd, and Levy (1988) showed that when simple texts switch reference predictably between individuals, changes of reference neither cost reading time nor degrade memory performance. The present experiments examine the effects of unpredictable referential change. Experiment 1 demonstrates that unpredictable reference change does cost processing time, as a function of the amount known about the referent to which attention shifts. Analysis reveals a distinction between primary and secondary individuals related to referential change. It also reveals word-length effects, both decelerations and accelerations proportional to description length, which are interpreted in terms of use of the articulatory loop (Baddeley, 1986). Experiment 1 reveals involvement of primary/secondary status in the process of switching reference and shows that the word-length effects cannot be interpreted in terms of frequency. Experiment 2 strengthens support for the primary/secondary distinction and confirms the use of the articulatory loop. The present results suggest a central role for distributed information about sequence in representing complex semantic structures both in immediate and in long-term memory. Predictable switching costs no time because the transparency of the relation between surface sequence and underlying semantic structure is preserved. The distinction between primary and secondary individuals emerges with unpredictable reference because it restores this transparency.

Connectionist techniques for modeling the temporal statistics of phonemically transcribed spoken discourse as described. The aim is to investigate the limits of modeling psycholinguistic data at this prelexical level. The training data respect the frequency with which phoneme strings occur in conventional speech. The general model proposed uses a backpropagation through time learning procedure to train a network that can predict the identity of the phoneme at the next time step, identify the current one, and confirm the last five, after training on noisy data. The model eschews local representations of words and will have implications for current models of word recognition which use such representations