Abstract not found

Since the introduction of hidden Markov modelling there has been an increasing emphasis on data-driven approaches to automatic speech recognition. This derives from the fact that systems trained on substantial corpora readily outperform those that rely on more phonetic or linguistic priors. Similarly, extra training data almost always results in a reduction in word error rate- “there's no data like more data”. However, despite this progress, contemporary systems are not able to fulfill the requirements demanded by many potential applications, and performance is still significantly short of the capabilities exhibited by human listeners. For these reasons, the R&D community continues to call for even greater quantities of data in order to train their systems. This paper addresses the issue of just how much data might be required in order to bring the performance of an automatic speech recognition system up to that of a human listener. 1.

If human behavior were list-like, accounting for human behavior would be simple: Just enumerate the list of possible stereotypies. Alternatively, if behavior were predictable on the basis of abstract, fully-productive, context-insensitive rules, our task would be different but similarly straightforward: just list the underlying rules.

Introduction The acquisition of word meaning has been extensively studied for the last thirty years in the field of language acquisition. However, the question of how children acquire word meaning remains highly controversial today. Recently, a number of computational studies have examined the emergence of lexical representations in connectionist networks or similar statistical systems, suggesting that word meaning can be acquired by the computation of statistical regularities inherent in the input data. In particular, Elman (1990, 1998) showed that categories of nouns and verbs, and subcategories of animates versus inanimates (within nouns), and transitives versus intransitives (within verbs), can emerge from the network's computing of the lexical co-occurrence properties in the input. Redington, Chater, and Finch (1998) also demonstrated that the use of distributional properties in large-scale speech corpus allows a statistical system to acquire syntactic categories. These s

The central issues in the study of speech recognition by human listeners (HSR) and of automatic speech recognition (ASR) are clearly comparable; nevertheless the research communities that concern themselves with ASR and HSR are largely distinct. This paper compares the research objectives of the two fields, and attempts to draw informative lessons from one to the other. 1.

Abstract not found

What is the role of the linguistic environment in children’s early word learning? Here we provide a preliminary analysis of one child’s linguistic development, using a portion of the high-density longitudinal data collected for the Human Speechome Project. We focus particularly on the development of the child’s productive vocabulary from the age of 9 to 24 months and the relationship between the child’s language development and the caregivers ’ speech. We find significant correlations between input frequencies and age of acquisition for individual words. In addition, caregivers ’ utterance length, type-token ratio, and proportion of single-word utterances all show significant temporal relationships with the child’s development, suggesting that caregivers “tune ” their utterances to the linguistic ability of the child.

Based on the apparent paucity of input, and the nonobvious nature of linguistic generalizations, Chomskyan linguists assume an innate body of linguistically detailed knowledge, known as Universal Grammar (UG), and attribute to it principles required to account for those “properties of language that can reasonably be supposed not to have been learned ” (Chomsky, 1975). A definitive account of learnability is lacking, but is implicit in examples of the application of the logic. Our research demonstrates, however, that important statistical properties of the input have been overlooked, resulting in UG being credited for properties which are demonstrably learnable; in contradiction to Chomsky’s celebrated argument for the innateness of structure-dependence (e.g. Chomsky, 1975), a simple recurrent network (Elman, 1990), given input modelled on child-directed speech, is shown to learn the structure of relative clauses, and to generalize that structure to subject position in aux-questions. The result demonstrates that before a property of language can reasonably be supposed not to have been learned, it is necessary to give greater consideration to the indirect positive evidence in the data and that connectionism can be invaluable to linguists in that respect.

This thesis addresses the cognitive basis of syntactic adaptation, which biases speakers to repeat their own syntactic constructions and those of their conversational partners. I address two types of syntactic adaptation: short-term priming and longterm adaptation. I develop two metrics for syntactic adaptation within a speaker and between speakers in dialogue: one for short-term priming effects that decay quickly, and one for long-term adaptation over the course of a dialogue. Both methods estimate adaptation in large datasets consisting of transcribed human-human dialogue annotated with syntactic information. Two such corpora in English are used: Switchboard, a collection of spontaneous phone conversation, and HCRC Map Task, a set of task-oriented dialogues in which participants describe routes on a map to one another. I find both priming and long-term adaptation in both corpora, confirming well-known experimental results (e.g., Bock, 1986b). I extend prior work by showing that syntactic priming effects not only apply to selected syntactic constructions that are alternative realizations of the same semantics, but still hold when a broad

Prior research has shown that people can use the cooccurrence statistics of words and referents in ambiguous situations to learn word meanings during a brief training period. The present studies investigate the effects of allowing some words and referents to appear more often than others, as is true in real learning environments. More frequent wordreferent pairs are often—but not always—learned better, and also boost learning of other pairs. Superior learning for training sets with varying pair frequency may be a result of learning frequent pairs first, and using this knowledge to reduce ambiguity in later trials to learn other items. However, contextual diversity – the number of other pairs a given pair appears with – is naturally confounded with frequency, and presents an alternative explanation. The experiments in the present study systematically manipulate three critical factors in cross-situational learning – frequency, contextual diversity, and within-trial ambiguity – and measure their individual and combined effects on statistical word learning.