Jeffrey L. Elman - US grants

Many of the most important things people learn are learned without explicit instruction; often, people are not even aware that they are learning something. One of the most dramatic examples of this phenomenon is language. Virtually all speakers of a language learn to speak and understand before formal instruction even begins, and the learning process seems to be largely unaffected by attempts of adults to guide it. This style of learning contrasts with the more explicit learning which also occurs when people consciously attempt to master a domain, and which typically involves explicit formulation and testing of hypotheses. Although considerable research has been carried out on explicit learning, implicit learning has been investigated only recently. Furthermore, most of this research has been experimental, with relatively little work focussed on developing theories or computational models. The purpose of this research is to develop both a theory and a testable computational model of implicit learning of sequential behaviors. The research will use experimental techniques to ask the following questions: Under what conditions is implicit learning triggered? Are there domains in which implicit learning is more effective than explicit learning? What are the constraints on the types of things which can be learned with implicit learning? A second component of the research will focus on trying to understand the possible mechanism for implicit learning by developing a computational model. This model, based on an artificial neural network architecture proposed by Elman, has been shown to have interesting properties which resemble the behaviors of humans engaged in implicit learning tasks. Elaboration of this model should allow for better understanding of the circumstances which facilitate learning of domains such as language and might make it possible to structure training environments in order to maximize learning. This work should also provide a foundation for the construction of machine-based systems for learning domains currently only well-mastered by humans.

DESCRIPTION (provided by applicant): The broad aim of this research program is to develop a deeper understanding of the representations, mechanisms, and computations that underlie people's ability to comprehend and produce language. Specifically, this research will focus on expectancy generation in sentence comprehension, with particular attention given to structural ambiguity resolution and thematic role assignment. We focus on expectancy generation because there is empirical evidence that it plays a key role in language processing, and because studying a comprehender's expectations as s/he hears or reads a sentence provides a valuable diagnostic for addressing three central questions in sentence processing: what information is available to the comprehender; when do different sources of information become available; and how do classes of information interact. Theoretically, our perspective reflects an emphasis on early information use, nonlinear interactions among knowledge sources, and the importance of both event-based semantic knowledge and statistical patterns of language usage, all of which are characteristics of constraint-based approaches and connectionist models. Our research methodology involves a combination of computer simulations, corpus analyses, and human experiments, including extensive norming procedures and on-line methodologies. Five specific areas will be studied: (1) the role of verb meaning in the generation of expectations regarding upcoming subcategorization frames; (2) the effect of verb-specific semantic distributions of arguments on subcategorization preferences; (3) developing a precise definition of plausibility by comparing six operational definitions in terms of their efficacy for predicting subjects' behavior; (4) the role of event structure and grammatical cues in generating expectations about thematic role assignment; and (5) the consequences of viewing expectations in terms of dynamics in semantic space, rather than as lists of possible lexical items.

Almost as remarkable as the ability to know one language is the ability to learn a second. However, current understanding of how the human brain acquires, organizes, and processes multiple languages is highly limited. A fundamental issue is whether multiple languages are represented in common brain areas. When monolinguals read, words enter the cortex at its posterior tip, and work their way forward. Within two-tenths of a second, the words are encoded visually, and then in the next three-tenths of a second they are encoded for meaning in specialized areas of the left temporal and frontal lobes. With support from the National Science Foundation, Dr. Eric Halgren and Dr. Jeffrey Elman of the University of California at San Diego, with their colleagues, will study word encoding in young adults reading words in their native Spanish, or in their second language, English, which most have been using since they started school. Cortical neurons process information using electrical currents, which in turn produce minute magnetic fields. These will be detected using arrays of superconducting quantum interference devices as the magnetoencephalogram, and then mapped to particular cortical areas using magnetic resonance imaging. Experiments will determine if the English and Spanish words are encoded in the same areas. Specifically, experiments will test a model that hypothesizes that the second language does engage the same areas as the first, but in addition accesses the corresponding areas in the right hemisphere. Experiments will also attempt to confirm the suggestion that the second language uses brain areas which are otherwise engaged in high level vision, and explore if the second language is characterized by perceptual representations of words. Experiments will test how early in the processing stream English and Spanish words diverge, and whether this divergence is due to top-down strategic control or quick categorization. Subjects will vary in how well they know English and when they began learning it, so that the effects of age of acquisition, order of acquisition, and proficiency can be determined.

This research brings together neural, cognitive, language and imaging sciences, providing interdisciplinary training, especially for bilingual Spanish-English students, and introducing neuroimaging to a societal problem previously approached mainly behaviorally. Over two-thirds of the global population is multi-lingual. Bilingualism binds together the multiple American subcultures, being essential for socioeconomic integration of recent immigrants, and for the access of American goods and services to a global marketplace. Yet, second language education is often ineffective. A clearer picture of how the brain organizes multiple languages will form part of the scientific basis for developing effective second language teaching methodologies.