Language learning is a grand challenge problem for Artificial Intelligence because it encompasses concept development and perceptual development, social learning and imitation, as well as learning the lexicon, the grammar, and other aspects of language; because it drives new technologies that apply widely to other kinds of sequential data; and because most of the world's knowledge is represented linguistically, so machines are limited by their inability to understand language.
The symposium is intended to bring together representatives of several communities --- the corpus-based and grounded language learning communities, and the developmental psycholinguistics and language education communities --- to assess progress in machine language learning and how what we know about human linguistic development might speed that progress.
Three kinds of interdisciplinary discussions are likely to be productive. In grounded language learning, language describes a present scene and is often learned in a language game of some sort with a competent language user. Corpus-based approaches work with corpora of language dissociated from a present scene and not generated in a language game that includes the learner. Learning rates may be higher for grounded language learning; corpus-based approaches may learn a wider range of word classes, including words with abstract semantics that do not refer to a present scene. Both approaches are inherently statistical and much can be shared between the practitioners of each. A second integration is between lexical acquisition and grammatical inference. Knowing word meanings can help one acquire grammatical rules, and the assignment of words to grammatical categories should help acquire their meanings. A third discussion is between language learning researchers and those who work on large, commonsense knowledge bases. Language is layered on a conceptual system, and depends on that system for its interpretation; and language conveys new concepts and distinctions; so language learning both depends on and extends commonsense knowledge.
Submit abstracts (2 - 4 pages) to Paul Cohen (cohen@isi.edu) by October 3, 2003. Abstracts that integrate approaches to machine language learning, that inform these approaches with knowledge and theories of human language learning, and that describe empirical results, are most welcome; though other kinds of abstracts will be considered.
Organizing committee: