R. J. Mammone
Rutgers Univ., CAIP Ctr., Core Bldg., Frelinghuysen Rd., Piscataway, NJ 08855
A. L. Gorin
AT&T Bell Labs., Murray Hill, NJ 07974
A new iterative approach to adaptive language acquisition is presented. Prior methods for adaptive language acquisition have used statistical associations to determine the connection between words and actions. However, the performance of these methods degrades when dealing with small-sample statistics. An alternative approach for language acquisition has been recently developed, which circumvents the problems of small-sample statistics. This alternative approach consists of modeling the language acquisition problem as a linear system of equations (i.e., algebraic learning), where each equation represents a sentence. The pseudoinverse solution for this linear system of equations, as obtained by a singular value decomposition (SVD), has been found to provide connection weights that are less subject to the effects of small-sample statistics. However, the batch-mode form and computational requirements of the SVD make it unsuitable for adaptive on-line language acquisition. A logical method for overcoming this limitation is to iteratively solve the linear system of equations. An iterative approach based on this concept is now presented. The iterative approach consists of solving for the connection weights on a ``sentence-by-sentence'' basis as opposed to requiring a block of sentences as in the SVD approach. The iterative approach is shown to converge to a pseudoinverse solution. The iterative approach provides significant improvement over the SVD approach due to the availability of on-line training and in addition provides improved generalization. Techniques for iterative training and simulation results are discussed.