Tech Report CS-91-14

A Structured Neural Network Approach to Robust Parsing

Eugene Santos Jr.

February 1991


Practical natural language interfaces must exhibit a robustness in dealing with ungrammatical data. Existing approaches considered ungrammaticality as some form of ``noisy'' data. Extensions were then made to existing language systems to handle some particular type or class of ungrammaticality. Most of the original systems were built from traditional symbolic approaches that are poorly suited for handling noise. The resulting augmented language systems were generally inefficient and cumbersome. Connectionism offers an alternative approach to this problem with its natural ability to handle noise. However, very few successful connectionist systems have been built that can adequately model complex tasks such as parsing. In this paper, we present a robust parser built from the precepts of structured neural networks that combines the desirable properties of traditional symbolic approaches and the connectionist approaches.

(complete text in pdf)