A New Linguistic Engine for NooJ: Parsing Context-Sensitive Grammars with Finite-State Machines

Abstract

NooJ is a linguistic development environment that allows linguists to construct large linguistic resources of the four types in the Chomsky hierarchy. NooJ uses a bottom-up, “cascade” approach to sequentially apply these linguistic resources: each parsing operation accesses a Text Annotation Structure, and enriches it by adding or removing linguistic annotations to it. We discuss the drawbacks of this approach, and we present a new approach that requires that all NooJ linguistic resources be represented by a single type of finite-state machine. In order to do that, we must solve theoretical problems such as “how to handle Context-Sensitive Grammars with finite-state machines”, as well as some engineering problems such as “how to compose sets of large dictionaries and grammars into a single finite-state machine”. Our first experiments show that although that composing large finite-state machines is extremely costly theoretically, the fact that linguistic resources in a typical NooJ cascade depend on each other heavily keeps the size of all intermediary machines manageable. Once the final resulting finite-state machine has been compiled and loaded in memory (e.g. on a webserver) it can be used to parse large texts in linear time.