Comments on Roche

As the original announcement for this Workshop observed, 1996 marks the fortieth anniversary of Chomsky's demonstration of the non-regularity of natural language | more specifically natural language syntax. 1 At rst glance the computational import of this result may seem obvious: clearly one could not construct a nite-state automaton capable of handling all sentences of a natural language. Yet from the point of view of building practical systems with wide coverage, the import is actually far from obvious. For even if language were regular, representing all of the legal sentences of a language via a single nite-state network would be impossible, because the size of the network would be astronomical. As a result, one only nds grammars modeled by single nite-state networks in trivial applications, such as limited-domain small vocabulary speech recognizers, where one can severely limit the set of sentences. Wide-coverage nite state syntactic analyzers must necessarily adopt other strategies. Apart from Roche's contribution, two other approaches to wide-coverage nite-state syntactic analysis come to mind. One is the Finnish nite-state Intersection Grammar work (e.g., 3]); the other is work on Local Grammars (e.g. 1]). 2 Despite superrcial diierences, all these approaches are related in that they each involve the intersection of a text automaton, containing the possible lexical analyses of an input sentence, with a series of nite-state machines which encode syntactic constraints. This method eeectively allows one to sidestep the question of whether language as a whole is nite-state; instead , one performs syntactic analysis with a lot of little constraints , each one of which is nite-state. The parallel between such computational approaches to syntax, and theoretical approaches under the general \principles and parameters" rubric is striking: in both cases complex systems of rewrite rules have been abandoned in favor of a scheme whereunder there is a massive overgeneration of output structures, which are then ltered by a series of simple, often lexical, constraints. Roche's approach diiers from Intersection Grammars and Local Grammars in one important respect: since he uses FSTs, rather than FSAs, he is able to build structure, as well as restrict it. The primary function of Intersection Grammars and Local Grammars is to weed out possible analyses, and while 1 This commentary has beneeted from discussions with Mehryar Mohri. 2 Of course, all of these approaches, and especially Roche's, owe some debt to the early work of Woods on ATNs 4] such analyses may …