Work on GATE has been partly supported by EPSRC grants GR/K25267 (Large-Scale

This thesis defines the boundaries of Software Architecture for Language Engineering (SALE), an area formed by the intersection of human language computation and software engineering. SALE covers all areas of the provision of infrastructural systems to support research and development of language processing software. In order to demonstrate the theory developed in relation to SALE, we present the design, implementation and evaluation of GATE, a General Architecture for Text Engineering, which illustrates in practice many of the theoretical points made.

Measures of the speed andacg&U:W ofproc:)gH3fiPW&gcc with nonadjacgH dependencH3 derived from the response -signal speed-acgnalg tradeo#proco#g& were used to examine the nature of the memory system that underlies sentenc ctencesg):WP Three experiments with di#erent sentenc strucg9P demonstrated that theacg))W& of procWPPgH3 dependenc decndenc as more material was interpolated betweennonadjac33 cnadjac33(g However, procer,g):WPWg wasuna#ecUP by the amount of interpolated material,indicl,g39UPfi memory representations for previously procious ciouslygfi9 cfi beac9((fi9 direcfi9g These results suggest that acfi99WgH339fi)gcgc memory system mediates sentenc ctenc&(3:WgH inwhic syntac9g andsemantic information providedirec acec to memory representations without the need tosearc through extraneous representations. Notably,cably,g3Pfi3)gH3&(:&gc appears to underlie the interpretation ofsentenc strucgH9 that also require therec33&U of order information, a type of operation that has been shown tonec:&(gH9fi a slowsearc proc3 in list-learningexperiments(Mct-lea 2001;Mc1;g) & Dosher, 1993).

We describe a technique for the robust interpretation of newswire texts which uses semantic role information about verb complements together with a general co-reference mechanism to extend the constituent structure analysis produced by a partial parser. This technique has the advantage that failure to find a spanning parse of an entire sentence does not necessarily preclude correct semantic interpretation of, for example, key subject-verb-object relations. An information extraction system employing this technique has been evaluated in the Sixth Message Understanding Conference (MUC-6), and while the scoring protocols in that exercise do not allow a direct assessment of the technique, we can use them to obtain indirect performance measures which give some indication of how much the technique is contributing to overall system performance. 1 Introduction 1.1 Terms of Reference In this paper we describe a technique for the robust interpretation of newswire texts which uses semantic role ...

and by majority vote has been found to be satisfactory.

plex (the feature structures 2 and 3 ). If we take FSs to be sets of paths and values, we will characterize (1) by saying, e.g. that the value of the path head j agr j num is sing. If we take them as functions, we will say (where f 1 means the feature structure tagged 1 above, and (f 1 cat) means the application the function f 1 to the argument cat). (2) a. (f 1 cat) = s b. (f 2 agr) = (f 3 agr) c. (f 2 agr) = f 4 d. ((f 1 head) agr) = (f 2 agr) If we take them to be DAGs we will think of them as follows: 1 (3) 1 subj 2 cat np agr cat s head 3 agr 4 num sing per 3 cat vp Thinking

in one or more steps). Similarly, we use S ) abc to mean S derives abc in zero or more steps: A ) A derives in one step A + ) A derives in one or more steps A ) A derives in zero or more steps We write L(G) for the language generated by the grammar G. A string of terminal elements ! is in L(G) if S ) !, where S is the start symbol of G. If is a string over T [ NT , and S ) , then is a sentential f

Many natural languages exhibit a much higher degree of freedom than English in the ordering of constituents within a clause. In order to use these so-called `free word order' languages in natural language processing applications, we need grammar models that are adequate from both a linguistic and a computational point of view. I examine Slovene, one such language, and propose that it is best treated by flattening the traditional hierarchical syntactic structure. I argue that there is little empirical evidence for the finite verb phrase constituent in the clause, and I show that several problems disappear if this assumption is rejected. Instead, I present a model whereby the verb combines with its subject and other complements in a single step. In both finite and nonfinite clauses, there is only one verbal projection in the syntactic structure. The clause structure is described within the constraint-based formalism of Headdriven Phrase Structure Grammar (HPSG) which is not confined to b...