摘要：In generative grammar the competence of an ideal speaker is related to a set of language invariant but parametricizable principles. This knowledge is represented in a multilevel theory, where the levels build a hierarchy in which the different stages are connected by wellformedness conditions on structural descriptions. A competence based theory in that sense requires a view that focuses on the abstract knowledge a speaker has about complete sentences. From the point of view of an on-line operating performance system knowledge of language has to be applied not only to complete sentences but far more often to parts of the phonetic/graphematic input in a timelinear fashion. The assumption that language processing is on-line implies that the structural descriptions for a given sentence at any level have to be computed simultaneously. Given the view that syntactic structures are projected from the lexicon the structure of lexical items plays a main role in determining the resulting X-bar structures. The performance system therefore needs a specific procedure where syntactic structures are projected by these lexical items. Especially in the case of head final constructions, the performance system needs some kind of memory in order to store constituents that have no clearly defined status until the head (lexical or functional) of the projection they are linked to appears in the course of the time sequential signal. Furthermore, the performance system has to have a linking module which joins these structurally undetermined projections. In this paper I will show how a linking system based on government directions and saturation principles works in an LR (0) parser. This is a subtype of the class of deterministic push down automata which parse an input string from left to right without using any look-ahead ability. This means that the parser processes one word at a time without any need to see elements following in the input string.
