Mechanisms for Sentence Processing

this paper we argued that one of the most salient properties of the human sentence parser was that it appears to operate incrementally. Both of the above algorithms are incremental in the sense that they process each word of the sentence one-by-one from left-to-right. But that only means they treat the input incrementally. If we assume, quite reasonably, that the output of the parser is a parse tree, and that this is to be used for subsequent semantic interpretation, then the two models differ substantially. In the case of the top-down parser there is a single `connected' parse tree for the sentence after each word is processed (although it is not complete, until the end of the parse). In the bottom-up parser, however, adjacent constituents may be left on the stack for an arbitrarily long period. For example, the NP subject of a sentence will not be reduced with the VP (via the rule S \Gamma! NP V P ) until after the entire VP has been parsed (in the above example, the VP only contains the verb, but typically it will also contain a number of complements). Under the standard assumption that we cannot begin to evaluate the semantics and plausibility, until these constituents are connected in the parse tree, the bottom-up parser will lead to a psychologically implausible delay in interpretation for all sentences (see Stabler (1991) for discussion). So from the standpoint of incrementality, the top-down parser fares much better than the bottom-up parser. The top-down parser, however, has problems of its own. To begin with, it attempts to construct large portions of the tree before even looking at the words in the sentence. In other words, given several (possibly dozens of) rules that expand a particular category, the parser simply has to choose one arbitrarily. Given that i..