Yet Another Sentence Processing Model of English Kei Takahashi In the theoretical syntactic literature, hierarchical structures have played a central role in accounting for various phenomena. However, this does not mean that linear order has no role to play. Given that sentences are processed from left to right, it would be rather surprising if linear order played absolutely no role when native speakers make (un)acceptability judgments. In fact, some researchers (e.g. Hawkins 1994) have attempted to explain syntactic phenomena in terms of real-time processing. This presentation reports our ongoing attempt to construct a formalized model which is based on the notion of working memory, activation value and well-formedness. Both activation value and the well-formedness value are annotated to all nodes. In this presentation, at first, we summarize the kinds of linear order effects to be accounted for by our model and pointing out that they could not be accounted for in structural terms. We will deal with certain phenomena such as toplicalization, complement-coordination, right node raising (RNR), and insertion that are able to be explained in terms of only word order or linear order. Some of the linear order effects we deal with have already been pointed out in the literature (Moosally 1996, Sadock 1998), but to the best of our knowledge there exists no formalized account of the observed linear order effects. Also, the (un)acceptability of some of those data which we deal with could not be predicted by grammar. Then we consider that these phenomena are not matter of syntax but rather real-time processing. Secondly, we illustrate the observational generalizations in processing terms. Then, we propose a formalized sentence processing model with activation value and well-formedness and discuss a relation between our model and and (un)acceptability. Activation value is represented by real number and, in our model, activation value is ``deactivated" mainly in three cases: processing time, the completeness of semantic predicate-argument structure, and insertion of overt phrase between arguments. These assumptions are supported by the previous psycholinguistic researches such as Sachs (1967) and Sturt and Lombardo (2005). On the other hand, well-formedness indicates the bias between the label of expected category and the label of category which is actually encountered to the parser. The value of well-formedness is calculated by the activation value(s) and represented by real number. In our model, (un)acceptability is determined by these two values. Fourthly, we demonstrate how our model accounts for the data in question uniformly. Then we mention some remaining problems and conclude our presentation.
|