|
A dynamic tree adjoining grammar as the basis for the human parsing mechanism Vincenzo Lombardo In this paper, we present a syntactic formalism and a principle of semantic interpretation that yield interesting predictions for sentence processing complexity. The syntactic formalism is called DVTAG (Dynamic Version of Tree Adjoining Grammar). DVTAG is a dynamic grammar which shares with standard TAG the formal operations of substitution and adjunction, but, in contrast with standard TAG, these operations are defined so as to allow fully connected left-to-right grammatical derivations in the formal system. This allows for a very close correspondence between the operations of the formal system and those of human parsing models, which are also typically highly incremental.Being a formalism, DVTAG does not directly provide a full theory of human parsing. For example, it is neutral with respect to the parallel/serial and modular/interactive dimensions. The formalism is rather intended as a framework within which theories of incremental processing can be stated. We will however discuss three aspects of the formalism which are of particular theoretical interest to the AMLaP audience. These are (1), the large domain of locality afforded by TAG elementary trees, allowing a direct implementation of fully- connected parsing, (2) the TAG adjunction operation, allowing for the incremental processing of post-modifiers and left-recursive structures, and (3), the linguistic motivation of TAG elementary trees in terms of lexical anchors and predicate argument structures, offering a natural characterisation of predictive processing behaviour. The Principle of Semantic Interpretation (PSI) describes how the syntactic structure is related to predicate-argument structure, leading to predictions for processing complexity. The PSI is related to Joshi's (1989) Principle of Partial Interpretation, which was expressed in terms of an embedded pushdown automata. The difference with Joshi is that our principle acts on fully connected syntactic structures, which are augmented with head information. The complexity of a sentence is proportional to the number of constituents that are not part of a predicate-argument structure, i.e. are in unheaded constituents. This accounts for the lower perceived complexity of cross-serial dependencies in Dutch, compared with the equivalent nested dependencies in German (Bach et al, 1986). Crossed dependencies have the following structure NP1 NP2 NP3 V1 V2 V3 When the verb V1 is inserted into the structure, it provides a predicate-argument structure for (NP2,V2) and V2 provides a predicate-argument structure for (NP3,V3). At each step, the completed structure can be interpreted immediately within the larger predicate-argument structure. On the contrary, nested dependencies have the structure NP1 NP2 NP3 V3 V2 V1 When V3 is inserted into the structure, the substructure (NP3,V3) has no predicate where it can be interpreted. The same holds for (NP2,V2). The head of the larger constituent comes later, leading to greater processing load. A similar motivation holds for center-embedding clauses in comparison with right and left branching structures (Abney, Johnson 1991). |