On modularity and compilation in a government-binding parser
Paola Merlo · 1992
This thesis describes a parser for English that implements Government Binding (GB) theory, adopting a design which is based on the structure of the principles of the theory. The main question addressed in the thesis can be formulated as follows: given a certain theory of grammar, in this instance GB, what are the design principles for a parser which maintain the explanatory power of the theory, in a computationally efficient way? The present work contributes to the empirical solution of this problem by making the observation that linguistic principles belong to five main classes. These classes are defined according to their informational content, for example topological properties of the tree or lexical information. Linguistic support for this partitioning is provided by facts related to word order. The classes defined in this way are taken to be the level that the parser must directly mirror to implement the theory of grammar. This hypothesis is then tested by implementing a parser. The data structures and the architecture of the parser mirror the partitioning of linguistic principles according to their informational content. The hypothesis is confirmed by the following results. Computationally, this organization of the parser is compact and non-redundant: the parser is implemented as an LR parser which is encoded in only a small number of states; information about category and other lexical properties are encoded in a different table which interacts with the LR table on-line. Moreover, a complex phenomenon, long-distance dependencies, can be computed efficiently, by making use of information available in the local context of the application of parsing rules. We also show how this design could be extended to other languages: algorithms are provided for long distance, cyclic movement in Italian and English. Finally, we argue that, psycholinguistically, the proposed design captures some experimental evidence about the interaction of lexical ambiguity with structural ambiguity.