Natural Language Understanding Using Generative Dependency Grammar
Research and Development in Intelligent Systems XIX
This document presents a new kind of grammars: the Generative Dependency Grammar (GDG) and GDG with Feature Structure (FS) represented as Attribute Value Tree (AVT). This type of grammars is based on dependency trees (DT) and a generative process. GDG will eliminate some issues of Dependency Grammars DG (for example the missing of phrasal categories) and Generative Grammars GG (the problem of discontinuous structures) and will merge the advantages of the two types of grammar (GG -the
... ion of phrasal categories; GDG -the handling of discontinuous structures as gaps and non projective constructions). We present the generation process and the unification in GDG. Some properties of AVT and a logical representation of AVT are presented too. GDG is useful in grammar representations for natural language understanding, machine translation and data retrieval. reality). The entities names and the syntactic structures are specific to each natural language. Let us take an example. We suppose that the "cats eat mice". This is the level 0 that is formed from three entities (the cats, the mice and the fact to eat) and two relations (between the cats and the fact to eat and between the fact to eat and the mice). A representation about these three entities and two relations is formed in the mind of an observer. This is the level 1 of analysis. Finally, the observer can build the statement "Cats eat mice". This is the level 2 of analysis. In this statement we find the three entities and two relations. The syntax served to build the statement as a sequence of words. The relations and the entities founded in the level 1 of analysis served to chose the appropriate words for the level 2. We will try to find a representation for the level 1 and a method to obtain the level 1 from the level 2. The level 1 of analysis can be considered the semantic level. The relation between the level 1 and the level 0 is not a semantic problem but a truth problem (is there a level 0 corresponding to level 1?). The level one of analysis has as best representation the Generative Dependency Grammar (GDG). GDG is a sort of combination between Dependency Grammar (DG) and Generative Grammar (GG). DG (Tesnière , Gaifman , Mel'cuk  , Hellwig , Hudson , McCord , Starosta , Milward ) tries to represent entities (as words) and relations between entities (words). They use dependency trees (DT) that contains usually words as nodes and relations (oriented links between words). We will take from DG the idea of the "relation" that we consider that express somehow the semantics. GG (Chomsky ) try to express a process to obtain phrases starting from a root symbol and using some generative rules. We will take from GG the idea of the "sequence" of words that we consider that express somehow the syntax. In this paper we will present the background elements of GDG.