Friday, August 30, 2019

JG-TAG

From teaching prospective natural language processing is superior due in large part to the â€Å"domain of locality â€Å"in this theory. Also it provides a brilliant framework to represent different verb classes using tag trees. TAG has always excelled in providing context sensitivity to a basic rule system and a lexicalized JG grammar implementation would allow JG structures that have previously been represented programmatically to be described in a more easily visualized and maintainable data structure format. The verb class JG-TAG trees would also simplify the lexical rules by attaching them to specific verbs and allowing them to be limited to the context of a specific verb. One of the exercises in creating such a system would involve the format of lexical rules that would be attached to the JG-TAG trees. Each JG-like rule in the tree specifies left-to-right, right-to-left or discontinuous ordering. Recall that the JG approach involves in-situ wh-elements and a specific traversal order without creating target nodes for movement. Thus the algorithm for deciding traversal would reflect, but not implement, movement. The documentation and implementation papers for the JG ordering algorithms and transfer language used in an early machine translation project could be a good starting point for a JG-TAG system (Melby 1974, Gessel 1975). Another challenge would be matching and using features attached to JG nodes with the TAG feature capabilities. TAG unification features that prevent more than one tense-bearing verb to be attached usually would be implemented by JG lexical agreement rules. However, the feature unification approach from TAG provides a straightforward manner to keep track of main and auxiliary verbs and their inflections as a sentence is created from the tree. Mandatory, optional and null adjunction constraints allow the relationships between the various TAG tree sets to be carefully defined, linked together and maintained. Expert rule systems generally need these kinds of constraints in order to assure tractable development and maintenance. These same capabilities would be very advantageous to link together JG tree fragments that would define a working grammar for a particular language. The power of the MC-TAG trees that encapsulate semantic relationships would then output not just a surface ordered derived tree but an order-independent syntax/semantics representation less dependent on th derivation tree for semantic relationships. The JG trees are not at as low a semantic level as the derivation tree but provide structure related to the original utterance (e.g. active vs. passive) and are very rich in specific syntax and semantic relationships (e.g. themes and verb classes with thematic roles (Millett, 1975)) between the concepts of the utterance. Comparative and quantifier structures have a particularly rich semantic structure in JG (Lytle 1985) and a JG-TAG system could facilitate comparison of the capabilities of a JG-based text-understanding application to other standard approaches. A JG-TAG system could also provide a standardized application and coding framework for using Junction Grammar. Conclusions As TAG formalisms have been applied to natural languages, their advantages over context-free phrase structure rules have become more apparent. Many useful re- finements to the basic TAG formalism have supported a wide variety of structures. Meanwhile JG embodies rather different assumptions than do traditional theories: a separation of linguistic data via conceptual and articulation trees, junction operators on non-terminal nodes, multiple-linked tree structures, and flexible traversal of lexical rules. The appreciable overlap of approaches with TAG and JG has prompted this discussion on combining the benefits of both theoretical systems to represent and process Junction Grammar trees. The advantages of the mildly context sensitive lexical JG-TAG system proposed in this paper can expand the domain of locality for JG trees, simplify lexical rules by attaching them to supertag class trees and draw on the extensive NLP experience using TAG based systems to benefit JG. TAG could likely also benefit from junctions, ordering, and multiple tree enhancements from Junction Grammar.

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.