Unification Grammars: A Unifying Approach

Abstract

Unification-based grammar formalisms have recently received great interest in computational linguistics, because they allow for a very high-level, declarative description of grammatical relations over linguistic objects. Feature structures are used to encode all kinds of information about any given linguistic object, e.g., certain syntactic properties or semantic content. Grammar rules define the relation between surface strings of a language and the information associated with it. Special purpose theorem provers — parsers and generators — operate on these grammar rules to compute this relation in either direction, from a string to a feature structure or vice versa.

Recent formalisms, like the STUF formalism developed and used within the LILOG project at IBM Germany, go even one step further: surface strings are made an integral part of informational structures and general relations can be defined on them. There is no longer a special notion of ‘grammar rule’ relating surface strings to feature structures. Consequently, there is also no parser or generator. A very general inference mechanism evaluates the relational specification when given an input structure representing, e.g., a string. Aspects of controlling the inference process thus become a crucial research topic.

Work on the STUF formalism revealed a lot of interesting similarities between some of the most active research areas in AI: computational linguistics, constraint logic programming, functional programming, and knowledge representation systems.