In the mid-1980s, the field of computational linguistics faced a crisis of complexity. Researchers were inundated with a "plethora of acronyms"—GPSG, LFG, DCG, FUG—leading to what Gerald Gazdar described as an "alphabet soup" of competing frameworks. In his tutorial survey, Gazdar cuts through this noise to identify the shared DNA of these emerging formalisms. He argues that despite their diverse origins, these systems represent a fundamental shift toward declarative, unification-based models of syntax.
The Shift from Procedural to Declarative
Before this era, the 1970s were dominated by two main traditions. In theoretical linguistics, "transformational" approaches were often informal or overly complex. In computational circles, the standard was the Augmented Transition Network (ATN), which relied on procedural languages like PIDGIN. In these systems, the grammar was effectively the program; the rules of the language were inextricably intertwined with the code used to parse them.
The "New Grammar" formalisms broke this link. Influenced heavily by the rise of the Prolog programming language and logic programming, researchers moved toward declarative semantics. In a declarative system, the notation has a clear, mathematical meaning independent of any specific parsing algorithm. This allows a linguist to describe what a language looks like, while a general-purpose engine handles how to process it.
The Five Pillars of Modern Formalisms
Gazdar identifies five defining characteristics that unify these new approaches:
- Declarative Semantics: As noted, the grammar is a set of static constraints rather than a sequence of instructions.
- Type 2 Rule Format: There is a move toward context-free-style rules (single mother nodes), avoiding the explicit context-sensitivity that plagued earlier models.
- Node Admissibility: Rather than viewing a grammar as a set of "rewrite rules" that generate strings, these formalisms treat grammar as a set of conditions that "admit" or validate specific tree structures.
- Complex Categories: Categories are no longer simple atomic labels like "Noun" or "Verb." Instead, they are recursively defined structures, often represented as Directed Acyclic Graphs (DAGs) or feature structures.
- Unification: This is the central engine. Unification is the operation used to merge syntactic information from different sources, ensuring consistency across complex categories.
Tools vs. Theories
A crucial insight in Gazdar’s survey is the distinction between the intent of different formalisms. He categorizes them into two distinct groups:
- Linguistic Theories: Frameworks like Generalized Phrase Structure Grammar (GPSG) and Lexical Functional Grammar (LFG) are designed to make empirical claims about the nature of human language and cognitive architecture.
- Computational Tools: Frameworks like Definite Clause Grammars (DCG), Functional Unification Grammar (FUG), and PATR-II are intended as computationally tractable languages for writing grammars, without necessarily claiming to be models of the human mind.
A Converging Landscape
Gazdar also maps the geographical trends of the time, identifying two primary schools of thought that began to converge in the 1980s. The European tradition was largely focused on Prolog and logic-based formalisms, such as Extraposition Grammar and DCGs. Meanwhile, the North American tradition leaned toward functional and lexical approaches, exemplified by Martin Kay’s Functional Unification Grammar and LFG.
The "Unification Revolution"—pioneered largely by Martin Kay—served as the bridge between these worlds. It provided the primary operation that allowed these disparate systems to converge on a shared mathematical foundation.
Conclusion
The survey concludes that the field is moving toward a "compilation" model. In this paradigm, high-level, expressive linguistic theories are compiled into lower-level, efficient computational formalisms for processing. While the empirical claims of frameworks like GPSG and LFG remained under scrutiny at the time of writing, the shift toward declarative, unification-based systems provided the theoretical clarity needed to move past the procedural limitations of the previous decade.
Summary updated