Introduction
Relational Grammar (RG) provides a powerful framework for describing syntactic transformations by tracking changes in grammatical relations—such as a "deep" indirect object becoming a "surface" subject. However, the multi-layered or "stratal" nature of RG has historically posed a challenge for computational implementation. Most modern parsers rely on monotonic unification, which struggles with the perceived "deletion" or "overwriting" of information across strata. This paper presents an implemented parser that reconciles RG's linguistic depth with the efficiency of unification-based frameworks using Stratified Feature Grammar (SFG).
Stratified Feature Grammar (SFG)
The core innovation is the use of Stratified Feature Graphs (S-graphs). Unlike standard feature structures that use atomic labels, S-graphs employ sequences of "Relational Signs" (R-signs). For example, a label like [3, 2, 1, 0] represents an element's progression through various grammatical roles (e.g., Indirect Object $\to$ Direct Object $\to$ Subject $\to$ Chômeur).
This sequence-based approach solves the "carry-over" problem. In previous systems, explicit rules were needed to maintain relations across levels. In SFG, the persistence of a relation is inherent in the sequence, allowing for a monotonic representation of change. The use of a "0" sign allows the system to represent elements that exist in deep structure but disappear on the surface, maintaining monotonicity even when things seem to "disappear."
The Parsing Algorithm
The system utilizes a Head-Driven Chart Parser. Parsing begins with the lexical "head" of a phrase (typically the verb) and expands bidirectionally to satisfy the requirements encoded in the S-graphs. This process is facilitated by a specialized Label Unification operation. While standard unification merges identical atoms, SFG unification merges compatible sequences to find the "maximal nonempty overlap," effectively stitching together different relational perspectives of a word.
As shown in the chart parsing of "John seemed ill," the parser builds the structure by anchoring on the head "seemed" and integrating the subject "John" and the predicate "ill" based on their pre-compiled lexical entries. By pre-compiling complex rules like Passive or Dative shifts into lexical entries, the parser avoids the exponential search spaces that often plague multi-level linguistic theories.
Key Insights and Results
The authors demonstrate that even with the complexity of multi-stratal relations, the "Surface Graph" (the final layer of the S-graph) always forms a tree structure. This allows the use of efficient chart-parsing techniques originally designed for context-free grammars.
The primary result is a feasibility proof: a unification-based parser can indeed handle the full multi-stratal complexity of Relational Grammar. The S-graph serves as a single data structure that simultaneously represents the predicate-argument structure (the "meaning" level) and the surface syntactic structure.
Conclusion
By leveraging Stratified Feature Grammar and a head-driven approach, the authors provide a computationally elegant solution to a long-standing problem in theoretical linguistics. While the current implementation requires output filters to ensure well-formedness and handles extractions through more complex mechanisms, it proves that the rich descriptive power of Relational Grammar is compatible with modern, efficient parsing algorithms.
Summary updated