An Experimental Parser for Systemic Grammars

Robert T. KasperCOLING 1988

TL;DR

This paper introduces a novel method for parsing systemic grammars, a linguistic framework that explicitly models grammatical choices and their structural consequences. By compiling systemic grammars into Functional Unification Grammar (FUG) and extending existing parsing techniques, the authors demonstrate a practical approach to analyzing language that can leverage grammars originally developed for text generation.

What Problem Does This Paper Address?

Systemic grammar (SG) offers a rich, explicit representation of how language functions in communication, making it excellent for text generation. However, using SG for parsing (analyzing text) has been challenging. Traditional parsers focus on constituent structure, but SG also requires assigning grammatical features and functions to constituents, reflecting their meaning and role. Furthermore, SG's "system networks" distribute structural information across many choices, making it difficult to extract a coherent parsing strategy. This paper addresses the challenge of building a general-purpose parser for SG that can effectively utilize existing, complex grammars like Nigel, which was developed for generation.

Methods

The core of the approach is to compile systemic grammars into the notation of Functional Unification Grammar (FUG). This involves representing the system network and its "realization statements" (which define structural consequences) as FUG formulas. The authors then extend existing FUG parsing methods, specifically adapting the PATR-II chart parsing algorithm. Key extensions include:

  • Handling Disjunctive and Conditional Descriptions: SG systems involve choices (disjunctions) and dependencies between features (conditionals). The authors developed a unification algorithm that doesn't require expanding these into a massive disjunctive normal form (DNF), which would be computationally infeasible for large grammars.
  • Skeletal Phrase Structure Component: To manage the complexity of optional grammatical elements (like adjuncts), they introduce a minimal phrase structure (PS) component. This component, unlike in other grammars, acts as a "skeleton" that recognizes broad structural patterns, while the FUG unification handles the fine-grained feature assignments and constraints. This avoids generating an explosion of PS rules for every possible combination of optional elements.
  • Compiled Tables: Realization statements are compiled into tables that store information about possible feature conflations and lexical item associations, making the parsing process more efficient.

Datasets

Not discussed in the paper. The paper focuses on the methodology and testing with the Nigel grammar of English.

Key Insights and Ideas

  • Bidirectional Grammar is Achievable but Requires Tuning: The paper strongly supports the idea that a single grammar can be used for both generation and parsing (bidirectional grammar). However, it highlights that grammars optimized for generation may need significant "tuning" (modifications) to be efficient for parsing, and vice-versa.
  • Distributed Information is a Parsing Challenge: The distributed nature of structural constraints across SG's system networks is a fundamental challenge for parsing. The FUG compilation and the skeletal PS approach are key to overcoming this.
  • Efficiency Gains from Avoiding DNF: The authors' extension to FUG parsing, which avoids the computationally prohibitive DNF expansion of large grammars, is a critical technical contribution.
  • Skeletal PS for Optionality: Using a minimal PS component to handle optional elements, rather than generating exhaustive combinations, is an elegant solution to combinatorial explosion.

Key Results

The authors successfully developed and tested a parser for the Nigel grammar of English. While specific quantitative results like parsing speed are mentioned qualitatively (e.g., "less than a minute to parse simple single-clause sentences"), the primary result is the demonstration of a working, general method for parsing systemic grammars. The paper shows that a grammar adequate for text generation can be adapted for text analysis, albeit with effort.

Limitations and Future Work

  • Real-time Performance: The current parser is described as reasonably efficient for grammar testing and linguistic research but would require further refinement for real-time applications.
  • Bottom-Up Strategy: The parser currently uses an exhaustive bottom-up strategy, which is good for grammar testing but less efficient for applications needing selectivity. Future work could explore other parsing strategies.
  • Incorporating Extra-Grammatical Information: The parser does not yet incorporate semantic or pragmatic information for ambiguity resolution, which is noted as a necessary step for practical applications.
  • Automating PS Component Generation: A significant future goal is to automatically generate the skeletal phrase structure component from the systemic description, which would fully automate the creation of a parsing grammar. This is acknowledged as a difficult problem.

Summary updated

BibTeX
@inproceedings{kasper1988experimental,
  author = {Robert T. Kasper},
  title = {{An Experimental Parser for Systemic Grammars}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
}
Figure