An HPSG Parser Based on Description Logics

J. Joachim QuantzCOLING 1994

TL;DR

This paper introduces a novel parser for German HPSG-style grammar that leverages Description Logics (DL) for syntactic parsing and disambiguation. By representing linguistic information declaratively within DL, the parser can utilize powerful inference mechanisms to handle ambiguity, particularly through a proposed exception minimization approach for interpretation.

What Problem Does This Paper Address?

Parsing natural language expressions, especially those with syntactic ambiguity, is a significant challenge in computational linguistics. Traditional approaches often struggle to integrate diverse linguistic information (syntactic, semantic, encyclopedic) into a unified framework for effective disambiguation. This paper addresses the need for a homogeneous and declarative approach to parsing and interpretation, specifically focusing on how to handle ambiguity in German HPSG-style grammar by exploiting the inferential capabilities of Description Logics. The goal is to build a parser that can manage complex preferences and resolve ambiguities by minimizing exceptions to default rules.

Methods

The core of this work is the development of an HPSG parser that is entirely based on Description Logics (DL). Instead of traditional unification-based parsing, this approach represents linguistic knowledge, including syntactic categories and phrase structure rules, as DL concepts and roles.

  • DL Representation: Syntactic categories (like noun, verb, NP) are modeled as DL concepts, and phrase structure relations (like daughters) are represented as DL roles. This allows for a uniform representation of linguistic information.
  • Inferential Parsing: The parser leverages the inference capabilities of a DL system to perform unification and check grammatical constraints. This means that instead of explicit unification algorithms, the DL reasoner handles consistency checks and derivations.
  • Situations for Ambiguity: To handle ambiguity, the paper introduces the concept of "situated descriptions" (o :: c in s), where s represents a specific interpretation or perspective. This allows for representing different possible parses of the same sentence within the DL framework.
  • Exception Minimization for Disambiguation: The paper proposes extending the parser with Preferential Default Description Logics (PDDL). Ambiguity is resolved by assigning weights to preference rules (defaults) and selecting the interpretation that violates the fewest or least significant defaults, effectively minimizing exceptions.

Datasets

Not discussed in the paper. The paper focuses on a theoretical framework and a small fragment of German grammar, using an example sentence for illustration.

Key Insights and Ideas

  • DL as a Unified Formalism: The paper demonstrates that DL can serve as a powerful and uniform formalism for representing diverse linguistic information, including syntax, which is often handled by specialized feature formalisms.
  • Inferential Power for Parsing: By relying on DL inference, the parser offloads complex grammatical checks and unification to the DL reasoner, offering a declarative alternative to explicit parsing algorithms.
  • Situations for Representing Ambiguity: The use of "situations" in DL provides a clean way to represent and manage multiple interpretations of a sentence, aligning with chart-parsing concepts but within a logical framework.
  • Exception Minimization as a Disambiguation Strategy: The proposal to use weighted defaults in DL for exception minimization offers a principled and formally grounded approach to resolving linguistic ambiguity.

Key Results

The paper presents a conceptual framework and a prototype parser. The key result is the demonstration that an HPSG-style parser can be implemented using DL inference, and that this approach can be extended to handle syntactic ambiguity through exception minimization. The paper provides an example of parsing the German sentence "Die schöne Frau sieht sie" and illustrates how different parses can be represented and potentially disambiguated. No quantitative performance metrics are provided.

Limitations and Future Work

  • Efficiency: The authors explicitly state that efficiency aspects have been deliberately ignored in the presentation to keep it simple. This is a significant limitation for practical applications.
  • Completeness: The paper focuses on a small fragment of German HPSG and does not aim to model all aspects of relevant knowledge. Information like linear precedence and traces are implicitly handled or not fully encoded in the DL model.
  • Future Work: The authors suggest integrating chart parsing techniques to improve performance. They also highlight the potential for extending the framework to incorporate semantic and encyclopedic information, and further developing the exception minimization approach for more robust disambiguation. The paper serves as a basis for future implementations and research into efficient and comprehensive natural language processing systems.

Summary updated

BibTeX
@inproceedings{quantz1994hpsg,
  author = {J. Joachim Quantz},
  title = {{An HPSG Parser Based on Description Logics}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1994},
}
Figure