Entity-Oriented Parsing

Philip J. HayesCOLING 1984

TL;DR

This paper introduces "entity-oriented parsing," a novel approach for building natural language interfaces (NLIs) that focuses on defining domain-specific entities rather than traditional grammatical structures. This method promises greater robustness in handling ungrammatical input and simplifies language definition by integrating structural and surface representations of entities.

What Problem Does This Paper Address?

Building effective natural language interfaces for restricted domains (like querying a database) is challenging. Traditional methods, such as "semantic grammar," excel at exploiting domain knowledge to reduce ambiguity in grammatical input. However, they often struggle with ungrammatical or incomplete user input, leading to brittle systems. Furthermore, defining the language for these interfaces can be cumbersome, as it often requires separate specifications for the domain's objects/actions and their linguistic manifestations. This paper aims to bridge this gap by proposing a parsing paradigm that is both robust to input variations and easier to define.

Methods

The core of the proposed approach is entity-oriented parsing. Instead of relying on syntactic categories (like nouns or verbs), this method defines the "entities" (objects, events, commands, states) that the NLI needs to recognize. Each entity definition bundles together:

  • Internal Structure: How the entity is conceptually represented.
  • Surface Representation: How the entity is expressed in natural language (e.g., as a noun phrase, a specific sentence structure).
  • Correspondence: How the surface form maps to the internal structure.

This is achieved through a scheme reminiscent of conceptual or case-frame parsers. A key innovation is a parser control structure that uses an "agenda mechanism" with flexibility levels. This mechanism prioritizes parsing continuations that imply less grammatical deviation. When input doesn't match expectations, the parser can explore continuations with increasing flexibility levels, allowing it to handle spelling errors, missing words, or unusual phrasing. The paper also highlights the use of multiple construction-specific recognition strategies that can be invoked based on entity definitions, further enhancing robustness.

Datasets

Not discussed in the paper. The paper focuses on the methodology and provides example definitions and parses rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Entity-centricity is Key: Shifting focus from grammar rules to domain entities simplifies the problem by directly aligning linguistic recognition with the system's underlying knowledge and capabilities.
  • Integrated Definitions: Grouping structural and surface information for each entity streamlines language definition and ensures consistency. For example, defining how a "Course Department" is represented linguistically ensures that this representation is used consistently across all entities that involve a department.
  • Flexibility for Robustness: The agenda-based control structure with explicit "flexibility levels" provides a principled way to handle ungrammatical input. It allows the parser to explore increasingly "flexible" interpretations without wasting effort on overly complex deviations when simpler ones suffice.
  • Bottom-Up and Top-Down Synergy: The approach supports both bottom-up indexing from input words to potential entities and top-down recognition of specific entity structures, allowing for efficient parsing of grammatical input and robust handling of deviations.
  • "Split" for Alternatives: A crucial primitive in the control structure is the "split" function, which allows a strategy to branch into multiple alternative continuations, each with a potentially different flexibility increment, enabling the exploration of various parsing paths.

Key Results

The paper presents a conceptual framework and demonstrates its application through worked examples. It shows how an entity-oriented parser could successfully parse both a perfectly grammatical command ("Enrol Susan Smith in CS 101") and a significantly ungrammatical one ("Place Susan Smith in computer science for freshmen"). The parses are expressed directly in terms of the underlying structural representation of the entities, eliminating the need for a separate semantic interpretation step. The authors claim that this approach offers greater robustness and ease of language definition compared to traditional semantic grammar techniques.

Limitations and Future Work

The authors acknowledge that the proposed approach is still under implementation and a more rigorous test is needed. They also note that the performance of the parser, particularly in handling complex deviations, might depend on careful tuning of the flexibility level increments, which could potentially lead to instability. Future work will focus on testing and refining the parser's performance and stability as it becomes operational. They also suggest that the flexibility level increments might need to be fine-grained, and if this proves impractical, the system might need to interact with the user to resolve ambiguities.

Summary updated

BibTeX
@inproceedings{hayes1984entity,
  author = {Philip J. Hayes},
  title = {{Entity-Oriented Parsing}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1984},
}
Figure