Features and Values

Lauri KarttunenCOLING 1984

TL;DR

This paper introduces a generalized and expanded feature computation facility, building on prior work in unification-based grammars. It addresses limitations in existing formalisms by incorporating negative and disjunctive feature specifications, enabling more expressive and linguistically accurate representations for phenomena like agreement.

What Problem Does This Paper Address?

Linguistics, across phonology, morphology, syntax, and semantics, heavily relies on feature structures to represent linguistic information. Traditionally, these features are attribute-value pairs, where values can be atomic (like "sg" for singular) or complex (another set of attribute-value pairs). While powerful, existing formalisms for handling these feature structures, particularly those based on unification (like LFG, UG, GPSG, and PATR-II), faced limitations. Specifically, they struggled with:

  • Cyclic Structures: Representing features that could refer back to themselves, which can arise during grammar construction.
  • Negation: Expressing that a feature does not have a certain value (e.g., "case is not dative").
  • Disjunction: Expressing that a feature can have one of several values (e.g., "case is nominative or accusative").

These limitations made it difficult to capture certain natural language phenomena accurately and efficiently, forcing linguists to resort to less elegant or more cumbersome representations.

Methods

The paper presents a new generalized feature computation facility, an evolution of Stuart Shieber's PATR-II system. The core innovation lies in extending the unification mechanism to handle:

  • Negative Constraints: This allows specifying that a feature's value must not be a particular value. For example, [case: -dat] means "case is not dative." This is implemented by marking nodes with constraints that prevent them from becoming identical to a specified value.
  • Disjunctive Specifications: This allows a feature to have one of several possible values, represented using curly braces, e.g., [case: {nom acc}]. The system handles unification with disjunctions by creating complex disjunctions of tuples, where each tuple represents a potential unification path. If one path becomes inconsistent, the system can still proceed with the remaining valid paths.

The paper also discusses how these new capabilities interact with existing operations like generalization (similar to set intersection).

Datasets

Not discussed in the paper. The paper focuses on the formalism and its theoretical implications rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Expressive Power of Negation and Disjunction: The paper convincingly argues that negative and disjunctive feature specifications are not just theoretical niceties but are crucial for accurately representing real-world linguistic phenomena, particularly agreement and case marking in languages like German.
  • Unification as a Core Operation: The paper reinforces the power of unification as the fundamental operation for combining feature structures, extending its applicability to more complex linguistic scenarios.
  • Connection to Logic Programming: The authors note a strong similarity between the problems and solutions in feature-based systems and those in logic programming, suggesting potential for cross-pollination of ideas and techniques.
  • Revisiting Markedness: The proposed system offers a way to model linguistic markedness hierarchies (e.g., 3rd person being more marked than 1st) using negative constraints, preserving traditional linguistic intuitions while offering a computationally elegant solution.

Key Results

The paper doesn't present quantitative results in terms of accuracy or performance metrics. Instead, its key results are:

  • A functional implementation: The paper describes a "new Texas version of the 'DG (directed graph)' package" that successfully implements these extended feature specifications.
  • Demonstration of linguistic applicability: Through examples from German case marking and English verb agreement, the paper shows how the new formalism can elegantly capture phenomena that were problematic for previous systems.
  • Theoretical justification: The paper provides a strong theoretical argument for the necessity of negative and disjunctive specifications, demonstrating their utility in simplifying grammar rules and avoiding redundant lexical entries.

Limitations and Future Work

  • Cyclic Structures: While the PATR-II system at SRI and its expanded version at Texas allow cyclic structures for practical reasons, the paper notes that it's not theoretically justified whether such structures are necessary.
  • Negative Constraint Handling in Generalization: The paper acknowledges that the current implementation's policy of dropping redundant negative constraints might need modification to align with proposed generalization mechanisms.
  • Computational Cost: The paper mentions that unification with negative or positive constraints naturally takes more time, though the Texas implementation aims to minimize this impact for non-constrained values.
  • Future Work: The authors suggest that the system could be used for certain kinds of inferencing, drawing parallels to logic programming. They also express interest in exploring other cases of markedness that could be analyzed using their approach.

Summary updated

BibTeX
@inproceedings{karttunen1984features,
  author = {Lauri Karttunen},
  title = {{Features and Values}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1984},
}
Figure