Abstract
This paper introduces Nonmonotonic Sorts (NSs), a formal framework for integrating default information into unification-based grammars. By redefining sorts as pairs of strict and default information, the proposed method maintains the declarative, order-independent nature of unification (commutativity and associativity), addressing a critical limitation in previous procedural approaches to defaults in computational linguistics.
Introduction
In computational linguistics, unification-based grammars model language by combining feature structures—complex bundles of information such as person, number, or tense. A primary goal for these systems is to remain declarative, meaning the final result of combining information should be independent of the order of operations. However, natural languages are replete with defaults and exceptions. For instance, while most German verbs follow a "weak" pattern, others are "strong" or "mixed."
Historically, handling these defaults has relied on procedural methods like "priority union," which are order-dependent. If a default is applied before a strict rule, the outcome may differ from the reverse order. This fragility makes large-scale grammar development difficult. This paper introduces Nonmonotonic Sorts (NSs) to move "default-ness" from the unification operation itself into the underlying sort system, preserving the mathematical elegance of standard unification.
Methods: Nonmonotonic Sorts
The core innovation is the representation of a node's type as a pair $(f, \Delta)$, where $f$ represents strict information and $\Delta$ is a set of default sorts.
When two NSs are unified, their strict components undergo standard unification. The default components are aggregated into a set, and any default that contradicts the new strict information is filtered out. This approach ensures that unification remains commutative and associative. To derive a final feature structure, the system identifies "solutions"—standard structures that incorporate the maximum number of non-conflicting defaults. This allows the system to represent multiple valid interpretations if defaults conflict.
Case Study: German Verb Morphology
The framework is demonstrated using a toy lexicon of German verb morphology, focusing on how different verb classes inherit suffixes and prefixes.
As shown in Figure 2, a general VERB template defines default suffixes like +te for the past tense. A MIXED-VERB inherits from VERB but overrides the past participle suffix with +en. A STRONG-VERB further overrides the past tense suffix to be null ($\emptyset$). The NS approach handles these overrides gracefully: if a specific lexical entry (like zwing) provides strict stem information, it naturally takes precedence over inherited defaults without requiring a specific sequence of rule applications.
Key Insights and Results
The NS approach offers several advantages:
- Declarative Defaults: By changing the data representation rather than the operator, standard unification remains "clean."
- Conflict Preservation: Defaults are not "overwritten" and lost; they are merely filtered during solution generation. This allows the system to handle complex multiple inheritance scenarios.
- Order Independence: The author provides a formal proof (Theorem 1) that NS unification is commutative and associative, guaranteeing stability regardless of rule application order.
- Handling Linguistic Complexity: The paper demonstrates a specific linguistic example (the verb forbid) where existing systems fail or require manual, arbitrary ordering. The NS approach handles it automatically by correctly prioritizing strict information over inherited defaults.
Limitations and Future Work
Currently, the method primarily addresses atomic values. Extending this to path equations—where two different parts of a structure are "by default" identical—is more complex due to potential non-local effects. Additionally, while the current model treats all defaults as equal, future work could incorporate a priority system where defaults from more specific levels of a hierarchy outrank more general ones.
Summary updated