TL;DR
This paper argues that lexical rules in Head-Driven Phrase Structure Grammar (HPSG) should apply based on subsumption rather than unification with lexical entries. This change prevents overgeneration of ungrammatical sentences and spurious structural ambiguities that arise when using unification, without sacrificing the ability to capture grammatical generalizations.
What Problem Does This Paper Address?
Linguistic theories often use "lexical rules" to capture generalizations between related words (e.g., active vs. passive forms of verbs). In constraint-based grammar formalisms like HPSG, these rules operate on feature structures. A key question is: when should a lexical rule be applied to a specific word's entry? The paper identifies two main hypotheses:
- Hypothesis A (Unification): A rule applies if the word's feature structure unifies with the rule's description. This is the commonly assumed approach.
- Hypothesis B (Subsumption): A rule applies if the word's feature structure is subsumed by the rule's description. Subsumption means the word's entry is more specific than the rule's description.
The authors demonstrate that Hypothesis A, while intuitive, leads to significant problems in analyzing German grammar. Specifically, it causes the grammar to either incorrectly accept ungrammatical sentences or produce multiple, unnecessary structural analyses for grammatical ones.
Methods
The paper's core method is to analyze specific lexical rules used in HPSG analyses of German and show the empirical consequences of applying them under unification versus subsumption.
- Case Studies: The authors examine two types of lexical rules:
- Passive Lexical Rules (PPLR & IPLR): These rules handle the active-passive voice alternation in German.
- Complement Extraction Lexical Rule (CELRG): This rule handles phenomena like topicalization, moving elements from a verb's complements to a "SLASH" set.
- Hypothesis Testing: They analyze how these rules interact with underspecified lexical entries (e.g., for auxiliary verbs) under both unification and subsumption.
- Empirical Argumentation: The argument is primarily empirical, showing how specific linguistic examples lead to ungrammaticality or ambiguity under unification, but are correctly handled by subsumption.
Datasets
Not discussed in the paper. The paper focuses on theoretical linguistic analysis and uses example sentences from German grammar rather than formal datasets.
Key Insights and Ideas
- The Subsumption Principle: The central idea is that lexical rules should be seen as templates that generalize over specific lexical entries. A rule should only apply if the lexical entry is a specific instance of the general pattern described by the rule, which is precisely what subsumption captures.
- Underspecification is Key: The problems with unification arise when lexical rules interact with highly underspecified lexical entries, particularly those for auxiliary verbs. These entries are designed to be flexible, but unification makes them too "open" for passive or extraction rules.
- Overgeneration and Spurious Ambiguity:
- Overgeneration: Unification can lead to rules applying where they shouldn't, generating ungrammatical sentences (e.g., passivizing an auxiliary verb in German).
- Spurious Ambiguity: Unification can create multiple, linguistically meaningless structural analyses for the same sentence, complicating parsing.
- Subsumption as a Solution: Hypothesis B (subsumption) elegantly solves these problems by ensuring rules only apply when the lexical entry is sufficiently specified to match the rule's requirements, thus preventing incorrect applications and unnecessary analyses.
Key Results
- Applying the Passive Lexical Rule (PPLR) under unification to an auxiliary verb like "können" (can) leads to the generation of ungrammatical sentences like "Das Auto wurde kaufen gekonnt" (The car was bought can). This is avoided under subsumption.
- Applying the Complement Extraction Lexical Rule (CELRG) under unification to auxiliary verbs creates spurious structural ambiguities for sentences involving topicalization and auxiliaries. Subsumption correctly restricts the rule's application to main verbs, resolving this ambiguity.
- The paper demonstrates that the subsumption test for LR applicability can be integrated into existing computational HPSG processing frameworks (van Noord & Bouma 1994, Meurers & Minnen 1995).
Limitations and Future Work
- Computational Implementation: While the paper argues for subsumption and shows it can be integrated into existing frameworks, it does not provide a fully worked-out computational proposal for processing LRs under subsumption. This is noted as future work.
- Formal Semantics: The paper acknowledges that a consensus on the formal semantics of lexical rules has not yet been reached, and a full treatment of this is beyond its scope.
- Broader Applicability: The paper focuses on German examples. While the principles are general, further testing on other languages and phenomena would be valuable.
- Other Linguistic Phenomena: The authors briefly mention that subsumption has also been found useful for analyzing coordination, suggesting its broader applicability in linguistic theory.
Summary updated