Argument Optionality in the LinGO Grammar Matrix

Safiyyah Saleem, Emily M. BenderCOLING 2010

TL;DR

This paper introduces a new library for the LinGO Grammar Matrix, a system for automatically generating grammars for natural languages. This library models "argument optionality" – the phenomenon where subjects or objects can be omitted – across a wide range of typological variations, achieving high coverage on test suites for 10 natural languages. This significantly enhances the Grammar Matrix's ability to handle diverse linguistic phenomena.

What Problem Does This Paper Address?

Many languages allow speakers to omit subjects or objects when they are understood from context, a phenomenon known as "argument optionality" or "pro-drop." This is incredibly common, with studies suggesting around 79% of languages allow for this. Before this work, the LinGO Grammar Matrix, a valuable tool for creating computational grammars, couldn't handle this crucial linguistic feature. This limitation meant it couldn't accurately model many common sentence structures in languages like Arabic or Hausa, where argument dropping is prevalent. The challenge was to develop a flexible and typologically comprehensive way to model the diverse constraints that govern when and how arguments can be dropped across different languages.

Methods

The authors tackle argument optionality by extending the Head-Driven Phrase Structure Grammar (HPSG) framework within the LinGO Grammar Matrix. Their core innovation is the use of unary (non-branching) phrase structure rules that can "cancel off" valence elements (arguments required by a verb) without requiring an overt realization. This is controlled by a boolean feature OPT on valence elements.

They introduce new phrase structure rules like head-opt-subj-phrase and head-opt-comp-phrase that allow verbs to satisfy valence requirements if the corresponding argument is marked as optional (OPT +). To handle the vast typological variation, they employ several strategies:

  • Lexical Typing: Verbs are classified into types based on whether they allow subject/object dropping, and specific lexical items instantiate these types.
  • Feature Manipulation: The OPT feature is manipulated to enforce constraints. For instance, verbs that don't allow dropping are constrained to have arguments that are not optional (OPT -), preventing them from undergoing the dropping rules.
  • Lexical Rules: For complex co-occurrence restrictions (e.g., between affixes and overt arguments), they use "no-marker" lexical rules to enforce constraints associated with markerless verbs.
  • Syntactic Context Constraints: For languages with context-dependent dropping (e.g., based on tense or person), they constrain the unary dropping rules to apply only in specific contexts (e.g., requiring a non-third person subject for dropping in Finnish).

This approach allows for a flexible system that can model various combinations of optionality, requirements, and prohibitions for both subjects and objects, as well as their interaction with affixes and overt arguments.

Datasets

The paper does not introduce new datasets but rather uses existing linguistic descriptions and typological surveys to create test suites. They developed three sets of test suites:

  • Set 1: Pseudo-Languages: 38 artificial languages designed to exhaustively cover all possible combinations of lexical, syntactic context, and affix co-occurrence restriction patterns described in their typological analysis (Table 1). This allowed them to test the analyses in isolation.
  • Set 2: Illustrative Languages: Test suites for four natural languages (Arabic, Finnish, Hausa, Tamil) chosen for their interesting argument optionality patterns, based on existing descriptive grammars.
  • Set 3: Held-out Languages: Test suites for six diverse languages (Abkhaz, Chemehuevi, Jingulu, Malayalam, Nkore-Kiga, West Greenlandic) not considered during development, to assess the system's generalization capabilities.

Key Insights and Ideas

  • Argument Optionality is Highly Typological: The paper strongly reinforces that argument optionality is not a monolithic phenomenon but rather a complex system with diverse constraints across languages, involving syntax, lexicon, and morphology.
  • Unary Rules for Flexibility: Using unary rules to cancel valence elements provides a powerful and flexible mechanism for modeling argument dropping, allowing for cleaner integration with existing HPSG structures.
  • The OPT Feature is Key: The OPT feature, previously underutilized, becomes central to managing the various constraints on argument optionality, enabling fine-grained control over when arguments can be dropped.
  • Interplay of Constraints: The work highlights the critical interplay between lexical constraints (e.g., specific verbs), syntactic constraints (e.g., context), and morphological constraints (e.g., affix marking) in determining argument optionality patterns.
  • Typological Breadth is Achievable: The success in modeling a wide range of typological variations demonstrates that it's possible to build robust, typologically informed linguistic resources within a customizable grammar framework.

Key Results

  • 100% Coverage on Pseudo-Languages: The system achieved 100% coverage on the 38 pseudo-languages, with 0% overgeneration and no spurious ambiguity, validating the analyses across the known typological space.
  • High Coverage on Illustrative Languages: Achieved 100% coverage on three out of four illustrative languages (Finnish, Hausa, Tamil), with 90% coverage for Arabic. No overgeneration or spurious ambiguity was observed in these sets.
  • Good Coverage on Held-out Languages: The customized grammars accounted for the majority of patterns in the held-out languages, achieving 100% coverage on four languages with zero or moderate overgeneration.
  • No Negative Impact on Other Libraries: Crucially, the addition of the argument optionality library did not reduce coverage on other parts of the test suites, indicating proper integration with existing grammar components.

Limitations and Future Work

  • Arabic Item Parsing: One specific Arabic item failed to parse due to the current implementation's difficulty in elegantly handling obligatory object marking with object drop on some transitive verbs and optional marking on others. This points to potential improvements needed in the morphotactic system.
  • Word Order Constraints: The system struggled with word order constraints in held-out languages, specifically in licensing or restricting argument optionality based on the relative positions of verbs and arguments. Future work on the Grammar Matrix's word order library is needed to address this.
  • Semantic Distinctions: The current work focuses on syntactic variation. The authors intend to expand the library to include semantic distinctions, such as modeling the discourse status of NPs and its relation to null instantiation.
  • Unaware of All Variants: The authors acknowledge that there may be argument optionality variants they are unaware of and hope to learn of more patterns as the Grammar Matrix system is applied to even more languages.

Summary updated

BibTeX
@inproceedings{saleem2010argument,
  author = {Safiyyah Saleem and Emily M. Bender},
  title = {{Argument Optionality in the LinGO Grammar Matrix}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2010},
}
Figure