TL;DR
This paper introduces the LinGO Grammar Matrix Customization System, a web-based tool that helps linguists rapidly prototype grammars for new languages. By answering a typological questionnaire, users can generate a foundational grammar fragment compatible with existing linguistic software, significantly reducing the effort required to build comprehensive grammars.
What Problem Does This Paper Address?
Developing computational grammars for natural languages is a time-consuming and labor-intensive process. Creating a broad-coverage, precision grammar from scratch for a new language requires extensive linguistic expertise and significant engineering effort. This is particularly challenging for low-density languages or when aiming for cross-linguistic coverage. The authors address this by aiming to reduce the cost and effort of grammar creation, enabling linguists to quickly build and test grammatical analyses for a wide range of languages.
Methods
The core of the system is a web-based, typological questionnaire. This questionnaire elicits detailed linguistic information from the user, covering various grammatical phenomena. The system interprets these answers without human intervention and generates a grammar fragment in TDL (Type Description Language), compatible with the LKB (Lexical-Knowledge Builder) grammar development environment.
Key methodological innovations include:
- Cross-linguistic Core Grammar and Libraries: The system is built upon a reusable core grammar and a collection of "libraries" that encapsulate analyses of typologically variable phenomena (e.g., word order, tense/aspect, case). This modular approach allows for efficient reuse of linguistic knowledge.
- Iterator Mechanism: To handle an unbounded number of lexical entries or morphological structures, the questionnaire employs an iterator mechanism. This allows users to create arbitrarily many lexical rule "slots" and morphemes, enabling richer and more complex linguistic descriptions.
- Dynamic Content in Questionnaire: The questionnaire is dynamic, adapting its presentation and questions based on user input and system-provided information. This includes expandable lists for unbounded entries and automatically populated drop-down selectors, reducing user burden and preventing invalid entries.
- Validation Component: Before grammar generation, a validation component checks the user's answers for consistency and potential issues, flagging problematic entries with visual cues and explanations.
- Test-by-Generation Feature: A novel "test-by-generation" feature allows users to test the generated grammar fragment directly within the browser. This process generates sentences from underspecified Minimal Recursion Semantics (MRS) representations, providing a quick way to identify overgeneration (ungrammatical sentences being produced) and undergeneration (expected sentences not being produced).
Datasets
The paper does not explicitly mention specific datasets used for training or evaluation in the traditional machine learning sense. Instead, the "dataset" is the collection of linguistic knowledge encoded within the LinGO Grammar Matrix itself, comprising its core grammar and libraries. The system is designed to create grammars for new languages based on user input, rather than training on pre-existing corpora of those languages. The paper does, however, present a table comparing the size of grammars generated by the system for 11 languages to the English Resource Grammar (ERG), providing a quantitative comparison of the complexity of these generated grammars.
Key Insights and Ideas
- Democratizing Grammar Engineering: The system significantly lowers the barrier to entry for computational grammar development, making it accessible to a broader range of linguists, including graduate students.
- The Power of Typological Questionnaires: Eliciting linguistic information through a structured, typological questionnaire is an effective way to capture cross-linguistic variation and automate grammar fragment generation.
- Iterators for Richness: The iterator mechanism is crucial for enabling the creation of detailed and complex linguistic analyses, moving beyond simple, finite descriptions.
- Generation as a Stringent Test: The test-by-generation feature offers a more insightful testing method than parsing alone, particularly for identifying grammatical errors and understanding the system's behavior with ungrammatical inputs.
- Bridging Typology and Formal Syntax: The Grammar Matrix project aims to integrate insights from typological research with formal syntactic analysis, creating a resource that is both broad in coverage and deep in analytical rigor.
Key Results
The paper demonstrates the system's ability to generate "starter grammars" for over 40 languages. Table 1 provides a quantitative comparison of the size (number of types, lexical rules, phrasal rules) of grammars generated for 11 languages over three weeks by graduate students, compared to the much larger and more established English Resource Grammar (ERG). The average generated grammar had significantly fewer types and rules than the ERG, but the system successfully produced functional grammars that could be further developed. The test-by-generation feature is highlighted as a practical tool for identifying issues in these generated grammars.
Limitations and Future Work
- Limitations:
- The generated grammars are "starter grammars" and require further hand-editing and development to become broad-coverage, precision grammars.
- While the system supports many phenomena, it's not always immediately apparent to users how different linguistic analyses will interact.
- The current system does not include visualization tools for designing type hierarchies or morphotactic dependencies.
- Future Work:
- The authors anticipate future improvements, including visualization tools to aid in grammar design.
- They believe the system is general enough to support the addition of analyses for many more linguistic phenomena.
- The test-by-generation methodology is expected to encourage users to develop more complex grammars within the system before moving to hand-editing.
Summary updated