TL;DR
This paper introduces a computational approach to testing generative grammars, specifically a functional generative description of Czech. By developing a computer program to simulate grammar generation, the authors aim to pragmatically assess whether grammars function as intended and to gain deeper insights into rule interactions.
What Problem Does This Paper Address?
The authors tackle a fundamental challenge in computational linguistics: how to empirically verify the correctness and behavior of complex generative grammars. Traditionally, evaluating such grammars relied heavily on linguistic intuition, which can be subjective and difficult to scale. The paper addresses the need for a systematic, computer-aided method to test whether a grammar "works" as designed, allowing linguists to move beyond theoretical descriptions to practical, data-driven analysis. This is particularly relevant for functional generative descriptions, which aim to capture semantic representations of sentences.
Methods
The core of the paper describes the development of a computer program designed to generate sentences based on a given generative grammar. This approach contrasts with recognition procedures (which check if a given string can be generated). The program simulates the application of rewriting rules, specifically focusing on the generative component (GC) of a functional generative description. Key aspects of their method include:
- Type 2 Grammar: The GC is classified as a context-free phrase structure grammar (Chomsky's Type 2).
- Dependency Syntax: The grammar incorporates principles of dependency syntax alongside immediate constituents.
- Rule Types: The grammar utilizes modifying, substitutional, and selectional rules.
- Probabilistic Generation: To avoid infinite loops and ensure representative samples, rules are not chosen randomly but with prescribed probabilities. This allows for controlling the frequency of certain phenomena in the generated output.
- Leftmost Derivation: The generation process uses a leftmost derivation strategy.
- Rule Schemes: To enhance legibility and efficiency, rule schemes are employed, which represent generalizations of multiple rules.
- Data Structures: The grammar is represented in a structured format suitable for computer processing, including name-symbols, indices, and their values.
Datasets
Not discussed in the paper. The paper focuses on the methodology of testing a grammar rather than its application to a specific, large-scale dataset. The "lexicon" size is mentioned as 275 semoglyphs, which is noted as a relatively small extent for computer experiments.
Key Insights and Ideas
- Pragmatic Grammar Evaluation: The paper champions a shift from purely theoretical evaluation to pragmatic, computer-driven testing of generative grammars.
- Understanding Rule Interaction: Computer simulation allows linguists to observe how rules interact and influence the generation process, providing insights that are hard to glean from static rule sets.
- Controlled Generation: The probabilistic approach to rule selection is crucial for generating meaningful samples and avoiding degenerate outputs, enabling targeted analysis of specific linguistic phenomena.
- Efficiency through Representation: The use of rule schemes and references for index values and symbols demonstrates a focus on computational efficiency and memory management.
Key Results
The paper doesn't present quantitative results from a completed, large-scale experiment. Instead, it details the design and preparation for such an experiment. The key "results" are the architectural design of the testing program and the characterization of the grammar being tested:
- Grammar Characteristics: The tested variant of the functional generative description of Czech has:
- 62 non-terminal name-symbols.
- A maximum of 15 indices per non-terminal name-symbol (average 5).
- A maximum of 30 indices per terminal name-symbol (average 3).
- 304 terminal name-symbols in total.
- 275 "proper" terminal name-symbols corresponding to lexemes.
- A binary context-free grammar structure (max two non-terminal symbols on the right-hand side of rules).
- Program Status: The paper describes the ongoing development of the computer system, including the GRAMMAR TRANSDUCER for converting the grammar into a computer-readable format and the main DERIVATION OF SR's program. The programming language chosen is PL/1 for IBM 360/370 computers.
Limitations and Future Work
- Partial Goal Achievement: The authors acknowledge that choosing the most adequate way to transfer linguistic knowledge into a computer-usable format is challenging and they have only "reached this goal only partially."
- Experimental Nature: The described system is still under development. The paper outlines the planned system and the initial steps taken, rather than presenting final experimental outcomes.
- Future Work: The immediate future work involves completing the programming of the system and conducting experiments. The authors suggest that the system will allow them to answer questions like "What will be the effect of a modification of the probability in the rule number n, upon the generation of SR's?" by running further experiments. They also envision testing the "whole generative grammar" and preparing data for other components of the description.
Summary updated