TL;DR
This paper explores the use of computational methods to simulate historical language change, specifically focusing on phonological shifts in Russian. The authors propose a "top-to-bottom" generative model that uses reconstructed proto-language forms and ordered phonological rules to predict modern language forms, aiming to improve the accuracy and understanding of relative chronologies in linguistic reconstruction.
What Problem Does This Paper Address?
Understanding how languages evolve over time (diachronic linguistics) is crucial for explaining the structure of modern languages. Historically, two main challenges have plagued this field: accurately describing the sound changes themselves, and determining the order in which these changes occurred (relative chronology). While significant effort has gone into describing individual sound changes, establishing their sequence has been difficult due to the vast amount of data required and the combinatorial complexity of rule ordering. This paper aims to address these challenges by proposing a computational approach to simulate historical language change, specifically focusing on phonological changes in Russian, to better understand and potentially automate the process of linguistic reconstruction.
Methods
The paper proposes a "top-to-bottom" generative model, reversing the traditional "bottom-to-top" comparative method. Instead of starting with modern languages and reconstructing older forms, this model begins with reconstructed proto-language forms and a set of ordered phonological rules. The program then applies these rules sequentially to simulate the synchronic stages of language development, with the ultimate goal of generating the modern language form. The authors developed a program in SNOBOL4 to implement this model. The core idea is to use a computer to test the accuracy of postulated rules and, crucially, to help discover or verify the relative chronology of these rules by examining the output at each stage.
Datasets
The study uses approximately five hundred reconstructed Proto-Indo-European (PIE) forms from Walde and Pokorny (1932). These forms were chosen to be as long as possible to provide a variety of phonetic environments. A separate Russian gloss was also used. The paper does not introduce a new benchmark dataset but rather uses existing reconstructed forms as input for their simulation model.
Key Insights and Ideas
- Computational Approach to Diachrony: The paper pioneers the use of computational simulation for historical language change, moving beyond manual analysis to rigorous testing of linguistic hypotheses.
- Top-to-Bottom Generative Model: Reversing traditional reconstruction methods, this model starts with proto-forms and rules to generate later stages, offering a new perspective on linguistic evolution.
- Importance of Rule Ordering: The authors highlight that the relative chronology of sound changes is as critical as the changes themselves, and their computational model is designed to test and potentially discover this ordering.
- Identifying Sources of Error: The model provides a framework for diagnosing errors in linguistic reconstruction, categorizing them into issues with rule formulation, rule existence, rule ordering, or reconstructed forms.
- Testing Hypotheses: The computer can be a powerful tool for testing hypothesized sound changes and their sequences, especially under simplifying assumptions.
Key Results
The programmed model, using a set of 21 abridged phonological rules for Russian, successfully predicted the Modern Russian form from PIE input in several cases. Examples include:
- PIE
*bhelgh-> Mod. Russianboloz- - PIE
*medhi--> Mod. Russianmež- - PIE
*korm--> Mod. Russiansram
However, a significant number of PIE forms were not successfully related to Modern Russian. The authors attribute this primarily to the loss of original PIE forms, but also acknowledge issues with the simplification of consonant clusters and the handling of certain sound alternations that lacked explicit rules or had exceptions. No errors in rule ordering were found in this specific sample.
Limitations and Future Work
The authors acknowledge several limitations:
- Language Dependency: The current program is specific to Russian, though the underlying conception is broadly applicable.
- Incomplete/Incorrect Rules: Documented historical changes are often not stated precisely, leading to potential inaccuracies in the rule set.
- Assumption of Uniform Proto-Language: The model assumes a uniform, single-stage proto-language, which may require further restrictions.
- Difficulty in Error Diagnosis: Pinpointing the exact cause of an incorrect output (rule formulation vs. ordering vs. input form) can be challenging, especially when multiple rules are involved.
Future work suggested includes:
- Incorporating Borrowing: Simulating the effect of lexical borrowing by introducing items at specific stages.
- Analyzing Loss of PIE Forms: Further study is needed to understand the effects of losing original PIE forms.
- Improving Error Analysis: Developing more sophisticated methods, like KWIC indexing of phone occurrences, to simplify error discovery.
- Generalizing Rules: Investigating whether rules need to be generalized or if new, unpostulated changes are responsible for discrepancies.
- Investigator Consistency: The computer operates best when rules and proto-forms are established by the same investigator, adhering to a consistent theory of historical change.
Summary updated