TL;DR
This paper explores how computational linguistics can be applied to study language contact and change in low-resource historical languages. By building a model for Early-Modern Frisian verb clusters based on insights from modern Dutch, the authors provide evidence for language contact influencing grammatical structures, specifically supporting a "learned borrowing" hypothesis over "contact through bilingualism."
What Problem Does This Paper Address?
Many linguistic phenomena of interest, particularly those related to language contact and change, are difficult to study computationally. This is because these phenomena often occur in low-resource languages or historical language varieties, which lack the extensive digital corpora and linguistic annotations that modern data-driven NLP methods rely on. The authors aim to bridge this gap by demonstrating how computational techniques can be adapted to investigate such cases, using Early-Modern Frisian as a testbed. Specifically, they focus on verb cluster word order variation, a phenomenon well-studied in resource-rich languages like Dutch, to understand how contact with Dutch might have influenced Frisian grammar.
Methods
The core of the methodology involves adapting computational models developed for modern Dutch to analyze Early-Modern Frisian. The authors first identify features known to be relevant for verb cluster word order variation in Dutch. They then apply these features to a corpus of Early-Modern Frisian texts, using a rule-based script to automatically extract verb clusters and their associated features. Logistic regression is employed to model the probability of different verb cluster orders (1-2 vs. 2-1) based on these features. This approach allows for a comparison of the factors influencing word order in both languages, enabling them to test hypotheses about the nature of language contact.
Datasets
The primary dataset used is the Early-Modern Frisian section of the Frisian Integrated Language Database (Taaldatabank, TDB) compiled by the Fryske Akademy. This corpus contains approximately 125,842 tokens and 10,405 types of texts from 1550-1800, which have been manually tokenized, lemmatized, and part-of-speech tagged. The authors also draw upon existing research and models developed for modern Dutch corpora, though specific details of these Dutch datasets are not elaborated upon in this paper beyond their use in prior work.
Key Insights and Ideas
- Computational Linguistics for Low-Resource Historical Languages: The paper demonstrates the feasibility of applying computational methods to study historical language change in low-resource languages, even with limited data and annotation.
- Testing Hypotheses on Language Contact: By comparing computational models of Frisian and Dutch verb cluster word order, the study provides a quantitative way to distinguish between different theories of language contact (e.g., "learned borrowing" vs. "contact through bilingualism").
- Feature Transferability: The success in adapting features from a well-resourced language (Dutch) to a low-resource one (Frisian) highlights the potential for cross-lingual learning in historical linguistics.
- Stylistic Marking vs. Processing Ease: The findings suggest that verb cluster order variation in Early-Modern Frisian might have been driven more by stylistic factors (like rhyme) rather than by processing ease, a key driver in modern Dutch.
Key Results
The study found that while some features from the Dutch verb cluster model were not significant in Early-Modern Frisian, others showed significant associations. Notably, rhyming text was found to strongly favor the 1-2 order (odds ratio of 18.69), suggesting a stylistic function. The authors also observed that the association of copular verbs with the 2-1 order in Frisian was the reverse of what is seen in Dutch, which they interpret as evidence supporting the "learned borrowing" hypothesis. Aspectual and to-infinitival verb effects were consistent with previous observations suggesting these constructions might have been borrowed.
Limitations and Future Work
The authors acknowledge several limitations. The lack of comprehensive syntactic annotation in the Frisian corpus meant that some features used in the Dutch model (e.g., tree depth, extraposition) could not be extracted. The limited size of the Early-Modern Frisian corpus also prevented the reliable operationalization of factors like information density through language modeling. Direct, number-by-number comparison to the Dutch model was also hindered by different corpus annotation schemes and feature sets.
Future work could involve studying older Frisian texts to investigate whether non-contact hypotheses for verb cluster variation are plausible. Further research could also explore other linguistic phenomena in low-resource historical languages using similar computational approaches.
Summary updated