TL;DR
This paper introduces DiaHClust, a novel data-driven method for identifying distinct stages of language change. By iteratively applying hierarchical clustering to syntactic features, DiaHClust offers a more nuanced and accurate periodization than traditional, arbitrary methods, revealing gradual shifts and potential genre effects in language evolution.
What Problem Does This Paper Address?
Historical linguistics often relies on pre-defined, often arbitrary, time periods (e.g., "Old English," "Middle English") to study language change. This "periodization problem" can obscure gradual shifts, misrepresent transitional stages, and lead to inaccurate conclusions about the pace and nature of linguistic evolution. Traditional methods may also incorporate non-linguistic factors, further divorcing the periodization from the actual linguistic data. While data-driven approaches have emerged, this paper aims to refine them, particularly for syntactic change, by offering a more granular and iterative method that avoids initial data aggregation biases.
Methods
The core of DiaHClust builds upon the Variability-based Neighbor Clustering (VNC) approach developed by Gries and Hilpert. VNC is a hierarchical clustering method sensitive to temporal ordering, grouping temporally adjacent data points with similar linguistic characteristics.
DiaHClust extends this by:
- Vectorizing Syntactic Change: Instead of focusing on a single linguistic feature, DiaHClust creates vectors for each text that capture the relative frequencies of multiple interrelated syntactic features. This provides a richer model of the language's syntactic system at different points in time.
- Iterative Clustering: A key innovation is the addition of an iterative layer of clustering. After an initial clustering (using VNC), if the number of identified clusters is large (e.g., >10), the method iteratively aggregates data points within clusters and re-clusters. This allows for tracing the clustering process from the individual text level up to broader time stages.
- Silhouette Coefficient for Validation: To objectively determine the optimal number of clusters at each iteration, DiaHClust uses silhouette coefficients. This metric measures how well an object (text) is clustered with respect to its own cluster versus other clusters, providing a quantitative basis for cluster validation.
- Handling Outliers and Genre Effects: The iterative approach and the ability to examine cluster composition at each stage facilitate the identification of outliers and genre-specific effects that might otherwise be masked.
Datasets
The primary dataset used is the Icelandic Parsed Historical Corpus (IcePaHC). This corpus contains approximately 1,000,000 words from 61 text extracts spanning 10 centuries (1150-2008), covering various stages of Icelandic. The authors note that while valuable, IcePaHC has limitations, including a relatively small sample of historical texts, uneven distribution across time, and a bias towards narrative texts, making it an ideal test case for their method.
Key Insights and Ideas
- Gradualism in Icelandic Syntax: Contrary to some previous assumptions of abrupt change, DiaHClust reveals that syntactic change in Icelandic is more gradual, with shifts occurring over extended periods rather than at specific turning points.
- Genre Effects are Detectable: The iterative nature of DiaHClust allows for the easy examination of cluster composition, revealing how specific genres (e.g., religious texts) can influence the linguistic patterns within a particular time stage. This is a significant advantage over top-down periodization, which can obscure such effects.
- Outlier Identification: The method effectively highlights outlier texts that deviate from the general trends, providing opportunities for further investigation into their unique characteristics.
- Data-Driven Periodization is Feasible for Sparse Data: The successful application to IcePaHC, a corpus with data sparsity and subtle syntactic changes, demonstrates the robustness of DiaHClust for languages where data is less abundant or changes are less dramatic than in, for example, English.
Key Results
The paper demonstrates DiaHClust's effectiveness through a case study on syntactic change in Icelandic. After an initial clustering yielding 28 potential stages, the iterative process refined this to 5 well-clustered time stages: 1150-1210, 1250-1450, 1475–1630, 1650–1882, and 1883–2008. This periodization provided a more nuanced view of syntactic changes (dative subjects, expletives, V1, prefinite subjects, VO order, Stylistic Fronting) compared to previous studies using traditional periodization schemes. For instance, the frequency of expletives showed a more gradual increase from 0.2% to 1.5% across the DiaHClust stages, whereas a traditional scheme might have shown a sharper jump. The average silhouette coefficient improved from 0.4 to 0.5 after removing an identified outlier, indicating a more coherent clustering.
Limitations and Future Work
- Computational Cost: While not explicitly detailed as a limitation, iterative clustering and calculating silhouette coefficients can be computationally intensive, especially with very large datasets.
- Subjectivity in Iteration Cut-off: Although silhouette coefficients provide objective guidance, the decision to stop the iterative clustering (e.g., at <10 clusters) still involves a researcher's judgment based on interpretability and the purpose of the analysis.
- Future Work: The authors suggest that testing DiaHClust on a language like English, which has more extensive corpora and more pronounced syntactic changes, would be a valuable next step. They also imply that the method could be adapted for other types of linguistic change beyond syntax.
Summary updated