TL;DR
This paper uses computational methods to analyze how the meanings of two Dutch near-synonyms for "to destroy" (vernielen and vernietigen) have evolved over centuries. By applying distributional semantics to historical texts, the authors demonstrate how semantic and onomasiological changes interact, leading to a clearer divergence in the core meanings of these words over time.
What Problem Does This Paper Address?
The paper tackles the challenge of understanding and computationally modeling semantic change in language, specifically focusing on how the meanings of near-synonyms diverge over time. While Cognitive Linguistics offers theoretical frameworks for analyzing semantic change (semasiology and onomasiology), bridging these insights with computational methods for diachronic analysis has been an ongoing area of research. The authors aim to show how distributional models can be used to automatically track these changes in historical corpora, providing empirical evidence for theoretical claims about how language evolves. They focus on the Dutch near-synonyms vernielen and vernietigen, which both translate to "to destroy," to investigate how their distinct prototypical meanings have solidified and their interchangeability has decreased over time.
Methods
The authors employ a computational approach based on distributional semantics to analyze diachronic semantic change. They construct vector space models for tokens (word forms) within different historical subcorpora. The core of their method involves:
- Bag-of-words model: Representing word meaning based on the surrounding context words.
- First-order and second-order vectors: Capturing both the immediate context of a target word and the semantic relationships between context words.
- Positive Pointwise Mutual Information (PPMI): An association strength measure used to weight context words, favoring those that are semantically relevant.
- Hierarchical Clustering: Grouping similar token vectors to identify semantic clusters within each time period.
- Contextual Analysis: Analyzing the most frequent context words within each cluster to interpret the semantic meaning of that cluster and how the target verbs are used within it.
This methodology is largely based on prior work by Montes et al. (2021) for contemporary data, adapted for historical corpora by using word forms instead of lemmas and omitting part-of-speech filters due to the lack of lemmatization and POS-tagging in historical Dutch.
Datasets
The study utilizes a corpus of Dutch prose texts from the Digitale Bibliotheek voor de Nederlandse Letteren (DBNL). The corpus was divided into subcorpora for the 16th, 17th, 18th, 19th, and 20th centuries. Due to data sparseness, the 16th and 17th-century subcorpora were combined for analysis. A random sample of N = 400 tokens for each of the target verbs (vernielen and vernietigen) was extracted from each subcorpus.
Key Insights and Ideas
- Divergence of Near-Synonyms: The study provides computational evidence that near-synonyms, even those with similar core meanings, can diverge significantly over time, solidifying their distinct prototypical uses.
- Interaction of Semasiological and Onomasiological Change: The findings support the theoretical claim that semasiological change (a word's meaning changing) and onomasiological change (the reorganization of semantic fields) are intertwined. As the core meanings of vernielen and vernietigen became more distinct, their interchangeability decreased.
- Prototypicality as a Driver of Change: The analysis highlights how the prototypical semantic core of each verb influences its diachronic development. Vernielen increasingly became associated with concrete destruction (e.g., buildings), while vernietigen gravitated towards abstract destruction (e.g., plans, decisions).
- Computational Tracking of Semantic Nuances: Distributional models, when applied to historical data, can effectively capture subtle semantic shifts and the emergence of new contextual uses for words.
- Retreat to Prototypical Cores: Over time, both verbs appear to have retreated to their more prototypical uses, leading to less overlap and interchangeability.
Key Results
- 16th/17th Century: High interchangeability between vernielen and vernietigen, with vernielen being slightly more frequent in contexts related to killing persons and concrete objects.
- 18th Century: The variants begin to recede to their prototypical cores. Vernielen becomes more associated with concrete objects (buildings, often by fire), while vernietigen starts appearing more in abstract contexts.
- 19th Century: This period aligns with previous linguistic analyses. Vernielen remains dominant for concrete destruction, while vernietigen takes over contexts related to the destruction of persons and abstract concepts.
- 20th Century: Vernietigen is significantly more frequent. Vernielen has decreased dramatically in frequency, with its prototypical use (destruction of buildings) no longer clearly distinguishable in the data. Vernietigen shows a diverse range of uses, including abstract concepts and war-related contexts.
Limitations and Future Work
- Parameter Sensitivity: The authors acknowledge that the chosen parameter settings for the distributional models might not be optimal for all near-synonym pairs or linguistic phenomena. Future work could explore a broader range of parameter settings.
- Diachronic Contextualized Embeddings: The current analysis uses word forms. The availability of diachronic contextualized word embeddings for Dutch would allow for a more nuanced analysis, potentially treating lemmas and accounting for syntactic filters.
- Broader Linguistic Phenomena: The methodology could be extended to analyze other types of linguistic alternations and near-synonym pairs in Dutch and other languages.
- Interplay of Formal and Conceptual Change: While the study demonstrates the interaction between semasiological and onomasiological change, further research could delve deeper into the specific mechanisms driving these conceptual shifts.
Summary updated