TL;DR
This paper uses dynamic word embeddings to track the semantic evolution of Dutch words related to nationalism ("nation," "people," "fatherland") in literature from 1700-1880. They found that "vaderland" and "volk" experienced significant semantic shifts, particularly around periods of rising nationalism, while "natie" showed less change.
What Problem Does This Paper Address?
The paper tackles the challenge of understanding how abstract concepts like nationalism evolve within a culture over time, specifically focusing on the Dutch context during the "Sattelzeit" (roughly 1750-1850), a period marked by the rise of modernity and nationalism. While historical and conceptual analysis methods exist, this work aims to leverage computational linguistics, specifically dynamic word embeddings, to quantify and analyze semantic shifts in key nationalistic terms within a large literary corpus. The goal is to provide a data-driven perspective on how the meaning and connotations of words like "nation," "people," and "fatherland" changed as nationalistic sentiments grew.
Methods
The core methodology involves using Dynamic Bernoulli Word Embeddings, a technique designed to capture how word meanings change over time, even with sparse data. This model is an extension of Bernoulli embeddings that provides a conditional model for text data across different time periods. The authors chose this model because it has shown good predictive performance and the ability to track changes in rare words while keeping stable words consistent.
The process involved:
- Data Collection: Gathering a large corpus of Dutch fiction literature from the Digital Library of Dutch Language (DBNL).
- Preprocessing: Normalizing spelling, removing stop words, and filtering out very short or infrequent words to create a cleaner vocabulary for the model.
- Time Slicing: Dividing the literature into bins per decade to represent different time periods.
- Model Training: Training the Dynamic Bernoulli Embedding model on these time slices.
- Semantic Drift Measurement: Quantifying the "absolute drift" of the target words ("natie," "volk," "vaderland") to measure how much their contextual usage changed over time.
- Nearest Neighbor Analysis: Examining the words that most frequently co-occur with the target words in each time slice to understand the qualitative nature of the semantic shifts.
Datasets
The primary dataset used is the Digital Library of Dutch Language (DBNL). This library contains thousands of literary texts, including fiction, plays, and youth literature, from the Netherlands and Belgium. The authors specifically selected fiction from 1700 to 1880, compiling a dataset of 414 books. This data was divided into training (80%) and validation (20%) sets for model evaluation. No new benchmark datasets were introduced; the focus was on applying existing methods to a historical literary corpus.
Key Insights and Ideas
- Quantifying Semantic Change: The paper demonstrates that dynamic word embeddings can effectively quantify semantic shifts in historical texts, providing a data-driven complement to traditional historical analysis.
- "Vaderland" and "Volk" as Dynamic Concepts: The significant semantic drift observed for "vaderland" and "volk" suggests these terms were more fluid and responsive to cultural and political changes related to nationalism than "natie."
- Nationalism's Impact on Language: The analysis of nearest neighbors reveals how the connotations of "vaderland" and "volk" shifted towards more nationalistic and sometimes even negative (e.g., "people as a mob") meanings over time.
- "Natie" as a More Stable Concept: The relatively small drift for "natie" indicates that its core meaning or institutional associations remained more consistent throughout the period, or that its semantic evolution was subtler.
- Historical Context Matters: The peaks in semantic drift for "vaderland" align with significant historical events like the Dutch enlightenment and constitutional reforms, underscoring the link between socio-political developments and linguistic change.
Key Results
- Absolute Drifts:
- "Volk" showed the largest absolute drift (0.1800).
- "Vaderland" followed with a significant drift (0.1128).
- "Natie" had the smallest drift (0.0644).
- For comparison, the word "we" had an absolute drift of 0.3864.
- Drift Over Time: "Natie" showed a gradual increase in drift, while "vaderland" exhibited multiple peaks, notably around the Dutch enlightenment (1780s) and the constitutional reform of 1848.
- Nearest Neighbors:
- For "vaderland," neighbors evolved from terms like "country of birth" to include national entities like "Nederland" and "Vlaanderland" from the late 18th century onwards.
- For "volk," neighbors shifted from biblical themes to words like "rebellious" and "rebellion," indicating a more negative or volatile connotation over time.
- For "natie," neighbors consistently referred to institutions (universities, courts, governments) and concepts like "nationality."
Limitations and Future Work
Limitations: * Sense-Agnostic: The method is sense-agnostic, meaning it treats all occurrences of a word as having the same meaning. This can obscure nuanced semantic shifts within specific contexts. * Vocabulary Constraint: The analysis is limited to words present in the entire corpus across all time slices, potentially missing shifts in words that became popular or fell out of use within the period. * Spelling Normalization: While spelling normalization was applied, historical texts can have non-standard spellings that might have been missed, potentially affecting word embeddings.
Future Work: * Deeper Analysis of Semantic Shifts: Further exploration into how semantic shifts reflect nation-building, nationalism, and the arrival of ideologies. * Hyperparameter Optimization: A more exhaustive exploration of hyperparameter settings for the Dynamic Bernoulli Embedding model. * Sense-Specific Analysis: Developing methods to analyze semantic shifts at a more granular, sense-specific level.
Summary updated