TL;DR
Researchers have developed a "Definitions as Anchors" approach that uses modern AI (BERT) to track how the meaning of sociopolitical concepts—like "privacy" or "climate"—evolves over centuries. By measuring a word's "distance" from two dictionary-defined anchor terms, they create intuitive visualizations that allow historians to see exactly when and how a concept shifted its cultural weight.
What Problem Does This Paper Address?
Conceptual history is the study of how the "big ideas" that shape society (e.g., justice, democracy, progress) change over time. Traditionally, historians did this by painstakingly analyzing old dictionaries and texts.
While modern Natural Language Processing (NLP) offers tools like word embeddings to automate this, existing methods often fall short for historians. Most current techniques are "ad hoc"—they look for a word’s nearest neighbors in a dataset without any prior context. This often results in "black box" outputs that are hard to interpret or require massive amounts of historical data to train specialized models for every decade. The authors wanted a method that is easy to use, interpretable, and grounded in the existing practices of historians.
Methods
The authors propose a simple but clever "anchor-based" approach:
- Identify Anchors: For a "key concept" (like Privacy), they look at a contemporary dictionary to find two primary, distinct dimensions of its meaning. For privacy, these were "Seclusion" and "Freedom."
- Contextual Embeddings: Instead of using static word vectors (which give a word one fixed value), they use BERT (bert-base-uncased). BERT creates "contextual" embeddings, meaning the vector for the word "privacy" changes based on the sentence it's in.
- The Similarity Metric: They calculate the average cosine similarity between the key concept (in its historical context) and the two fixed anchor words.
- The Difference Plot: The core innovation is plotting the difference between these two similarities over time. If the line moves up, the concept is becoming more like Anchor A; if it moves down, it's becoming more like Anchor B.
- Statistical Rigor: They use the Mann-Kendall test to confirm if the observed trends are statistically significant "monotonic" shifts (moving consistently in one direction) rather than just random noise.
Datasets
The study utilizes the Corpus of Historical American English (COHA), specifically focusing on the period between 1900 and 2000. COHA is a well-known diachronic corpus containing roughly 400 million words from various genres (fiction, magazines, newspapers, non-fiction).
Key Insights and Ideas
- Dictionary-Grounded NLP: By using dictionary definitions to set the "anchors," the researchers bridge the gap between traditional qualitative history and quantitative data science.
- Zero-Shot Historical Analysis: Surprisingly, the researchers found that a BERT model pre-trained on modern data (Wikipedia and BookCorpus) was effective at capturing historical shifts without needing to be fine-tuned on the historical corpus itself. This is a huge win for researchers working with "small data" historical periods.
- Interpretability over Complexity: While other models might produce complex 3D clusters of "word neighbors," the 1D "difference plot" used here is immediately readable for a historian. It tells a clear story of a concept's trajectory.
Key Results
The authors demonstrated their approach on several concepts with striking results:
- Privacy: In the early 1900s, "privacy" was more closely associated with seclusion (being alone). Around the 1960s, this reversed, and the concept became more strongly associated with freedom (the right to be left alone/autonomy). This shift was statistically significant ($p < .001$).
- Climate: This concept showed the most dramatic shift. Historically, "climate" was a local, geographic term (identical to geodetic position). In the late 20th century, it shifted sharply toward a global meaning associated with planetary weather systems and climate science ($p < .001$).
- Fairness: The data showed that "fairness" consistently leans more toward honesty than equality, though the gap between the two has narrowed slightly over the century.
- Peace: Interestingly, "peace" did not show a single monotonic trend over the whole century, highlighting that some concepts fluctuate rather than move in one steady direction.
Limitations and Future Work
- Monotonic Bias: The statistical tests used (Mann-Kendall) are designed to find trends moving in one direction. They struggle with concepts that "flip-flop" or change direction multiple times.
- Anchor Stability: The study uses modern dictionary definitions to define anchors. Future work could explore using historical anchors to see if the "goalposts" of meaning were also moving.
- Domain Specificity: While the general BERT model worked well, the authors suggest that "histBERT" (a model pre-trained specifically on historical text) might provide even more nuance in future studies.
- Change Point Detection: The authors aim to incorporate methods that can automatically detect the specific year a concept's meaning "tipped" from one anchor toward another.
Summary updated