How AI Tracks Word Meaning Changes Over Centuries

This article explores how computational historical linguistics leverages artificial intelligence and diachronic word embeddings to track the evolution of language. By converting historical texts into mathematical representations, researchers can measure precisely how words shift in meaning and connotation over centuries. You will learn the core mechanics behind word embeddings, how historical models are aligned across different eras, and how these technologies uncover large-scale cultural shifts.

Understanding Word Embeddings

To track how language changes, computers must first understand what words mean. Artificial intelligence achieves this through word embeddings. In natural language processing, a word embedding is a mathematical representation of a word, expressed as a vector—a long list of numbers.

AI models generate these vectors by analyzing massive amounts of text and applying the distributional hypothesis, which states that words used in similar contexts have similar meanings. The model places words into a multi-dimensional geometric space. In this space, words with similar meanings are grouped close together. For example, the vector for "king" will be mathematically close to "monarch" and "throne," while remaining far away from unrelated words like "bicycle."

What Are Diachronic Word Embeddings?

Standard word embeddings are static; they represent a snapshot of language at a single point in time. Diachronic word embeddings add the dimension of time to this process. Instead of training one AI model on all available text, researchers divide historical texts into distinct time slices, such as decades or centuries.

The AI then generates separate vector spaces for each era. By comparing the position of a target word in the 1800s vector space to its position in the 1900s vector space, linguists can mathematically observe how that word has moved. Its movement from one neighborhood of words to another directly reflects its change in meaning.

Aligning Historical Models

A major technical challenge in this process is that AI models trained on different eras generate incompatible geometric spaces. A word's coordinates in an 1850s model cannot be directly compared to its coordinates in a 1950s model because the AI initializes each space randomly.

To solve this, computational linguists use a mathematical technique called Orthogonal Procrustes alignment. This process rotates and scales the vector space of one time period to match another, using stable words—words whose meanings rarely change, like "tree," "water," or "hand"—as anchor points. Once the spaces are aligned on top of one another, any remaining distance between a word's past and present coordinates represents genuine semantic change.

Tracking Changes in Meaning

With aligned diachronic embeddings, tracking a word's definition becomes a matter of measuring geometric distance. Researchers calculate the cosine similarity between a word's vector in one century and its vector in another.

Consider the word "broadcast." In an early 19th-century vector space, "broadcast" sits firmly in the neighborhood of agricultural terms like "seed," "sow," and "field," reflecting its original meaning of scattering seeds by hand. By the mid-20th century, the vector for "broadcast" migrates across the geometric space to cluster with "radio," "television," and "news." AI allows linguists to pinpoint the exact decade this transition accelerated.

Measuring Shifts in Connotation

Words do not just change their literal definitions; they also change their emotional weight and social connotation. Diachronic embeddings track two primary types of connotative shift: pejoration (a word becoming more negative) and amelioration (a word becoming more positive).

To measure this, researchers establish "sentiment poles" within the vector space. One pole consists of consistently positive words (e.g., "good," "pure," "honest"), and the other consists of consistently negative words (e.g., "bad," "cruel," "evil"). Researchers then track how a target word moves in relation to these poles over time.

For instance, the word "awful" originally meant "worthy of awe" or "majestic." In a 17th-century vector space, it sits close to the positive pole and words like "divine." Over the centuries, AI models show the vector for "awful" drifting steadily away from the positive anchors and settling near the negative pole alongside "terrible" and "dreadful."

Moving to Contextual AI Models

Early diachronic research relied on static embedding models like Word2Vec, which assign a single vector to a word regardless of how many meanings it has. Today, computational historical linguistics increasingly uses advanced transformer models, such as BERT.

These modern AI systems generate contextual embeddings. Instead of giving the word "bank" one vector, the AI generates a unique vector for every single instance the word appears in a historical text, based on the surrounding sentence. This allows linguists to track polysemy—how words acquire multiple meanings over time. Researchers can observe a word birth a new slang meaning in a specific century while still retaining its original definition in formal contexts, offering an unprecedented, high-resolution view of human cultural and linguistic evolution.