Measuring the Semantic Structure and Evolution of Conspiracy Theories

The diagram summarizes four conspiracy categories and their neighboring words; Figure 1 situates those terms in the overall embedding space; below, choose a term to explore its top 10 most similar words.

The diagram groups the 19 conspiracy theories into four color-coded categories, each with illustrative neighboring words drawn from the embedding space:

The outer ring shows representative neighboring words for each conspiracy, illustrating how conspiratorial language forms coherent, semantically distinguishable regions.

Click any segment to zoom in, or click the center to zoom back out.

View poster online

Figure 1: Semantic landscape of conspiracy terms

Figure 1 maps the broader semantic landscape of conspiracy theories. Using a t-SNE projection of Word2Vec embeddings trained on Reddit data from 2012–2022, each point represents a word in a sampled vocabulary. Most points are non-conspiracy words (pink), forming a wide background of general political and cultural discourse.

Against that backdrop, colored points and labels mark where key conspiracy-related terms fall in embedding space. The figure shows that these terms do not occupy one uniform region: some cluster near related events and narratives (e.g., mass-shooting terms such as sandy_hook, boston_bombing, and sutherland_spring), while others sit in more distinct neighborhoods (e.g., flat_earth, chemtrails, illuminati, gamergate).

Interact with the plot: scroll to zoom, drag to pan, hover for word labels, and click legend entries to show or hide topics.

This global view motivates the neighbor analysis below. Select a conspiracy term to explore the words closest to it in the full high-dimensional embedding — the local semantic neighborhood that defines how that conspiracy is talked about in context.


Neighboring Words by Conspiracy Term

Left: interactive network with the conspiracy term at the center. Right: 2D map of the same words in embedding space (PCA); lines connect neighbors that are also similar to each other.

Neighbor network

Embedding map