How Different Communities Redefine the Same Words

34 concepts28 papers1 proposal

When two people from opposing political parties debate a topic, they appear to be speaking the same language. They use the same nouns, verbs, and grammatical structures. However, traditional natural language processing often falls victim to the illusion of a universal lexicon. Standard models treat text as a collection of static tokens, assuming that a word like "democracy" or "safety" maps to a single, shared semantic meaning.

In reality, words are highly malleable. Different demographic, political, and cultural groups frequently assign fundamentally different connotations, contexts, and functional roles to the exact same vocabulary. Rather than looking at how language changes over centuries—a historical process known as diachronic shift—a growing body of research focuses on synchronic semantic divergence: the way word meanings split across different communities at the exact same moment in time.

The Statistical Footprint of Perspective

How do we mathematically distinguish between text collections that cover different topics (such as sports versus finance) and those that cover the same topic from different perspectives (such as pro-choice versus pro-life)?

In a foundational study, Wei-Hao Lin 0003 analyzed document collections using Kullback-Leibler (KL) divergence and discovered that differing perspectives occupy a unique statistical "middle ground." Unrelated topics use entirely disjoint vocabularies, resulting in massive statistical distance. Conversely, opposing perspectives share a core vocabulary because they are fighting over the same conceptual turf. They use the same words but distribute their emphasis differently, creating a distinct, measurable footprint that separates ideological divergence from mere topical variation.

As Tianran Hu et al. point out in A World of Difference: Divergent Word Interpretations among People, traditional NLP relies heavily on word frequency counts. This approach completely misses how the underlying meaning of a single term shifts between social groups. To see these shifts, we must look at the relational structure of language.

Aligning Ideological Vector Spaces

If we train separate word embedding models on text from different communities—such as different political subreddits or news outlets—each model builds its own isolated vector space. Because the absolute coordinates in these spaces are arbitrary, we cannot compare them directly. However, we can align these multidimensional worldviews geometrically.

Once aligned, we can track how words shift their positions relative to their neighbors. As Jeremiah Milbauer et al. demonstrate in Aligning Multidimensional Worldviews and Discovering Ideological Differences, this alignment reveals "functional equivalents" across ideological dialects. For example, the word "Trump" in a conservative forum might occupy the exact same structural and relational position as "Clinton" in a liberal forum.

Similarly, Hosein Azarbonyad et al. developed a framework in Words are Malleable: Computing Semantic Shifts in Political and Media Discourse that combines linear mapping with neighbor-based graph analysis. The authors showed that words like "democracy" or "moral" carry distinct semantic signatures driven entirely by the speaker's political identity rather than the passage of time.

Quantifying Polarization Along Semantic Axes

Instead of just looking at a word's nearest neighbors, we can gain deeper intuition by projecting word vectors onto explicit semantic "axes" defined by antonym pairs (such as good vs. bad or vulnerable vs. threatening).

Jisun An et al. introduced this concept with SemAxis: A Lightweight Framework to Characterize Domain-Specific Word Semantics Beyond Sentiment. By projecting words onto hundreds of semantic dimensions, the authors mapped how community perceptions diverge on specific traits—such as whether a public figure is viewed as honorable versus creative across different online forums.

To measure how these associations differ between two distinct text corpora, Federico Bianchi et al. developed the Sliced Word Embedding Association Test (SWEAT) in SWEAT: Scoring Polarization of Topics across Different Corpora. SWEAT quantifies "relative polarization" by measuring how much a specific topic shifts toward opposing semantic poles when moving from one community's text to another.

A vivid real-world example of this phenomenon is documented by Keyu Chen et al. in Partisan US News Media Representations of Syrian Refugees. The authors found a stark semantic misalignment in media coverage: left-leaning outlets consistently aligned the term "refugees" with concepts of childhood and vulnerability, whereas right-leaning outlets shifted the exact same term toward security threats and terrorism.

Why Text Alone Fails

Because meaning is so deeply tied to the speaker's identity, standard text-only models face a fundamental bottleneck. Ideological differences leave measurable traces in the mathematical representations generated by modern language models. In One Word, Two Sides: Traces of Stance in Contextualized Word Representations, Aina Garí Soler et al. showed that contextualized language models represent identical words differently depending on the speaker's underlying stance, proving that individuals in a debate may use the same vocabulary while holding entirely different conceptual definitions.

This actor-dependent ambiguity means that ignoring who is speaking leads to severe classification errors. For instance, in legislative analysis, the exact same bill text can signal completely different political goals depending on its sponsor. To solve this, Anastassia Kornilova et al. developed a neural network model in Party Matters: Enhancing Legislative Embeddings with Author Attributes for Vote Prediction that explicitly injects author metadata into text embeddings, significantly improving vote prediction accuracy.

Ultimately, capturing the true nuance of language requires moving beyond isolated text strings. As Rajkumar Pujari et al. argue in Understanding Politics via Contextualized Discourse Processing, political discourse must be processed through holistic, graph-based frameworks that integrate the text with the speaker's identity, historical context, and ideological background. Without accounting for who is speaking, models remain blind to the hidden divides built into our everyday vocabulary.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_semantic_divergence_4b8b3334,
  title = {semantic divergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-divergence/rdr_4b8b333475f8465b8c13be401bd716f1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_semantic_divergence_4b8b3334_v1,
  title = {semantic divergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-divergence/rdr_4b8b333475f8465b8c13be401bd716f1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake