Discovering Conceptual Metaphors Across Topics and Media Types

Human thought often relies on a hidden scaffolding of comparisons. We understand complex, abstract ideas by mapping them onto simpler, physical experiences we know well. This is the core of Conceptual Metaphor Theory, a framework suggesting that when we speak about things like "paying taxes" or "political campaigns," we are not just using flowery language; we are unconsciously borrowing the logic of physical objects, journeys, or battles to make sense of the world. If a speaker describes taxes as a heavy burden they must carry, they are framing the issue as a physical weight, which naturally leads to a desire to put it down. If they describe the same taxes as an investment in a community, the logic shifts entirely to one of growth and future return. These underlying mental maps shape how people reason, argue, and perceive reality, yet they are difficult to see directly because they are buried beneath the surface of everyday speech.

Researchers at the University of Colorado Boulder have developed a new way to uncover these invisible maps by analyzing thousands of spoken and written words without human intervention. Instead of asking people what they mean, the team built a computer system that reads vast collections of text, identifies the specific words used metaphorically, and then groups them together based on the deeper logic they share. The system first finds pairs of words where a verb is used in a non-literal way, such as a politician "wielding" power or an economy "crashing." It then asks a large language model to explain what these phrases imply about the problem, the cause, and the solution, effectively translating the metaphor into a clear statement of the speaker's viewpoint. Finally, the system sorts these interpretations into clusters, but with a strict rule: it refuses to group metaphors together if they describe fundamentally different physical actions, even if they talk about the same topic. This ensures that the resulting groups reflect genuine similarities in how people are thinking, rather than just superficial similarities in the words they chose.

When the researchers applied this method to a massive dataset of political podcast transcripts, they found distinct patterns in how left-leaning and right-leaning speakers frame the world. The analysis revealed that left-leaning hosts frequently conceptualize information, media stories, and political power as weapons. They often use verbs like "crush," "wield," and "seize" to describe how these forces are used against one another. In contrast, right-leaning sources were far more likely to discuss economic issues as systems subject to vertical changes, using language that suggests things are rising or falling, or that the economy is a structure that can be built up or torn down. The study also showed that left-leaning speakers were more prone to discussing political violence through metaphorical language, while right-leaning speakers focused more heavily on economic themes.

The team tested their method on other topics, including immigration, gun control, and abortion, to see if it could recover well-known patterns of thought. In the immigration data, the system successfully identified the long-documented metaphor of immigrants as parasites or natural disasters, grouping together phrases about "flooding" borders or "draining" resources. It also found that right-leaning sources often framed immigration as an active invasion or a crisis being fought, whereas left-leaning sources used different metaphors to describe the same events. The researchers confirmed that these groupings were not random; when they used the frequency of these metaphor clusters to predict the political bias of a podcast episode, the system could accurately distinguish between left and right sources. This suggests that the way people choose their metaphors is a reliable signal of their underlying ideological framework.

The study does not claim to have solved the problem of media bias or to have found every metaphor in existence. The researchers acknowledge that their method relies on automated tools that can make mistakes, such as misidentifying a literal word as a metaphor or assigning a phrase to the wrong category. They also note that their analysis of podcasts covers a specific set of channels and a limited time period, so the findings may not apply to all media formats or all speakers. However, the results demonstrate that it is possible to use unsupervised machine learning to map the conceptual structures that guide public discourse. By revealing how different groups use the logic of physical objects, forces, and journeys to frame abstract issues, this approach offers a new way to understand the subtle, often unconscious, ways in which language shapes our political reality.