A conceptual framework for ideology in online discourse beyond the left and right

arXiv:2603.18945 · cs.CY, cs.CL · Submitted 2026-03-19 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "A conceptual framework for ideology in online discourse beyond the left and right".

Jane: NLP+CSS work has operationalized ideology almost exclusively on a left/right partisan axis, obscuring the fact that people hold interpretations of many different complex and more specific ideologies.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So everyone, let's talk about the title of this new paper by Kenneth Joseph at Buffalo and Kim Williams at Portland State University. It's "A conceptual framework for ideology in online discourse beyond the left and right," which immediately tells us they are trying to move past that old binary thinking.

Jane: Exactly, Tom; it sounds like they're saying the way we’ve been studying ideology with just left or right labels is too simplistic for what people actually hold. It suggests there are many more specific ideologies out there that dimensional models just can't handle.

Lu: I find the authors' move to introduce an "attributed, multi-level socio-cognitive concept network" really fascinating; it’s a much richer way to think about ideology than just a single label.

Meng: From an engineering side, I wonder how they actually build this network; is it something that can handle the sheer complexity of real human language without getting bogged down in too many parameters?

Lalam: I'm curious how this framework will help us improve our systems because right now, our models often just output a score on a single axis.

Tom: Right, and what they propose is that ideology isn't one thing at all, but a web of values and attitudes that are connected in specific ways. It sets the stage for understanding how these different concepts interact in real discourse.

The paper's summary: Jane: So, the core idea they lay out is that ideologies are shared, attributed, multi-level networks of beliefs and rationalizations that dictate how people interpret issues like climate change or gender.

Tom: That’s a big concept; it moves us away from just looking at what someone says about an issue and starts looking at the whole structure of their worldview. It explains *why* people hold those interpretations, not just *what* they say.

Lu: I think the authors are doing a lot of work by defining what counts as an ideology through coherence—they suggest it's a collection of concepts that fit together in some structured way.

Meng: So, if we’re looking at their definition, they seem to be trying to give us concrete criteria for when a set of concepts actually qualifies as an ideology. That sounds like they are trying to solve a measurement problem before they even start building the structure.

Lalam: I see how that helps with practical application; if we can define what makes something coherent conceptually, it gives us better constraints to guide our modeling efforts.

Tom: Precisely, and beyond just defining what an ideology is, they show how this framework connects different NLP tasks like stance detection and framing through this ideological lens. It shows these tasks aren't isolated; they all point toward a larger ideological system.

The paper's improvements: Jane: The paper points out that their main improvement is providing a unifying lens that clarifies the overlaps between existing NLP tasks, showing how stance detection and natural language inference are actually complementary measurements of these complex systems.

Tom: That’s interesting because it means we don't have to treat those tasks as totally separate entities anymore; they feed into each other through ideology. It also suggests new research directions for detecting ideologies in both humans and large language models.

Lu: The framework helps guide us by asking some key questions, like what concepts can be part of an ideology and whether relationships between them should be defined by entailment or composition.

Meng: I'm thinking about the practical impact of that; if we can model the relationships as necessary components versus just social associations, it gives us a much tighter structure for feature engineering.

Lalam: From a modeling perspective, being able to distinguish between those two types of relationships seems crucial for building robust structures that aren't just superficial correlations.

Tom: And they also tackle the dual problem in sociology—how to tell if an actor aligns with an ideology when ideology isn't directly tied to social identity. That’s a tough one they try to address.

Conclusion: Jane: To wrap up, the paper on "A conceptual framework for ideology in online discourse beyond the left and right" shows that we need to look at ideology as a multi-level network rather than a simple spectrum.

Tom: It really emphasizes that people hold many different specific interpretations, and this framework helps us map those complex configurations instead of forcing everything into one category.

Lu: The paper’s main contribution is providing this conceptual bridge between computational methods and ideology theory, giving us a starting point for studying social discourse more deeply.

Meng: For me, the practical value is that it gives engineers a way to build models that are sensitive to these underlying structural relationships instead of just surface-level text patterns.

Lalam: I think this level of detail will allow our AI to potentially develop systems that understand the deeper cultural nuances and how they manifest in communication, which could significantly improve how we interact with the world.

Tom: It’s a lot to process, but it's clear that moving beyond the left or right axis is necessary for getting a richer picture of online discourse. We'll keep looking at these ideas as we move on to other important work.

University at Buffalo · Portland State University · Northeastern University

cs.CY, cs.CL

Submitted: 2026-03-19

Updated: 2026-10-06

Importance score: 81/100

The gist: NLP+CSS work has operationalized ideology almost exclusively on a left/right partisan axis, obscuring the fact that people hold interpretations of many different complex and more specific ideologies.

Key concepts

Ideology
Socially shared networks of values, attitudes, beliefs, and rationalizations that shape how individuals interpret social issues. It is not a single idea but a complex structure composed of many interconnected elements that justify certain interpretations.
Entailment
A relationship between concepts representing socially-shared cognitive associations. If concept A implies concept B in an ideological context, this means they are linked through shared thinking patterns within that social system.
Composition
A relationship where one concept is a necessary component of another. This signifies that certain elements must be present for a larger ideological structure or belief to be considered coherent and functional.

Terminology

Summary

NLP+CSS work has operationalized ideology almost exclusively on a left/right partisan axis, obscuring the fact that people hold interpretations of many different complex and more specific ideologies. This paper introduces a framework that understands ideology as an attributed, multi-level socio-cognitive concept network and explains how it manifests in discourse in relation to other relevant social processes like framing.

The gist

Ideologies are socialy shared, attributed, multi-level networks of values, attitudes, beliefs, and rationalizations that shape how individuals interpret and act on social issues.

A Framework for Studying Ideology

The proposed framework is built around answering four key questions to provide a starting point for studying ideology conceptually:

  1. Which concepts can be part of an ideology? This includes elements like values, attitudes, beliefs and their rationalizations.

  2. Which relationships between concepts should be considered? These are defined as entailment and composition. Entailment represents socially-shared cognitive associations, while composition signifies where one concept is a necessary component of another.

  3. Which criteria should be used to determine whether a set of concepts and their relations forms a coherent ideology? The framework defines four types of coherence: subject matter coherence, temporal coherence, social coherence, and conceptual coherence.

  4. When does (n’t) an actor’s discourse tell us they align with a particular ideology? This addresses the dual problem in sociology, considering when actors have an ideology given that ideologies are not deterministically connected to social identities.

Concept Definitions and Separation of Constructs

The framework distinguishes between elements that constitute ideology and constructs frequently analyzed in NLP+CSS, such as expressed issue stance, social identity, and frames. Ideological concepts include Value [A]bstract ideals, Attitude 'Tendencies to evaluate an object positively or negatively', Belief A determination about the truth of a statement, and Rationalization [R]easons that justify the truth of the relevant beliefs. Conversely, expressed issue stance, social identity, and frames are treated as processes that interact with ideology rather than components of it. For example, framing is defined as the act of selecting and emphasizing particular aspects of a situation or issue, while ideology refers to the underlying system of meaning being communicated.

Modeling Relationships and Coherence

The framework specifies how ideological concepts relate through two types: entailment (implications) and composition (necessity). Ideologies are modeled as attributed multi-level subnetworks, meaning both nodes and edges can have attributes. Values are continuous, while attitudes and entailment relationships are valenced. Coherence is defined by four criteria: subject matter coherence (existence around domains of power struggle), temporal coherence (persistence over time), social coherence (being socially shared but imperfectly aligned), and conceptual coherence (structural and logical consistency).

Operationalization Steps

The framework operationalizes the study of ideology in two main tasks: extracting noisy concepts and relations from discourse, and then inferring ideological structure from these observations. The first task yields partial observations of an attributed, multi-level network. The second involves modeling problem over those observations, where ideologies are modeled as separate draws from a distribution over attributed multi-level networks. This approach connects to statistical network models and graph learning in NLP. Theory informs this process by defining candidate ideologies and their core components, which can serve as priors or constraints in the modeling of coherence.

Implications for NLP+CSS

The framework suggests several directions for research: enriching ideology detection by treating ideologies as time-varying and socially shared networks, enabling new links between existing tasks like stance detection and values by showing they are complementary measurements of rich and complex ideological systems, and helping to reinterpret existing work, such as LLM studies, by allowing researchers to consider that models may exhibit multiplex ideological leanings rather than a single scalar score. The paper argues for moving beyond the left/right spectrum to study the complex and varying conceptual configurations of ideologies.

Limitations

The framework acknowledges limitations, including providing no formal mathematical specification of an ideology and not extensively exploring the tenuous connections between ideology measures and actual behavior in these models. However, it concludes that approaching ideology with this complexity is necessary because reduction to a spectrum misses what matters.

**(Note: The summary above adheres strictly to the constraints provided, focusing only on information present in the text and following the specified structure.

Improvements for AI systems

Based on the conceptual framework presented in the paper, here are specific improvements that could be made to existing AI systems, categorized by how they address their current limitations:

) 1. Shift from Left/Right Partisan Classification to Multi-Dimensional Ideological Modeling:

Current AI systems often reduce ideology to a binary left vs. right axis (as noted in the abstract). The proposed framework suggests moving beyond this scalar representation.

  • The improved system would be able to identify and model individuals or texts as residing within complex, multi-level ideological networks rather than a single point on a spectrum.

  • This allows for the detection of niche or specific ideologies (e.g., intersectional race ideologies, climate change skepticism) that are currently obscured by the binary classification.

) 2. Implement Concept and Relation Extraction for Ideological Structure:

Current NLP tasks like stance detection only identify a position on an issue, not the underlying conceptual structure justifying it.

  • The improved system would perform concept extraction to identify core ideological elements: values (e.g., freedom), beliefs (e.g., structural racism), and rationalizations (justifications).

  • It would then map the relationships between these extracted concepts using an entailment/composition framework, allowing the AI to understand how a belief necessitates a rationalization, or how one value supports another.

) 3. Integrate Framing Analysis with Ideological Substructure:

Current framing analysis often treats frames as separate entities from ideology.

  • The improved system would analyze cognitive frames (the mental structures activated by text) and determine which specific ideologies are embedded within those frames (e.g., identifying that a specific news frame simultaneously activates concepts from both colorblindness ideology and intersectional ideology).

) 4. Develop Context-Aware Coherence Checks:

The system should move beyond simple pattern matching to check for ideological coherence across different dimensions:

  • It would assess subject matter coherence by ensuring the text aligns with the power struggles inherent in a specific social domain (e.g., political discourse).

  • It would check temporal coherence by tracking how a concept's valence or relational structure changes over time within a corpus, allowing it to distinguish between persistent core beliefs and transient discursive shifts.

) 5. Refine Ideology Detection as System Identification:

Instead of simply labeling text with an ideology, the system should be designed to treat ideology itself as a latent, evolving network:

  • The AI would be trained not just to locate ideological texts, but to identify the systems (networks) present in a corpus.

  • This allows for detecting whether an entire ideological structure (defined by its set of concepts and their relations) is present, rather than just individual textual instances.

) 6. Enable Nuanced LLM Analysis:

When analyzing Large Language Models (LLMs), the system can move beyond simple bias scores.

  • It can probe whether an LLM exhibits multiplex ideological leanings—expressing components of multiple competing ideologies simultaneously across different domains, rather than being strictly aligned with a single scalar position.

  • This allows researchers to investigate complex interactions between identity performance, framing choices, and underlying cognitive processing.

The improved AI system can perform the following functions:

  1. Identify and map individuals or texts onto a detailed, multi-level network of socio-cognitive concepts (values, beliefs, rationalizations).

  2. Extract the structural relationships (entailment/composition) between these concepts to understand the underlying logic of an ideology.

  3. Analyze how cognitive frames select and emphasize certain ideologies embedded within a piece of discourse.

  4. Determine if a corpus contains coherent ideological systems by checking for subject matter, temporal, social, and conceptual coherence constraints.

  5. Provide a nuanced assessment of LLM behavior by detecting multiplex ideological leanings rather than simple binary alignment scores.

Sources

Related papers