A conceptual framework for ideology beyond the left and right

1. Problem Statement

Current Natural Language Processing (NLP) and Computational Social Science (CSS) research predominantly operationalizes ideology along a single left/right partisan axis (or simple multi-dimensional extensions). The authors argue this approach is insufficient because:

• Oversimplification: It reduces complex, specific ideologies (e.g., regarding race, climate, gender) to a single spectrum, obscuring the nuanced, structured systems of thought individuals hold.
• Theoretical Gap: Existing NLP tasks (like stance detection or natural language inference) are often treated in isolation without a unifying theoretical framework that explains how they relate to the construction of ideology.
• Conflation: Current methods often conflate ideology with social identity, issue stance, and framing, failing to distinguish between the underlying belief systems and the discursive expressions of those systems.

The paper addresses the challenge of identifying specific ideologies, defining their contents, and mapping discursive patterns to these ideologies in a way that accounts for their fluidity, social sharing, and internal coherence.

2. Methodology: A Conceptual Framework

The authors propose a novel framework that redefines ideology not as a point in a latent space, but as an attributed, multi-level socio-cognitive concept network.

Core Definitions and Structure

• Ideology as a Network: Ideologies are modeled as subnetworks of concepts (nodes) and relationships (edges) that are socially shared and temporally stable.
• Concept Hierarchy: The framework distinguishes between:
  • Ideological Concepts: Values (abstract ideals), Attitudes (evaluations), Beliefs (truth determinations), and Rationalizations (justifications). These form a partial hierarchy where lower levels (beliefs) support higher levels (values).
  • Non-Ideological Constructs: Social Identity (who we are), Expressed Issue Stance (policy preferences), and Frames (how issues are presented). These interact with ideology to produce discourse but are distinct from the ideology itself.
• Relationship Types:
  • Entailment: Socially shared cognitive associations (e.g., if Belief A is true, Belief B must be true).
  • Composition: Necessary components (e.g., a belief requires a rationalization).
• Attributed Multi-Level Subnetworks:
  • Attributed: Nodes and edges have attributes (e.g., values are continuous; attitudes and entailments are valenced positive/negative).
  • Subnetworks: An ideology is a subset of a larger concept space; individuals may hold overlapping subsets of different ideologies.

Criteria for Coherence

To determine if a set of concepts constitutes a coherent ideology, the framework proposes four constraints:

1. Subject Matter Coherence: Ideologies exist within specific domains of power struggle.
2. Temporal Coherence: Core concepts and relations remain stable over time, even if surface expressions change.
3. Partial Social Coherence: Ideologies are socially shared but imperfectly; they emerge at scale despite individual variations.
4. Conceptual Coherence: Concepts must be structurally connected (stronger internal links than external) and logically consistent (e.g., avoiding contradictory valences).

Operationalization Strategy

The authors suggest a two-step computational approach:

1. Extraction: Use NLP to extract noisy concepts and relations from discourse (e.g., via stance detection, value identification, or entailment models).
2. Inference: Model ideologies as distributions over attributed multi-level networks. Actors are modeled as distributions over these ideologies, and texts are samples conditioned on them. This treats ideology detection as a network inference problem under uncertainty, utilizing statistical network models and graph learning.

3. Key Contributions

• Theoretical Bridge: Provides the first actionable framework linking high-level ideology theory (from political science, sociology, and psychology) with specific NLP+CSS tasks.
• Reinterpretation of NLP Tasks: Demonstrates how disparate tasks like stance detection, value identification, and natural language inference (NLI) are complementary measurements of a single complex ideological system, rather than isolated problems.
• New Research Directions:
  • Ideology Detection: Shifts the goal from placing actors on a pre-defined spectrum to discovering which coherent ideological systems exist in a corpus and how they are organized.
  • LLM Analysis: Offers a new lens for analyzing Large Language Models, suggesting that LLMs may express "multiplex" ideologies (components of multiple ideologies) rather than a single scalar position, challenging current methods of measuring AI bias.
  • Uncertainty Modeling: Explicitly addresses the uncertainty in determining whether a discourse reflects genuine ideological commitment or merely identity performance/framing.

4. Results and Findings

Note: As a conceptual paper, it does not present empirical experimental results (e.g., accuracy scores on a dataset). Instead, its "results" are the logical derivations and theoretical insights derived from the framework.

• Limitations of Dimensional Models: The authors demonstrate that while left/right models work for broad US partisan analysis, they fail to capture the complexity of specific issue-based ideologies (e.g., colorblind vs. intersectional racial ideologies) and global contexts.
• Framework Utility: The framework successfully differentiates between frames (discursive choices) and ideologies (underlying belief systems), allowing researchers to ask new questions about how cognitive frames induce specific ideological constellations.
• Feasibility of Network Modeling: The paper argues that existing NLP tools (graph learning, distant supervision) are sufficient to operationalize this framework, provided they are grounded in social theory to define candidate concepts and coherence constraints.

5. Significance

• Paradigm Shift: Moves the field away from "reductionist" dimensional models toward "complexity-aware" network models of ideology.
• Interdisciplinary Synthesis: Unifies insights from Van Dijk's discourse analysis, Converse's political psychology, and modern NLP, creating a shared vocabulary for computational social scientists.
• Improved Measurement: By treating ideology as a network with specific coherence constraints, the framework allows for more nuanced measurement of social constructs, potentially leading to better detection of extremism, polarization, and social change.
• Future-Proofing: As LLMs become central to social discourse, this framework provides a necessary theoretical basis for understanding how these models process and generate complex, multi-faceted ideological content beyond simple binary classifications.

In conclusion, the paper argues that to truly understand social discourse, NLP researchers must stop treating ideology as a fixed coordinate and start treating it as a dynamic, attributed, multi-level network that requires both top-down theoretical constraints and bottom-up data-driven discovery.