Computational Social Science & NLP

Learning the Geometry of Political Discourse through Contrastive Rhetorical Deformations

A continuous geometric modeling of doublespeak (langue de bois) in French political discourse (1981-1993 period). The approach replaces surface-level lexical analysis with contrastive representation learning based on the Multi-Contrastive Noise Contrastive Estimation (NCE) objective.

Author: Avner EL BAZ (Supplementary analysis: Gabriel FRANCOIS)

Data: 12,497 normalized documents, split into 450-token blocks

Actual Corpus Size

12,497

French political documents analyzed (Manifestos, Declarations).

Text Segmentation

450

Tokens per chunk, overlapping with a stride of 120 tokens.

Discursive Layers

3

Families of constraints studied: structural, conceptual, and informational.

Theoretical Activation Profile of Rhetorical Constraints

Multi-Contrastive NCE Objective & Functionals

The model learns to separate direct formulations \(P_i\) from their rhetorical deformations \(N_i\) via the following contrastive loss function:

\(\mathcal{L}(q,g_{i})=-log\frac{\sum_{p\in P_{i}}exp(sim(q,p)/\tau)}{\sum_{z\in P_{i}\cup N_{i}}exp(sim(q,z)/\tau)}\)

Two functionals allow for evaluating the latent space at the document level: the extreme sensitivity score (\(Score_{D}^{max}\)) and the distributional profile (\(Score_{D}^{prof}\)):

\(Score_{D}^{max}=\frac{1}{n}\sum_{j=1}^{n}\max_{g_i \in \mathcal{G}} \mathcal{L}(c_j, g_i)\)

Learning the Geometry of Political Discourse through Contrastive Rhetorical Deformations

Theoretical Activation Profile of Rhetorical Constraints

Multi-Contrastive NCE Objective & Functionals

Scientific Paper