Polarization captured by Community Notes’ latent ideology in multi-dimensional settings

Determine which political polarization dimension(s) are encoded by the latent ideology parameters (theta_n for notes and theta_r for raters) learned by X’s Community Notes matrix factorization (hat_eta_{rn} = beta_0 + beta_n + beta_r + theta_n · theta_r) in countries where political competition is organized along multiple ideological dimensions, and quantify how this encoding affects the system’s moderation outcomes (e.g., the likelihood that a proposed note attains Helpful Status).

Background

X’s Community Notes system learns latent ideological positions for notes and raters from rating data and uses these positions to select notes that garner cross-partisan support. This approach was developed and validated in the United States, where a single Left–Right ideological dimension often structures political disagreement.

In many countries, political competition is organized along multiple dimensions (e.g., Left–Right and anti-elite). Without explicit calibration, it is unclear which form(s) of polarization the learned latent ideology captures internationally and how this mapping impacts algorithmic moderation performance, particularly the selection of notes that reach Helpful Status.