POLAR:A Per-User Association Test in Embedding Space

A research team led by Pedro Bento and nine other authors has introduced POLAR (Per-user On-axis Lexical Association Report), a novel method for detecting subtle linguistic associations at the individual user level. Unlike traditional probes that operate on entire corpora, POLAR works in the embedding space of a lightly adapted masked language model, representing each author with private deterministic tokens. The system projects these user vectors onto carefully curated lexical axes—like sets of words associated with specific ideologies or bot-like behavior—and reports standardized effects with statistical rigor using permutation p-values and Benjamini-Hochberg control for false discoveries.

In practical applications, POLAR demonstrated significant utility. On a balanced benchmark of bot and human Twitter accounts, the method cleanly separated LLM-driven bots from organic human users. In a more sensitive case study on an extremist forum, POLAR successfully quantified users' strong alignment with pre-defined slur lexicons and, critically, revealed a measurable 'rightward drift' in linguistic association over time. The authors emphasize the tool's modularity, allowing researchers to plug in new attribute sets for different investigations. All code is publicly available, positioning POLAR as a concise, per-author diagnostic for the field of computational social science.