So we built an LLM framework that reads four dimensions instead of one: bias, sensationalism, emotional appeal, framing. Friday @LREC2026 (PoliticalNLP).
Partly inspired, years back, by late-night chats with my old friend @MarcoRoberti link here https://t.co/fW6Hh4EhMb
50 political news articles. 17 outlets. RoBERTa labels 35 of them "neutral."
Not because the articles are neutral — coverage of a female PM ousted, diplomatic crises, Franco-German military rifts. The model just can't see what makes them political. Wrong tool for the job. 🥷
We call it neutral collapse. Journalistic prose hedges, attributes, balances — and sentiment models read all that as "no valence." One BBC piece on Australia's Liberal Party leadership lost the negative label by 0.02. A researcher studying gender in politics would never know.
Back when I ran political propaganda, the unwritten rule was: the source matters more than the fact. Every disinfo tool still works that way. Wrong.
This week @LREC2026 (InDor) I'm presenting our Multi-Layer Framework. 11 layers, claim and source on separate tracks.
Tested on the Russian op that tied Macron to Epstein last February. Same words, 3 problems on top of each other: an active fabrication, real reporting about it, and a low-grade outlet republishing wire copy without checking. A binary verdict averages them into noise. We don't. 🥷
Un giornalista del Guardian (!) ha la faccia tosta di chiedere a Francesca Albanese se l’uso del termine «dominio», riferito agli israeliani, non sia un «tropo», lasciando maliziosamente intendere che si tratti di un pregiudizio antisemita. Lei gli risponde a tono. Ma che fatica.
“Bevendo champagne al Cairo, i giovani ricchi discettano del conflitto arabo israeliano”
Il mio racconto dal Cairo. Lo leggi gratis qui.
https://t.co/9EjPyHZf2D via @MagElecto#guerra#israele#Palestina#egitto