New preprint! Metaphors shape how people understand politics, but measuring them (& their real-world effects) is hard. We develop a new method to measure metaphor & use it to study dehumanizing metaphor in 400K immigration tweets Link: bit.ly/4i3PGm3 #NLP #NLProc #polisky #polcom #compsocialsci 🐦🐦 - ThreadSky

jmendelsohn2.bsky.social • 4 days ago

New preprint!
Metaphors shape how people understand politics, but measuring them (& their real-world effects) is hard.

We develop a new method to measure metaphor & use it to study dehumanizing metaphor in 400K immigration tweets Link: http://bit.ly/4i3PGm3

#NLP #NLProc #polisky #polcom #compsocialsci
🐦🐦

Comments

akhilarora.bsky.social•4 days ago

How are you still getting Twitter data? 🤨

anchor-model.bsky.social•3 days ago

their API still allows reading tweets (but it's now a billion dollars)

akhilarora.bsky.social•4 days ago

Super cool!

propjerry.bsky.social•4 days ago

"Where one sees these viruses, one can expect unchanging attitudes in support of conspiracies, of cults, of populists and of bigots. Should not be surprising.

"In these ... these viruses, in forms of reasoning, one can see the DNA of today’s politics."

https://www.pinoytoolbox.org/post/dna-of-today-s-politics

jmendelsohn2.bsky.social•4 days ago

Broadly, metaphor means talking about one concept in terms of another. Our approach is grounded in 2 ideas from linguistics and polcom: (1) Metaphor can be cued by both specific expressions and discursive context, and (2) Their political implications depend on the specific concepts evoked.

jmendelsohn2.bsky.social•4 days ago

We first identify seven dehumanizing "source domain" concepts in immigration discourse (e.g. "water" or "vermin"). Sometimes, dehumanizing metaphors are overtly hateful, but are often accepted as natural ways of talking about immigrants (e.g. "waves of immigration").

jmendelsohn2.bsky.social•4 days ago

We then propose and evaluate a new approach that combines word-level and document-level measures of metaphorical language, using LLMs and document embeddings (SBERT) respectively. While we focus on metaphorical dehumanization of immigrants, this method could be applied to many other contexts!

jmendelsohn2.bsky.social•4 days ago

We create a new dataset of 1,600 tweets, each annotated by ~8 people for judgements of metaphor with respect to a given source domain concept to evaluate our approach. We evaluate 3 LLMs, 2 prompt variations, and show that incorporating SBERT similarity consistently improves performance.

jmendelsohn2.bsky.social•4 days ago

Finally, we turn to the most interesting part: our analysis of the relationships between metaphor, authors' political ideology, and user engagement! This comes from 400K tweets from a dataset I compiled a few years ago (alas, not the most recent. thanks elmo). We fit lots of regressions.

jmendelsohn2.bsky.social•4 days ago

Unsurprisingly, we find that conservative ideology is associated with higher metaphor use, but the extent varies across concepts.
Interestingly, authors with stronger left-wing ideologies are actually more likely to use creature-related metaphor than moderate liberals (animal, vermin, parasite).

graubart.bsky.social•4 days ago

Fascinating; I may add a link to this in the post I just wrote about intentional manipulation of language in LLMs

https://open.substack.com/pub/unpromptedthoughts/p/rewriting-reality-the-hidden-influence?r=2bsm&utm_campaign=post&utm_medium=web&showWelcomeOnShare=false

Comments

Posting Rules

Reply