Ana Macanovic

Join us for a workshop on Immigration, Public Opinion, and Crime in the Digital Age with a focus on combining online and offline data @ISS_UniCologne April 28/29 2025 with two great keynotes by Christian Czymara and Isabelle van der Vegt @Isabellevdv, call is below, pls share.

📣 Applications are open for the EUI PhD programme! 🎢Start your academic journey in September 2025 and embark on the intellectual adventure of a lifetime at the EUI in Florence⚜️ 🔗 loom.ly/Qag4xdc ⏳ Deadline: 31/01/2025 (14:00 CET)

Abstract submissions for this @isa_sociology session on generative AI are open for another week, until the 15th of October. Consider sending in your abstract if you haven’t already! 📄

We are hiring a postdoc (2years+2years) @labss @ISTC-CNR in Rome to work on an experimental project on #social norms and #cooperation in collective action problems under risk. Deadline extended until October 10 to apply: istc.cnr.it/it/content/ass… Contact @giulia_ndr for details

Are you using generative AI to study social processes? 📃 If so, consider submitting an abstract to this session I'm organizing at the at the upcoming 2025 @isa_sociology Forum! Send your abstract in before October 15th here: isaconf.confex.com/isaconf/forum2…

"It's just a quick R plot. How long could it possibly ta--" 💀

At @COMPTEXTCONF, @JohannesBGruber showcases {rollama}, an R package enabling local use of LLMs—fully reproducible and no need to expose data. Best thing is: EASY to set up. I am certainly convinced: if we use LLMs, going local should be encouraged in social science research.

@AnnaLeptikon I took a two-year break before re-entering academia. Applied consistently for two years, if I recall correctly, it took 7 applications. The position I got was a perfect match, but it took a while. Just graduated! I’d say you shouldn’t give up! 💪🏼

Second year in a row with no PhD position offer. Unclear what to learn from this. Either I a.) overestimate myself, b.) what I have to offer isn't legible to the academic bureaucratic system, or c.) randomness and not enough applications sent out (have sent few, partially because I can only get myself to work on things I find truly important and interesting).

Can LLMs replace human subjects in social science? A thread on our latest paper. TL;DR: to date, they cannot really, because they exhibit an unnoticed pattern we call “machine bias”. osf.io/preprints/soca… 1/n

🤯

@bdomackenna Thanks you, Bernardo! Looking forward to hearing your thoughts

Ever tried identifying emotions, values, or hate speech in a large number of texts from social media? Should you use a dictionary?📕 A fine-tuned LLM? 🔨ChatGPT? 💻 We discuss this in our paper just out in Behavior Research Methods: link.springer.com/article/10.375… 🧵A thread: [1/9]

@infornomics @Prashant_Garg_ @fetzert Thank you! I am hopeful that it can help social scientists decide on the best method for their application.

@a_macanovic @Prashant_Garg_ @fetzert Really great paper, thanks for doing the work and writing it down!

@infornomics If you have been working on something along these lines, please do share your work, I'd be curious to read more.

@infornomics This sounds promising for CI/CD routines. King et al proposed something along these lines for keyword selection with supervised ML (doi.org/10.1111/ajps.1…). I am certain enriching this routine with LLMs would be promising, although we haven't really tested anything like this!

Should you remove stop words or lemmatize your texts?🧾How many epochs should you fine-tune your transformer models for? And how much does GPT improve if you provide it with more detailed instructions?💻Read the paper for more results here: rdcu.be/dErm1! [9/9]

Extensively validated dictionaries (e.g., the Moral Foundations Dictionary) work well, but still worse than LLMs.📚 We also build custom dictionaries, and these outperform off-the-shelf ones, primarily as they allow us to include words specific to social media.🗨️[8/9]