Another bonus: these ratings can be generated "in-context", because transformer word embeddings change depending on surrounding semantic information. (Think "I go to the bank to get money" vs "I walk on the sandbank" -> previous methods were not able to differentiate btw these)
The result surpasses SOTA and even the reliability of human raters (at least in English). Interestingly, we were able to extend ratings to other languages with a simple translation step (suggesting that there may be something universal about these ratings).
Why a multimodal transformer? We know that visual information is important for accurately rating concrete words. And because emotional information is important for accurately rating abstract words, we fine-tuned our model on a dataset of emotional image descriptions.
Our multimodal transformer tool for automating word-concreteness ratings is published today
📝 C-ratings are used in research across Cognitive Science 💶 They take time and money to collect ⚙️ Automation solves this + we get in-context ratings for free!
Sam Gershman writes beautifully about how theory-free neuroscience prevents the field from reaching its promise. Beautiful and true. Most folks do not test hypotheses. Running a NHST does not a hypothesis make. www.thetransmitter.org/theoretical-...
Most critiques of LLMs (e.g. @garymarcus.bsky.social) are ultimately about not wanting to accept that very dumb processes can lead to "intelligence". I think this paper offers a great perspective on how this could happen in both minds and machines. www.sciencedirect.com/science/arti...
I wish a smart mind like yours would not focus on the negative side for likes, but instead try to think about what we can do with the technology we have.
what are those other approaches and what impact have they had? I 100% appreciate the skepticism but it just works. No matter how many failures you quote it’s absolutely incredible we have a model as capable as today’s.
The neurobiology of language does not operate in a vat. An important perspective from the @thelablab.bsky.social and colleagues: "Language is widely distributed throughout the brain" www.nature.com/articles/s41...
@viktorkewenig.bsky.social shows that, while concepts generally encode habitual experiences, the underlying neurobiological organisation is not fixed but depends dynamically on available contextual information. 👏🍾🐐
