Something you can join on Tuesday:
This presentation will outline new research on the lived experience of creator-practitioners of open educational resources (OER) and open educational practices (OEP) in United States (U.S.) higher education institutions.
Despite being an article on Scholarly Kitchen this is actually quite good and provides a solid overview of the challenges academic libraries face as the filling in the scholarly publishing sandwich. Worth a read, especially if you don't understand why libraries won't just stop paying such extortionate fees to data miners uh I mean “scholarly publishers”.
I admit I don't understand a lot of this paper but the gist seems to be that it's possible to do some basic processing with
gzip and achieve pretty similar results to NLP tools that usually require enormous amounts of compute power:
Our method achieves an accuracy comparable to non-pretrained neural network classifiers on in-distribution datasets and outperforms both pretrained and non-retrained models on out-of distribution datasets. We also find that our method has greater advantages under few-shot settings.
Essentially, it's a smart hack to massively reduce the resources required to do large language model processing, especially with data sets that aren't well labelled (i.e. untrained). This has a lot of potential for smaller and niche pattern-matching tasks – which describes a lot of the potential applications of NLP/“AI” tools in GLAM.
Libraries and Learning Links of the Week is published every Monday by Hugh Rundle.
Subscribe by following
@email@example.com on the fediverse or sign up for email below.