can history
be written out of tweets?

frédéric clavert
frederic.clavert@uni.lu / inactinique.net

08/03/2024

who am i?

an obvious answer?

yes

no

understanding commemorations online

#ww1. online echoes of the centenary of the great war

analytical frame

information

propagation

vibrations

D. Boullier, « Les sciences sociales face aux traces du big data », Revue française de science politique, vol. 65, nᵒ 5, p. 805‑828, 2015.

collecting tweets

9 million tweets / 1.5 million users

apiv1.1 / hashtags-based / 2014-2019.

visualize, analyse, interpret

access

from dependence to apicalypse

The problem of this research is that it was depdendant on the willingness of Twitter to allow us access to its data. And for a long time, Twitter has been a researcher’s eldorado.

Twitter allowed access via its API – a way, here, to connect to Twitter to get its data in a quite structured form (135 metadata / json). For free: first via its streaming set of commands, and then, for a couple of years, via a specific product for researchers, still free, after a small procedure to get recognized as researcher by twitter (through the new APIv2).

Both kinds of access had their methodological limitations, but it made collecting data so easy.

With deborah paci, for instance, we could collect data on the commemoration of the end of the second world war in italy and france during the lockdowns in 2020. It was easy, fast, the data was well structured, etc.

But that is over. For two reasons:

Elon Musk
Large Language Models

Why LLM? Data is all the more valuable (financially speaking) today that the numerous large language models that are available or being developped incited many platforms to close their access to data (twitter and reddit in 2023 for instance). And they closed it fully: they closed their apis (or made it much more expensive) and limited the ways we had before to scrap data online.

this is the consequence of our dependance, in the age of big data, towards the big web platforms. and this might prevents us to access data and just do our work.

digital dark age

volume

big data

methods

ai

we use too much the acronym “ai”. Here, I’m speaking about machine and deep learning, which is already something wide.

and there are different ways to use that kinds of ai. MALLET for instance, for Latent Dircihlet Allocation, the dominant way to perform topic modelling, is already based on machine learning, with training, etc. But it’s a first step to analyse large dataset.

with more advanced ai-based systems, including some uses of large language models, problem is: infrastructure. Of course we can use those systems for our research, but we need a performant digital infrastructure.

and here, we go from access / volume to inequalities. All universities, all the more when those specialising in humanities and social sciences, are / will not be able to acquire and maintain such infrastructures.