Semantic distances between comments in online newspaper forums increase over time
Max Pellert (https://mpellert.at)
Barcelona Supercomputing Center
DerStandard
DerStandard is a major Austrian newspaper, online user can post below news articles
We created a dataset from the platform that spans ten years (2013–2022)
Includes over 75 million user comments, more than 400 million votes and detailed metadata, also on articles and user interactions
All in a nicely organised data repository: https://dataverse.bsc.es/dataset.xhtml?persistentId=perma:BSC/P32CXW
(Dis-)Similarity of postings over time
We embed the text of each posting with “KaLM-embedding-multilingual-mini-v1”
Next, we compute cosine similarity of all pairs of postings below each newspaper article
We average by month of each year
To check for divergence of contents, we compute similarity between short text summaries of each pair of article in each year
A proxy for polarization?
Interesting phenomenon that could be an artifact
Like singular topics that overlay everything for some time at least (COVID-19 or Russo-Ukrainian War)
But dominating news topics hardly a new phenomenon and that should also affect article summary similarity?
Otherwise, it could be a revealing trend
Some sign of polarization?
If yes, the trend should differ between article contents –> we can make use of the topical tags on articles provided by the platform
| Topic Tag (G) | Topic Tag (E) | Average increase of distance (in 10^−4) |
|---|---|---|
| diestandard | genderissues-political | 2.934 |
| diskurs | discourse | 1.716 |
| etat | media | 1.134 |
| gesundheit | health | 3.093 |
| immobilien | real estate | 2.496 |
| inland | national (domestic) | 1.822 |
| international | international | 1.342 |
| karriere | career | 2.424 |
| kultur | culture | 1.980 |
| lifestyle | lifestyle | 0.955 |
| panorama | reports | 1.157 |
| sport | sport | 1.982 |
| web | web | 1.632 |
| wirtschaft | economy | 2.072 |
| wissenschaft | science | 1.804 |
Further analyses
I speculate that within-group similarity may have become higher over time, leading to higher dissimilarity between groups and an increasing overall mean
To compare groups, we have (ideological) user factions based on signed networks
Also checking for Simpon’s paradox (unlikely)
Definitely, better metrics needed
What are the mechanisms behind that?
