I'm too dumb to understand the paper, but it doesn't feel unlikely that this is a misinterpretation.

What I've figured out:

• They're exclusively looking at text.
• Translations are an important factor. Lots of English content is taken and (badly) machine-translated into other languages to grift ad money.

What I can't quite figure out:

• Do they only look at translated content?
• Is their dataset actually representative of the whole web?

The actual quote from the paper is:

Of the 6.38B sentences in our 2.19B translation tuples, 3.63B (57.1%) are in multi-way parallel (3+ languages) tuples

And "multi-way parallel" means translated into multiple languages:

The more languages a sentence has been translated into (“Multi-way Parallelism”)

But yeah, no idea, what their "translation tuples" actually contain. They seem to do some deduplication of sentences, too. In general, it very much feels like just quoting those 57.1% without any of the context, is just a massive oversimplification.