A lot of data with a small amount of noise can be much more informative than a small amount of "perfect" data.
If 1-5% of your examples are noise, you can still extract meaning, and find patterns.
Also, that's not meaningless gibberish, there is still signal in those examples.
Also, they do note that, "'Semantically related' is purposefully left vague."
Tennis Playpen
Semantically related yes, meaningful no.
Or rather... Could I make a youtube comment classifier and enter it?
Much better would to have been take real sentences from wikipedia.
Here is a uniform, random sample of documents:
faststretch details
honourable judge
oriental museum http
london wc 1n 3 bg pdsg seminar
1857-8 indian mutiny
vera bell
katherine r.b
rare books collection
novy arbat street
universitys case
I took the UKWAC web-as-corpus crawl (2 billion words, crawled in 2008), ran it through the splitta sentence splitter, removed all funny characters, ran the Penn treebank word tokenizer, and perform term extraction with topia.termextract, discarding terms that will single words. I then lowercased the terms, sorted them, and uniqued them, to give the dataset.
From the website, here are some example "documents":
Now, what is the point in trying to generate any kind of "meaning" from those documents if they consist of completely meaningless gibberish?As I was reading this challenge, I immediately thought of spam filtering / youtube comment classification ("smartness" classification) / etc as a potential useful application of this technology.
For example if each "document" is a youtube comment, then in theory you could write an algorithm to examine each comment and output a "smartness guess" for each. Then you (as in, you personally, by hand) would look at the results and specify your own "smartness rating" for a few comments. Then you'd run an algorithm to look at the difference between your specified "smartness rating" and the "smartness guess". Then, using that difference, it would tweak the settings in the original algorithm until it outputs a "smartness guess" that more closely fits your "smartness rating". If you repeat that process enough times, and your original algorithm has enough modifiers to tweak, then you might wind up with an algorithm that can make a pretty good guess about whether any given youtube comment is retarded or not.
And that's just one example of a practical application for this kind of thing.
That said, if the input "documents" are completely and utterly meaningless, then there does not seem to be any point in trying to build "meaning" from those inputs. (Garbage in, garbage out.)