Thread by @emilymbender

I've seen several different #NLProc folks suggesting today that it would fun/interesting/worthwhile to use BERT or GPT-2 to fill in the redacted bits of the Mueller report. A short thread on why this is a terrible idea /1

First: consider the importance of the ability to find news sources that you trust and how much interest there is in the document. If you put out a version of that document with invented text in place of the redactions, how long before someone reposts it as the real thing? /2

How does that affect the discourse around what's actually contained in the (unredacted) version of the document, what it means, etc. both immediately and at some future point when the actual thing is available in full? How does it affect people's trust in reliable news? /3

Second, examine why you think that BERT or GPT-2 generated answers would be interesting at all. Do you think that a big language model somehow can guess what the truth is and reveal it to you based on the rest of the document? /4

If so, you are wrong. Those are language models. They can only come up with sequences that are probable based on what's seen in the training data, given the prefix fed in. /5

In other words, they can tell you about what's in the training data, not what's in the report. /6

I haven't looked at the report, but I'm fairly confident the people doing the redacting would have been careful to do it in such a way that the redacted info furthermore is not predictable.

(And, ahem, that the black-out can't just be deleted...) /7

So, please, just stop it with this idea. It's not funny nor helpful. If you're interested in applying #NLProc in ways relevant to the current political moment, how about working on e.g. rumor detection and tools that might help users think twice before retweeting/sharing? /fin