Transformers can write text, but they can also estimate scores / do regression for any text you give them. Here's a quick demo I made (RoBERTa fine-tuned on Yelp review data)

Jan 9, 2023 · 10:43 PM UTC

18
34
1
468
0
Replying to @alyssamvance
how did you get the distributions?
1
2
Trained it to output mean and s.d. instead of just mean, then used log(pdf(label)) as the loss function
3
17
more replies
Replying to @alyssamvance
is lower or higher score better
1
This is a Yelp review dataset, it's from 1 to 5 stars
1
2
more replies
Replying to @alyssamvance
love the interactivity!
2
5
Thanks! You don't even need AI for that part, though, just WebSockets ☺️
4
Replying to @alyssamvance @JgaltTweets
@Frances_wxh
1
Replying to @alyssamvance
Amazing how sensitive it is to punctuation and capital letters!
Replying to @alyssamvance
One thing you can build with this is by prefixing or not prefixing the text with a hidden prompt, e.g: "The following is written in excellent, idiomatic English with perfect spelling and grammar:" Then you can highlight tokens that are relatively less likely under the prompt.
1
2
Replying to @alyssamvance
generalized sentiment scorer for any text as browser plugin within reach. Grammarly was supposedly worth 13 billion, so, yk
2
Load more