Shift in text statistics?
#1
by
schaefer
- opened
Hey, I'm wondering whether any post-hoc detector could be trained to detect texts generated/paraphrased by DIPPER. I see that DIPPER has parameters for adjusting the lexical diversity and the amount of content re-ordering in the paraphrase, which I assume mitigates the effectiveness of such a detector.
Is there any published research on this? I'd be interested in seeing the detection performance of e.g. Ghostbuster by Verma et al. (https://github.com/vivek3141/ghostbuster) trained on text generated by DIPPER.
Best regards,
Jonas