Confusion matrix in Quality Estimation

[bittlingmayer]

In Outbound Translation User Interface Ptakopět: A Pilot Study, it mentions false positives.

Is a "positive" a good translation or a bad translation (like testing positive for a disease in a medical context)?

@zouharvi I assume your setup is in harmony with that in Lucia Specia's book on QE.

[zouharvi]

Negatives (in terms of QE classification) are segments/words/tokens classified as incorrect.

• True negative: segment is incorrect and is reported incorrect
• True positive: segment is correct and is reported as correct
• False negative: segment is correct but is reported as incorrect
• False positive: segment is incorrect but is reported as correct

[zouharvi]

An example: The kids fell asleep .

Reference	Translated	QE	Classification
Die	Die	OK	TP
Kinder	Kinder	BAD	FN
sind	ist	OK	FT
eingeschlafen	eingeschlufen	BAD	TN
.	.	OK	TP

[bittlingmayer]

Thanks!

At ModelFront, we've been doing the opposite!

[zouharvi]

@bittlingmayer Can this issue be closed?

[cefoo]

I think this is better fitted to be stored under Discussions instead of closing an issue?

I may be wrong, I'm still new to GitHub! 🙃

[bittlingmayer]

Sure!