Conversation
|
Prior to this PR, NER inferencing produced a single unnested list of Named Entities as predictions. After this PR, it is not a list of lists where each inner list corresponds to an input sample (i.e. if there are 3 input samples at inference time, the predictions will be a list populated by three lists). It seems odd to me that this was the case. @Timoeller Do you think this change might break anything that is downstream of FARM? |
|
Mhh, I do not see where we use FARM NER in downstream applications and also "NER inferencing produced a single unnested list of Named Entities as predictions" seems not useful in the first place. So lets go with your newly implemented approach. |
Timoeller
left a comment
Timoeller
left a comment
There was a problem hiding this comment.
LGTM.
How about adding a test case that checks if labels and probs are in line now (I think it is good practice to always add test cases when we bugfix).
2 participants
Previously, the NER probabilities being returned were often probabilities for the wrong entities. This PR addresses this issue (#658 )