Sequence Labeling Issue

[qhd1996]

How to label words that bert vocab does not contain?
This situation happens in English corpus. For example, the word 'jony' is tokenizered to 'jon', '##y'. If in the origin corpus 'jony' is labeled with B-PER, how to modify the corresponding sequence labels in tokens produced by BERT in your code?

[weizhepei]

In my understanding, there is no need to modify the corresponding sequence labels.
For example, given a sequence ['Jony', 'is', 'a', 'boy'] with label being ['B-PER', 'O', 'O', 'O'].
After tokenization, you will get ['Jon', '##y', 'is', 'a', 'boy'].

According to the BERT paper:

we use the representation of the first sub-token as the input to the token-level classifier over the NER label set.

In this setting, you are using ['Jon', 'is', 'a', 'boy'] as the representative of the original sequence ['Jony', 'is', 'a', 'boy']. Since the sequence length is unchanged, you don't need to change the original sequence labels either. Hope this discussion would be helpful to you.