Predictions NER_Ontonotes_BERT_Mult for entities with interpunction about deeppavlov

[ronaldvelzen]

The English config doesn't work for the sentence, it gives the following results:

deeppavlov_model = build_model(
        "ner_ontonotes_bert",
        install=True,
        download=True)
sentence = 'Today at 13:10 we had a meeting'
output = deeppavlov_model([sentence])

print(output[0])
[['Today', 'at', '13', ':', '10', 'we', 'had', 'a', 'meeting']]
print(output[1])
[['B-DATE', 'O', 'B-TIME', 'O', 'B-TIME', 'O', 'O', 'O', 'O']]

However, I expect 13:10 to be predicted as B-TIME, I-TIME, I-TIME.

I would also like to use the model for Dutch sentences, therefore I chose to use the multilingual version. In the older version of Deeppavlov these mistakes didn't occur. The problem also occurs for other entities with interpunctions.

Would you happen to have any suggestions on how to solve this problem?

[dmitrijeuseew]

The model ner_ontonotes_bert.json
https://github.com/deeppavlov/DeepPavlov/blob/master/deeppavlov/configs/ner/ner_ontonotes_bert.json
gives correct prediction for your example sentence. I suggest using English config for English sentences.

[ronaldvelzen]

@dmitrijeuseew @IgnatovFedor