This is a fine-tuned version of COVID-Twitter-BERT, which itself is a fine-tuned version of BERT-large-uncased.
It is designed with business-/brand-analysis purposes in mind and is meant to be used on short documents (maximum token length of 128).
It was built for Named Entity Recognition of the following entity types:
"corporation", "event", "location", "person" and "product".
Tag-label-dictionary:
0: "B-corporation",
1: "B-event",
2: "B-location",
3: "B-person",
4: "B-product",
5: "I-corporation",
6: "I-event",
7: "I-location",
8: "I-person",
9: "I-product",
10: "O"
The following datasets were used for fine-tuning:
TweetNER7, WNUT 2016, WNUT 2017 and a self-created dataset of synthetic, brand-related tweets.
