We can make the annotation experience fast and efficient. We unfortunately don't do office furniture

Everyone in #NLP talks about attention because of the amazing performance of #transformers. But are there other models that could be competitive ? A while back the team at Facebook AI published a paper on "dynamic convolutions" which run faster. buff.ly/3s7YpKH

How can you choose the right named entity recognition model for your use case ? With so many options in the #OpenSource #NLP world, it can be overwhelming to figure out which one is right for you. Read our article showing how you LightTag can help buff.ly/3rYqMLk

#ActiveLearning uses AI to train #AI. Instead of labeling all of our data, we ask our AI what is the highest value data to label, cutting costs and saving time. But Active Learning has some pitfalls that you should know before using it. buff.ly/3bejGfU

@Huggingface's transformers library makes modern NLP easy! But the special tokenization makes using it for #NER tricky. Read about best practices for aligning span annotations to transformers input format. buff.ly/3pUA6Oo

#Medical #NLP uses AI to understand medical documents. But collecting medical training data is particularly expensive. Read how Zebra Medical vision overcame cut their annotation costs with clever tricks and #NLP buff.ly/2yamjdz

#Labeling #Data for #NLP is a great start, but we need to do it efficiently. What are the best practices for making your data #annotation project cost effective and fast ? Read our article to find out. buff.ly/3hRYXzC buff.ly/3hRYXzC

After all that text (pun intended) there's even a small repo with some convenient code that encapsulates everything we learned. It will align annotations, implement a BIOLU scheme, pad, batch and window your data in just a few lines github.com/LightTag/sequence…

Just like "old school" NER had BIO tags per token, it makes sense to do the same for subword tokens. As a bonus, this let's us take advantage of the inductive bias of a CRF.

If you have annotations that look like this, getting them aligned with tokenizer's output is annoying. The latest versions of the tokenizers library have lesser known functions (char_to_token, token_to_chars) that make that much easier