[Feature] Named-Entity-Recognition

[do-me]

I recently found span-marker-xlm-roberta-base-multinerd, a multilingual model for Named-Entity-Recognition with 15 interesting labels:

• PER (person)
• ORG (organization)
• LOC (location)
• ANIM (animal)
• BIO (biological)
• CEL (celestial)
• DIS (disease)
• EVE (event)
• FOOD (food)
• INST (instrument)
• MEDIA (media)
• PLANT (plant)
• MYTH (mythological)
• TIME (time)
• VEHI (vehicle)

It's not perfect, but it still works ok.

Other NER-models I know of have fewer categories or less useful ones. The model size is around 1.1Gb and would probably be ~270Mb in the quantized version. Quite heavy for SemanticFinder but still worth a try.

Apart from the model itself, integrating it in the frontend would pose a bit of a challenge. Afaik e.g. displacy-ent (now integrated in spacy) is the state-of-the-art NER visualization frontend library. Not sure whether Gradio uses the same, but it looks pretty similar. However, I don't know whether it would integrate nicely with CodeMirror.

I was thinking of a similar listing feature for all named entities, so if you select e.g. PERSON you would see a list with all persons found in the text and a respective link to jump to the right section. Theoretically we already have all this functionality, just not the nice CSS that displacy-ent has which is secondary anyway.

In this context, I was thinking that we should also allow for a different font. I guess, the majority of users would rather work with normal text, not with code. Hence, a monospaced font is maybe not the best option as default (?).