https://kotlinlang.org logo
Title
р

Ролан

03/09/2021, 3:38 PM
@Maria Khalusova why not using transformers for that NLP task?
m

Maria Khalusova

03/09/2021, 3:41 PM
HMMs are simpler, fast and you can apply them to a ton of other things. They are more universal. POS is just one of the multitude of applications. Transformers will give you better results, sure, and I plan to write about them at some point later, but this time I was in the mood to write about a probabilistic approach. As simple as that.
👍 2
р

Ролан

03/09/2021, 3:47 PM
Fair enough
h

holgerbrandl

03/10/2021, 9:57 AM
And HMMs contain a lot less parameters and thus can be trained faster with less data.
р

Ролан

03/10/2021, 10:50 AM
You don't need to retrain the whole model. Just use those already packaged by huggingface or Google
Then you just fine tune.
On GPU they are pretty fast enough to run. Even spacy starts to move away from probabilistic models
Delivering those to Kotlin is very much on the plan here
h

holgerbrandl

03/10/2021, 10:59 AM
There are many sparse training data situations in time-series analytics (beyond NLP) where no base model for transfer learning is available, and there HMMs are a great tool. And when in doubt it's always worth starting with a simpler model unless metrics prove it wrong.