Financial Sentiment Analysis with FinBERT & HuggingFace + Analyzing Model Predictions w/ W&B Tables

[@carlolepelaars]

Great video as always, Ivan! W&B Tables is really becoming something special!

[@IvanGoncharovAI]

Thank you Carlo!

[@kingbach7309]

hello. thanks for the great vid! is it possible to fine-tune this network and add even more labels (other emotions)? thank you.

[@IvanGoncharovAI]

Thank you! Glad you enjoyed the video! I think it's possible, why not? FinBERT was fine-tuned on a general English language BERT model using an annotated dataset where financial folks labelled sentences as positive/negative/neutral with respect to how they' think the sentences would affect the stock of a given company. I imagine you could compile your own dataset and use FinBERT as the base model to fine-tune on, say, you'll have classes, like: confident/worrying/really bad/really optimistic. As long as you get the data you can totally leverage the pretrained FinBERT weights. One caveat: if you might not be able to keep the current classes if you retrain the model on a new dataset, so it's less about adding more classes to the 3 existing ones, and more about training a new neural with new classes off of FinBERT. To add classes to FinBERT you'd probably have to get the dataset that it was trained and run all of that training from scratch with your new classes.

[@kingbach7309]

@@IvanGoncharovAI Again, this is super helpful, Ivan. I appreciate you taking the time and replying to us here. I was wondering.. since my "goal" would actually to predict a sentiment of an individual answer to a certain question, would you suggest me trying to finetune a different network, or would Bert be to go to in this scenario as well? re-training a network from scratch sounds pretty difficult - isn't it? where would i even obtain the dataset & resources to train such a network? :-) I was wondering if you knew any other network that predicts sentiment labels that are more than just "positive" "neutral" and "negative"? Thanks again, Ivan!

[@WeightsBiases]

Hey, Ivan here. I'd say a great place to start looking for a pretrained network that classifies different types of sentiment would be to look at HuggingFace's model hub and the text classification section there. I can't think of the top of my head of one that does something other than positive/negative/neutral, but it's probably worth taking a look there. And let me know if you have other questions :)

[@christopher-george]

​​​@@IvanGoncharovAINice comment ! What are you currently working on ?

[@the_real_cookiez]

Do you think it can do large amounts of cleaned financial papers?

[@WeightsBiases]

W&B Tables can handle 200,000 rows at the moment, but that will increase over time

[@Kongsoon_shorts]

good bro!

[@金融数据]

Got an error message as following. Please take a look. --------------------------------------------------------------------------- NameError Traceback (most recent call last) in () 9 STRIDE = 100 10 ---> 11 model.eval() 12 13 n=0 NameError: name 'model' is not defined

[@sameedzahoor8753]

If you are just starting out with the model then I would reckon to use google colab. you wont get any errors.