jonathanbgn

Great summary! I also recently wrote a post digging into the internals of wav2vec with illustrations:

The Illustrated Wav2vec - https://jonathanbgn.com/2021/06/29/illustrated-wav2vec.html

Thanks! I used a combination of datasets that I found from Google's Dataset Search tool (https://datasetsearch.research.google.com).

For inference speed I recommend a Naive Bayes model. I've tried this on Twitter messages and got near ~90% accuracy with 3-class (positive, negative, neutral).

The easiest library to do that would probably be scikit-learn with their ComplementNB class: https://scikit-learn.org/stable/modules/generated/sklearn.na...

For the data you can use the SemEval 2017 Task4-A dataset (around ~10K labeled tweets): https://github.com/cbaziotis/datastories-semeval2017-task4/t...

Thanks for your message, like you said it's a real challenge when the domain is quite broad (such as social media). I'm currently based in Taipei.

The algorithm will try to give more importance to words which appear rarely and are only used with the chosen brandname (similar to TF-IDF). This is why sometimes weird words can surface to the wordcloud, especially when the sample size of messages is small.

To prevent those words from appearing, I was thinking to implement some dictionary-check to only allow for meaningful words. However this approach also have drawback as you restrict people's words and can miss important new concepts.

Thanks for the feedback.