Been trying out some examples of NLP sentiment analysis the last couple of days, ie CNN vs GRU vs bidirectional LSTM. I guess it’s not surprising but complex is not always best… I know it’s only running for 30 epochs but the dataset is the full GLUE/SST2 dataset so not too shabby. Plus not showing much in the way of improvement beyond 20-odd epochs with val_accuracy stuck at 75% odd. CNN is numerically best but some of the sentiment evaluations are just wrong. 2-layer bidirectional LSTM? More subtlety but also not the best. Have to say it’s difficult to tell what to use – seems like the journey into NLP is just getting started.
Was playing this in the background this morning while coding – latest techdub just out end of March from Berlin label Kynant Records with legendary artist Tikiman. Reminds me of May-time vibes back in Berlin ten years ago, DJs playing on massive soundsystems on a Sunday afternoon overlooking the river. Maybe need the same in Tokyo 🙂