Plan.txt

Feb 16 2020
Did dt for previous trop -> next trop. Were able to predict several trop, most it made no prediction:
SOF PASUQ --> END
MAHPAKH --> PASHTA
PASHTA --> MUNAH (PASHTA, ZAQEF QATAN is also common, but slightly <50% of time)
MERKHA --> TIPHA

We got 26% predictions, most trop no prediction at all.
Plan: need to try more complex combinations
1) Try next trop -> previous trop
2) Try several trop -> next or previous trop
How many do we need?

Do we need to move from white box to black box predictor for good predictions (which would mean we can't interpret the results)?

Split into trope system (EMeT, Nach, Chumash) and train models separately.

Look into GMM (Gaussian mixture model, to allow more than one mode: PASHTA --> either MUNAH _or_ ZAQEF QATAN)

Link to Intro: https://drive.google.com/drive/folders/0B1_EiCPi0cabdWNGOGR4enVHTXM