For viterbi backwards I kept getting the wrong pred output for the last 6 values (all of them were #).
At first I was confused as to why this was happening. Then I found that several of the last columns in the best_probs matrix were all zeros, even though the matrix passed the viterbi_forward unit test. Because all of these columns were zero, the “highest probability” POS tag was chosen as the 0th index, which correlated with “#”.
Anyways, this is more of a question as to what went wrong with the best_probs matrix. I’m confused as to why several of the last columns are filled with zeros even though the unit test was passed.