-
March 6 2018
- Edit distance is too slow
- Tfidf+cosine achieves 80% perfect predictions on the replacement benchmark with 4454 tasks)!
-
March 23 2018
- naive split is better than languageg tokenizer
- WMD distance works well on icse15-1LineRep
- WMD-distance combined wit hedit distance in a pseudo geometric mean works even better on icse15-1LineRep
- Decision: move from cumulative loss to average loss (resulting in loss in [0,1] thanks to tanh)
- Decision: when an algorithm fails to output something, we take the worst case loss, which is one