Luheng He on deep SRL: homes.cs.washington.edu/~luheng/files/… #acl2017nlp
previous SOTA is a CRF with neural potentials; here we're just going to train greedily (?) and Viterbi decode at test
NN is not smart enough to learn what well-formed BIO taggings look like, so add hard constraints
8 layers of LSTMs!
SOTA results with this simple model. What mistakes does it make? Do we still need syntax?
mistakes are things that people find hard: argument vs. adjunct, PP attachment
next, add a constraint requiring outputs to be well-formed constituents
this helps a lot with gold syntax, but not much with modern parsers---that last 6 F1 matters!
Share this Scrolly Tale with your friends.
A Scrolly Tale is a new way to read Twitter threads with a more visually immersive experience.
Discover more beautiful Scrolly Tales like this.