(Not sure what the problem is, so, not sure what the best category would be for this.)
I am experimenting with a new supertagger for ACE+ERG. What I am getting is: sequences of tags which look accurate to me and then parses which FFTB does not consider similar enough to the gold parses.
I need to understand what is going on and what to do about it, but there are too many pieces to it so I end up confused.
Suppose we have a sentence (from the pest dataset): Not all those who wrote oppose the changes.
The supertagger tags terminals for lexical type. I’ve extracted the gold lexical types using pydelphin (second column), and I supply the tags as listed in the third column:
not av_-_dg-det_le av_-_dg-det_le
all n_np_mc-a_le n_np_mc-a_le
those n_-_pr-dei-pl_le n_-_pr-dei-pl_le
who n_-_pr-rel-who_le n_-_pr-rel-who_le
wrote v_np*_le v_np*_le
oppose v_np_le v_np_le
the d_-_the_le d_-_the_le
changes n_pp_mc-of_le n_pp_mc-of_le
. pt_-_period_le pt_-_period_le
The tag sequences look identical to me? I do get parses, but when I compare the treebanks with FFTB, I see:
How should I approach this? I suppose I should be looking at the profile files themselves and compare things there. Which specific things?