summaryrefslogtreecommitdiff
path: root/decoder/hg.cc
diff options
context:
space:
mode:
Diffstat (limited to 'decoder/hg.cc')
-rw-r--r--decoder/hg.cc2
1 files changed, 1 insertions, 1 deletions
diff --git a/decoder/hg.cc b/decoder/hg.cc
index e57f7807..70511c07 100644
--- a/decoder/hg.cc
+++ b/decoder/hg.cc
@@ -1,4 +1,4 @@
-//TODO: lazily generate feature vectors for hyperarcs (because some of them will be pruned). this means 1) storing ref to rule for those features 2) providing ff interface for regenerating its feature vector from hyperedge+states and probably 3) still caching feat. vect on hyperedge once it's been generated. ff would normally just contribute its weighted score and result state, not component features.
+//TODO: lazily generate feature vectors for hyperarcs (because some of them will be pruned). this means 1) storing ref to rule for those features 2) providing ff interface for regenerating its feature vector from hyperedge+states and probably 3) still caching feat. vect on hyperedge once it's been generated. ff would normally just contribute its weighted score and result state, not component features. however, the hypergraph drops the state used by ffs after rescoring is done, so recomputation would have to start at the leaves and work bottom up. question: which takes more space, feature id+value, or state?
#include "hg.h"