diff options
Diffstat (limited to 'decoder/ff_lm.cc')
| -rw-r--r-- | decoder/ff_lm.cc | 4 |
1 files changed, 3 insertions, 1 deletions
diff --git a/decoder/ff_lm.cc b/decoder/ff_lm.cc index 03dc2054..e6f7912e 100644 --- a/decoder/ff_lm.cc +++ b/decoder/ff_lm.cc @@ -1,3 +1,5 @@ +//TODO: backoff wordclasses for named entity xltns, esp. numbers. e.g. digits -> @. idealy rule features would specify replacement lm tokens/classes + //TODO: extra int in state to hold "GAP" token is not needed. if there are less than (N-1) words, then null terminate the e.g. left words. however, this would mean treating gapless items differently. not worth the potential bugs right now. //TODO: allow features to reorder by heuristic*weight the rules' terminal phrases (or of hyperedges'). if first pass has pruning, then compute over whole ruleset as part of heuristic @@ -311,7 +313,7 @@ class LanguageModelImpl { double sum=0; for (;rend>rbegin;--rend) { sum+=clamp(WordProb(rend[-1],rend)); - UNIDBG(","<<TD::Convert(rend[-1])); + UNIDBG(" "<<TD::Convert(rend[-1])); } UNIDBG(")="<<sum<<endl); return sum; |