EXPLOITING SUBTREE POLYMORPHISM

• The previous techniques we have discussed only eliminate identical subtrees.
  
  
• There are many subtrees that share the same nodes and topology, but have different LM scores. Can we avoid redundant acoustic model state evaluations for such trees?
  
  
• A subtree is dominated when the best outcome in that subtree is not better than the worst outcome in another instance of that tree. State evaluation is redundant and unnecessary for the dominated tree.
  
  
• A polymorphic linguistic context assignment to reduce redundancy is employed.
  
  
• Polymorphic context assignment involves keeping a single copy of the tree and allowing each state to assume the linguistic context of the most promising history (in essence, a pruning technique).
  
  
• Each node in the tree is evaluated once. However, this approach can introduce search errors.
  
  
• One approach (WHISPER) to mitigate the effects of this pruning is to keep a heap of the most promising contexts, and to delay the decision regarding which context is most promising.
  
  
• For trigram decoding, many of the approaches we have just described are still not practical.