MEMORY ORGANIZATION AND EFFICIENCY

• A major drawback to the use of successor trees is the large memory overhead required.
  
  
• For example, the 1994 NAB LM contains 5M bigrams and over 70M bytes to store predecessor-dependent lexical trees.
  
  
• Need efficient ways to handle the multiple copies of lexical trees.
  
  
• Factorization of the LM scores pushes the application of probabilities earlier in the tree (before we leave the leaves and transition to the next word). Hence, we open the possibility to merge duplicated trees.
  
  
• These trees can be merged to avoid redundant state evaluation, thereby saving space and computation, with no loss of optimality.