COMPLEXITY OF N-GRAM SEARCH

• Although it is always desirable to use as many knowledge sources as possible, there are practical problems integrating such information into a time-synchronous search.
  
  
• One alternate strategy is to use a mult-pass search. However, the more accurate the first-pass, the better the performance on subsequent passes.
  
  
• One of the most critical parts of search is the tree lexicon.
  
  
• In a linear lexicon, each word is represented as a linear sequence of phonemes independent of other words. For example, though task and tasks share the same root, we do not share any of their history during the search process.
  
  
• Large lexicons introduce enormous complexity for a backoff N-gram language model because we must "start" all words in the lexicon for every unique history in our current search space.
  
  
• Can we share some of the underlying phonetic structure of words in the lexicon?