N-BEST DECODING IS USEFUL FOR INTEGRATING KNOWLEDGE SOURCES

• Often the search space becomes unmanageable for real problems due to complex language model constraints (e.g., trigrams) and acoustic models (e.g., cross-word context-dependent phones), and our interest in integrating multiple knowledge sources (e.g., phrase structured grammars).
  
  
• A pragmatic alternative is to use a multipass search approach:
  
  
  
• It is common to perform a first pass of decoding with a bigram language model (or a word-internal triphone/trigram system), and postprocess (or rescore) the output with a more sophisticated system. We refer to this process as multipass decoding.
  
  
• Stack decoding is naturally suited to generating N-best lists. Consider this example from the North American Business (NAB) Corpus:
  
  

  1.	I will tell you would I think in my office
  2.	I will tell you what I think in my office
  3.	I will tell you when I think in my office
  4.	I would sell you would I think in my office
  5.	I would sell you what I think in my office
  6.	I would sell you when I think in my office
  7.	I will tell you that I think in my office
  8.	I will tell you why I think in my office
  9.	I will tell you would I think on my office
  10.	I Wilson you think on my office

  
  
• N-best lists are very compact representations of the search space since timing information is discarded.
  
  
• One popular rescoring experiment that can be performed with N-best lists is referred to as an oracle experiment: How often does the correct hypothesis appear, and at what depth in the list does it appear?
  
  
  
• Oracle experiments are one form of a cheating experiment. Cheating experiments are very important diagnostic tools.