Natural Language

PSU CS441/541 Lecture 9
November 20, 2000

• Plan For Today
  • Natural Language
  • Scheduling (deferred)
  • Project discussion
  • HW3 discussion (eliminated)
• Natural Language
  • Why AI?
    • obvious: language implies intelligence
      • Turing test
      • Washoe
      • SHRDLU
      • ELIZA
    • language recognition as puzzle domain requiring reasoning
  • Two facets: recognition and synthesis
    • state of art in synthesis
      • acceptable phoneme-based word pronunciation
      • usable
        • spoken NL text using tricks
        • text generation via standard CS methods
      • sonics generally ahead of NL generation
      • NL generation generally ahead of general AI
    • state of art in recognition
      • much harder than synthesis
        • the crummy->good vs. good->crummy problem
        • requires much more integration with semantics
      • many axes on which to compare
        • for a good time, try 1-800-555-TELL
        • DragonDictate
      • big sponsor: the phone companies
  • levels of recognition
    • speech sounds
      • digitize
      • FFT (voiceprint = time/freq plot)
    • phonemes, phoneme groups
      • fit to vocal tract model
      • deal with transitions
    • words ("segments")
      • most words are not pronounced distinctly ("continuous speech")
        • force speaker to segment speech (e.g. menu systems)
        • try to find breaks bottom up (hard/impossible)
        • work back to breaks from syntax (HMMs, works better)
    • syntax ("grammar")
      • book is very explicit here
        • obtain word meanings from dictionary
        • walk nondeterministic ATN grammar to generate parse
        • grammar constrained by special rules
        • where do dictionaries and magic grammars come from?
    • semantics ("meaning")
      • goal: run special rules over parse to generate FOL
      • make sure to match conjunction of all triggered rules
    • pragmatics ("higher meaning")
      • narrow choices as far as possible using background knowledge
      • e.g.: pronoun referent problem (anaphora resolution)
      • discourse analysis (e.g. scene formation)
      • plan recognition
        • identify agent goals
        • find agent plan which would fit agent actions
        • give broader semantic analysis (e.g. need money, get gun)
        • also useful elsewhere
• Project + HW 3 discussion
• review of negamax
• from board value to move value
• general tricks
  • transposition tables
    • hash to position
    • open hash table
    • random replacement
    • with iterative deepening
    • graph-history interaction
  • alpha-beta with t-tables
  • incremental do-undo!
  • side-separated heuristic, differencing
• the rules of LOA
• LOA-specific problems
  • connectedness
  • move generation
  • heuristic evaluation
    • material (different for high/low)
    • connectedness (number/size of groups)
    • mobility
    • coordination
    • piece position