Training the System

• Despite existence of efficient estimation algorithm, still too slow

  • 1500 word document yields ~ 10k states
  • Use extract: down to 511 words ~ 4k states
  • Beam search: explore only most likely 50% of state space