First page Back Continue Last page Graphics
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
Model learns in an unsupervised manner on 2033 document/extract pairs
- 13k/41k sentences 261k/1m words
- 6.3/21.5 sents/doc 128/511 words/doc