There are different decode methods such as dynamic programming and beam search. I used dynamic programming, which is often called Viterbi-decode, to predict the output labels.
Then I used latent variables. I replaced ‘I-XXX’ label by ‘I-XXX’ or ‘E-XXX’. The ‘E-XXX’ means the end of the location, person name, or organization. So there are 10 labels in output set Y.
| Iteration | averaged | Latent var | precision | recall | F1 |
|---|---|---|---|---|---|
| 1 | yes | yes | 85.58 | 75.73 | 80.35 |
| 5 | yes | yes | 89.47 | 85.37 | 87.37 |
| 10 | yes | yes | 90.44 | 86.78 | 88.57 |
| 20 | yes | yes | 90.86 | 87.14 | 88.96 |
Zhang X, Du Y, Sun W, et al. Transition-based parsing for deep dependency structures[J]. Computational Linguistics, 2016.
I used two-stack transition system. All details are in the paper.