This model is reported in the paper: Nguyen Truong Son, Nguyen Le Minh, "Nested named entity recognition using multilayer recurrent neural networks", PACLING 2017, August 16 - 18, 2017, Sedona Hotel, Yangon, Myanmar (to be appear)
Requirements:
- Python 2.7, with Numpy and Theano installed.
Two implemented models:
-
lstm-tagger-v4: Implementation of single BI-LSTM-CRF with additional features to recognize named entites at the top level.
-
multi-lstm: Implementation of Multilayer BI-LSTM-CRF model to recognize nested named entities.
Our proposed models are based on Lample et al 2016.
- VLSP 2016: http://vlsp.org.vn/evaluation_campaign_NER
- Results on the official test set
| Model | POS | CHUNK | Pre-trained | F1 % | |
|---|---|---|---|---|---|
| 1 | 82.9 | baseline1 (Lample et. al) | |||
| 2 | X | 86.44 | +3.54% | ||
| 3 | X | 89.77 | +6.87% | ||
| 4 | X | X | 90.27 | +7.37% | |
| 5 | X | 86.84 | baseline2 (Lample et. al) | ||
| 6 | X | X | 88.66 | +1.82% | |
| 7 | X | X | 91.79 | +4.95% | |
| 8 | X | X | X | 92.97 | +6.13% |