Berserker
Berserker (BERt chineSE woRd toKenizER) is a Chinese tokenizer built on top of Google's BERT model.
Installation
pip install basakaUsage
import berserker
berserker.load_model() # An one-off installation
berserker.tokenize('姑姑想過過過兒過過的生活。') # ['姑姑', '想', '過', '過', '過兒', '過過', '的', '生活', '。']Benchmark
The table below shows that Berserker achieved state-of-the-art F1 measure on the SIGHAN 2005 dataset.
The result below is trained with 15 epoches on each dataset with a batch size of 64.
| PKU | CITYU | MSR | AS | |
|---|---|---|---|---|
| Liu et al. (2016) | 96.8 | -- | 97.3 | -- |
| Yang et al. (2017) | 96.3 | 96.9 | 97.5 | 95.7 |
| Zhou et al. (2017) | 96.0 | -- | 97.8 | -- |
| Cai et al. (2017) | 95.8 | 95.6 | 97.1 | -- |
| Chen et al. (2017) | 94.3 | 95.6 | 96.0 | 94.6 |
| Wang and Xu (2017) | 96.5 | -- | 98.0 | -- |
| Ma et al. (2018) | 96.1 | 97.2 | 98.1 | 96.2 |
| -------------------- | ---------- | ---------- | ---------- | ---------- |
| Berserker | 96.6 | 97.1 | 98.4 | 96.5 |
Reference: Ji Ma, Kuzman Ganchev, David Weiss - State-of-the-art Chinese Word Segmentation with Bi-LSTMs
Limitation
Since Berserker is muscular is based on BERT, it has a large model size (~300MB) and run slowly on CPU. Berserker is just a proof of concept on what could be achieved with BERT.
Currently the default model provided is trained with SIGHAN 2005 PKU dataset. We plan to release more pretrained model in the future.
Architecture
Berserker is fine-tuned over TPU with pretrained Chinese BERT model. It is connected with a single dense layer which is applied to all tokens to produce a sequence of [0, 1] output, where 1 denote a split.
Training
We provided the source code for training under the trainer subdirectory. Feel free to contact me if you need any help reproducing the result.
