Unofficial LemmaTag [1] implementation in TensorFlow 2
Each experiment can be run with the following command:
python train.py --scheduler <scheduler> --we-dim <we-dim> --encoder-layers <encoder-layers> --name <project>_<name>
For example the original results from [1] can be reproduced using:
python train.py --scheduler lemmatag --we-dim 768 --encoder-layers 2 --name tagger_lemmatag
The results including the pre-trained models can be found here: https://app.wandb.ai/kulhanek/lemmatag?workspace=user-kulhanek
| configuration | lemmatizer dev accuracy | lemmatizer test accuracy | tagger dev accuracy | tagger test accuracy |
|---|---|---|---|---|
| baseline | 99.02 | 98.75 | 97.00 | 96.67 |
| original lemmatag [1] | - | 98.37 | 96.90 |
In this experiment the tagger was trained without the lemmatizer. For tagger alone, the training results including the pre-trained models can be found on wandb: https://app.wandb.ai/kulhanek/tagger?workspace=user-kulhanek
| configuration | scheduler | encoder-layers | we-dim | dev accuracy | test accuracy |
|---|---|---|---|---|---|
| tagger-baseline | lemmatag | 2 | 768 | 97.03 | 96.79 |
| tagger-cosine | cosine | 2 | 768 | 97.05 | 96.74 |
| tagger-4-layers | cosine | 4 | 542 | 97.09 | 96.73 |
| original lemmatag [1] | lemmatag | 2 | 768 | - | 96.83 |
[1] LemmaTag: Jointly Tagging and Lemmatizing for Morphologically-Rich Languages with BRNNs, Daniel Kondratyuk, Tomáš Gavenčiak, Milan Straka, Jan Hajič
https://arxiv.org/abs/1808.03703