Experimental Indonesian language model for Spacy

Indonesian language model trained on UD_Indonesian_GSD v2.2 (2018-03-01) available here:

https://github.com/UniversalDependencies/UD_Indonesian-GSD

Lemmas, XPOS and morphological features comes from MorphInd (created by Septina Dian Larasati, run and converted by Dan Zeman, http://septinalarasati.com/morphind/).

Command used to train the model:

python -m spacy train id models id_gsd-ud-train.json id_gsd-ud-dev.json

Note that this is not an official spaCy model yet. More improvements are needed before the model is production-ready, hence countributions are welcomed. Open an issue or make a pull request to give feedback on POS / dependency parser predictions or discuss enhancements like hyperparameter tuning. The next thing this model needs is Named Entity Recognition, so it would be great to get a discussion going on where to find good data and what previous state-of-the-art results are for Indonesian. Get in touch!

Installation

To install the package use this command:
pip install https://github.com/jeannefukumaru/id_dep_ud_sm/tree/master/id_dep_ud_sm/id_dep_ud_sm-0.0.1/dist/id_dep_ud_sm-0.0.1.tar.gz

Troubleshooting:

1. If pip install doesn't work, manually download id_dep_ud_sm-0.0.1.tar.gz then at the command line:

   tar -xzvf id_dep_ud_sm-0.0.1.tar.gz
   cd id_dep_ud_sm-0.0.1.tar.gz
   python setup.py install

Usage

import spacy
id = spacy.load("id_dep_ud_sm")

doc = id("Bantulah kami melengkapi berita dan informasi")

for token in doc:
    print(token.text, token.lemma_, token.pos_, token.tag_, token.dep_,
            token.shape_, token.is_alpha, token.is_stop)

Training results:

• dropout_from = 0.2 by default
• dropout_to = 0.2 by default
• dropout_decay = 0.0 by default
• batch_from = 1 by default
• batch_to = 16 by default
• batch_compound = 1.001 by default
• max_doc_len = 5000 by default
• beam_width = 1 by default
• beam_density = 0.0 by default
  Warning: Unnamed vectors -- this won't allow multiple vectors models to be loaded. (Shape: (0, 0))
• learn_rate = 0.001 by default
• optimizer_B1 = 0.9 by default
• optimizer_B2 = 0.999 by default
• optimizer_eps = 1e-08 by default
• L2_penalty = 1e-06 by default
• grad_norm_clip = 1.0 by default
• embed_size = 7000 by default
• token_vector_width = 128 by default
• parser_hidden_depth = 1 by default
• parser_maxout_pieces = 2 by default
• token_vector_width = 128 by default
• hidden_width = 200 by default
• embed_size = 7000 by default
• history_feats = 0 by default
• history_width = 0 by default

Training results

the explanation of each metric can be found from spaCy's official docs: https://spacy.io/api/cli#train

Name	Description
Dep Loss	Training loss for dependency parser. Should decrease, but usually not to 0
NER Loss	Training loss for named entity recognizer. Should decrease, but not to 0
UAS	Unlabeled attachment score for parser. The percentage of unlabeled correct arcs. Should increase
NER P.	NER precision on development data. Should increase.
NER R.	NER recall on development data. Should increase.
NER F	NER F-score on development data. Should increase.
Tag %	Fine-grained part-of-speech tag accuracy on development data. Should increase.
Token %	Tokenization accuracy on development data.
CPU WPS	Prediction speed on CPU in words per second, if available. Should stay stable.
GPU WPS	Prediction speed on GPU in words per second, if available. Should stay stable.

Evaluation results

Time               1.67 s         
Words              11780          
Words/s            7040           
TOK                100.00         
POS                89.51          
UAS                82.33          
LAS                73.74          
NER P              0.00           
NER R              0.00           
NER F              0.00