Skip to content

Latest commit

 

History

History
102 lines (77 loc) · 2.95 KB

TRAIN_README.md

File metadata and controls

102 lines (77 loc) · 2.95 KB
Raw

Training Twitter-Stanza

NER

You can specify two data setttings:

  • en_tweetwnut17: trained on TB2+WNUT17
  • en_tweet: trained on TB2
# prepare NER data
cd ./twitter-stanza/data/ner
python prepare_ner_data.py
cd ../..

# train the NER model
shorthand=en_tweetwnut17
python stanza/utils/training/run_ner.py ${shorthand} \
--wordvec_file ./data/wordvec/English/en.twitter100d.xz \
--eval_file data/ner/en_tweet.dev.json

Syntactic NLP Models

You can specify two data settings:

  • en_tweetewt: trained on TB2+UD-English-EWT
  • en_tweet: trained on TB2

1. Tokenizer

## assign the shorthand name
shorthand=en_tweet

## Data Preparation
python -m stanza.utils.datasets.prepare_tokenizer_treebank ${shorthand}

## Train
python stanza/utils/training/run_tokenizer.py ${shorthand} --no_use_mwt

2. Lemmatization

## assign the shorthand name
shorthand=en_tweet

## Data Preparation
python -m stanza.utils.datasets.prepare_lemma_treebank ${shorthand} 

## Train
python stanza/utils/training/run_lemma.py ${shorthand}

3. POS Tagging

Note that for POS and depparse, if there are a pretrained word2vec files in the target folders, Stanza will prioritize using them even if you give the --wordvec_file argument. To avoid accidentally using a wrong word2vec, remember to add --no_pretrain.

## assign the shorthand name
shorthand=en_tweetewt

## Data Preparation
python -m stanza.utils.datasets.prepare_pos_treebank ${shorthand} 

## Train
python stanza/utils/training/run_pos.py ${shorthand} --wordvec_file ../data/wordvec/English/en.twitter100d.xz --no_pretrain
## We didn't use a pretrained wordvec file in our training process. To specify one, use --wordvec_pretrain_file.

4. Dependency Parser

## assign the shorthand name
shorthand=en_tweetewt

## Data Preparation
## --gold would give you gold data. But according to conventions we did not use gold data for our depparse model.
python -m stanza.utils.datasets.prepare_depparse_treebank ${shorthand} 

## Train
python stanza/utils/training/run_depparse.py ${shorthand} --wordvec_file ../data/wordvec/English/en.twitter100d.xz --no_pretrain 
## We didn't use pretrained word2vec file in our parser training, but the pretrain.pt generated in training the pos_tagger can be re-used here.

Summary

In general, we follow these steps:

1. Prepare the treebank

  • {shorthand} should be selected from en_tweet, tweet_ewt
  • {model} should be one of tokenizer, lemma, pos, depparse
python -m stanza.utils.datasets.prepare_{model}_treebank {shorthand}

2. Train and evaluate on the default dev set

python stanza/utils/training/{model}.py {shorthand}

Compared to Stanza, we do not include sentence splitting, so we commented out the error check for stanza.utils.conll18_ud_eval.UDError: There are multiple roots in a sentence.

3. Evaluate on the tweetbank v2 test set

python stanza/utils/training/run_{model}.py {shorthand} --mode predict --score_test