Skip to content

CLI

By using the CLI, you can pre-annotate texts with the trained model. Currently, login and predict commands are supported. login is the command to login to doccano and must be executed first. predict is the command to annotate data with the model. The usage is as follows:

Named Entity Recognition

To use this feature, you need to install doccano-client as follows:

pip install doccano-client[spacy]

First, you need to train a model for NER. You can use the following command to train a model:

python -m spacy train config.cfg \
  --output outputs \
  --paths.train /path/to/train.spacy \
  --paths.dev /path/to/dev.spacy \
  --gpu-id 0

Then, you can annotate data with the trained model:

docli login \
  --host http://127.0.0.1:8000 \
  --username admin \
  --password password
docli predict ner \
  --project <project_id> \
  --model <model_name> \
  --mapping [mapping.json] \
  --framework [spacy]

Currently, ner is supported as a task and spacy as a framework. Also, the framework and mapping options can be omitted.

$ docli predict ner \
    --project <project_id> \
    --model <en_core_web_sm>

Mapping can be a path to a JSON file. This JSON file consists of keys and values as shown below. The key is the label names output by the model, and the value is the label names you defined in doccano. In this way, the label names output by the model is converted to the defined label names so that they can be labeled.

{
   "LOCATION":"LOC",
   "ORGANIZATION":"ORG",
   "GPE":"LOC"
}

Note that if the label names output by the model are not defined in doccano and are not converted using mapping, the labels will be discarded.

This feature is integrated with spacy-partial-tagger, which allows you to train models with just a dictionary and some text. Please see the repository for more details.

Automatic Speech Recognition

To use this feature, you need to install whisper as follows:

pip install git+https://github.com/openai/whisper.git

In this feature, you can annotate audio file with OpenAI Whisper. For a list of available models, please see the official repository.

docli predict asr \
  --project <project_id> \
  --model base

Active Learning

To use this feature, you need to install doccano-client as follows:

pip install doccano-client[al]

First, you need to login to doccano:

docli login \
  --host http://127.0.0.1:8000 \
  --username admin \
  --password password

Then, you can use active learning as follows:

docli teach \
  --task <ner> \
  --project <project_id> \
  --lang [en] \
  --query_strategy [MNLP] \
  --train_frequency [50] \
  --transformer_model [bert-base-uncased] \
  --patience [-1]

Currently, only ner is supported as a task.