Bunruija

Bunruija is a text classification toolkit. Bunruija aims at enabling pre-processing, training and evaluation of text classification models with minimum coding effort. Bunruija is mainly focusing on Japanese though it is also applicable to other languages.

See example for understanding how bunruija is easy to use.

Features

• Minimum requirements of coding: bunruija enables users to train and evaluate their models through command lines. Because all experimental settings are stored in a yaml file, users do not have to write codes.
• Easy to compare neural-based model with non-neural-based model: because bunruija supports models based on scikit-learn and PyTorch in the same framework, users can easily compare classification accuracies and prediction times of neural- and non-neural-based models.
• Easy to reproduce the training of a model: because all hyperparameters of a model are stored in a yaml file, it is easy to reproduce the model.

Install

pip install bunruija

Example configs

Example of sklearn.svm.SVC

data:
  label_column: category
  text_column: title
  args:
    path: data/jsonl

output_dir: models/svm-model

pipeline:
  - type: sklearn.feature_extraction.text.TfidfVectorizer
    args:
      tokenizer:
        type: bunruija.tokenizers.mecab_tokenizer.MeCabTokenizer
        args:
          lemmatize: true
          exclude_pos:
            - 助詞
            - 助動詞
      max_features: 10000
      min_df: 3
      ngram_range:
        - 1
        - 3
  - type: sklearn.svm.SVC
    args:
      verbose: false
      C: 10.

Example of BERT

data:
  label_column: category
  text_column: title
  args:
    path: data/jsonl

output_dir: models/transformer-model

pipeline:
  - type: bunruija.feature_extraction.sequence.SequenceVectorizer
    args:
      tokenizer:
        type: transformers.AutoTokenizer
        args:
          pretrained_model_name_or_path: cl-tohoku/bert-base-japanese
  - type: bunruija.classifiers.transformer.TransformerClassifier
    args:
      device: cpu
      pretrained_model_name_or_path: cl-tohoku/bert-base-japanese
      optimizer:
        type: torch.optim.AdamW
        args:
          lr: 3e-5
          weight_decay: 0.01
          betas:
            - 0.9
            - 0.999
      max_epochs: 3

CLI

# Training a classifier
bunruija-train -y config.yaml

# Evaluating the trained classifier
bunruija-evaluate -y config.yaml

Config

data

You can set data-related settings in data.

data:
  label_column: category
  text_column: title
  args:
    # Use local data in `data/jsonl`. In this path is assumed to contain data files such as train.jsonl, validation.jsonl and test.jsonl
    path: data/jsonl

    # If you want to use data on Hugging Face Hub, use the following args instead.
    # Data is from https://huggingface.co/datasets/shunk031/livedoor-news-corpus
    # path: shunk031/livedoor-news-corpus
    # random_state: 0
    # shuffle: true

data is loaded via datasets.load_dataset. So, you can load local data as well as data on Hugging Face Hub. When loading data, args are passed to load_dataset.

label_column and text_column are field names of label and text.

Format of csv:

category,sentence
sports,I like sports!
…

Format of json:

[{"category", "sports", "text": "I like sports!"}]

Format of jsonl:

{"category", "sports", "text": "I like suports!"}

pipeline

You can set pipeline of your model in pipeline section. It is a list of components that are used in your model.

For each component, type is a module path and args is arguments for the module. For instance, when you set the first component as follows, TfidfVectorizer is instanciated with given arguments, and then applied to data at first in your model.

  - type: sklearn.feature_extraction.text.TfidfVectorizer
    args:
      tokenizer:
        type: bunruija.tokenizers.mecab_tokenizer.MeCabTokenizer
        args:
          lemmatize: true
          exclude_pos:
            - 助詞
            - 助動詞
      max_features: 10000
      min_df: 3
      ngram_range:
        - 1
        - 3

Prediction using the trained classifier in Python code

After you trained a classification model, you can use that model for prediction as follows:

from bunruija import Predictor

predictor = Predictor.from_pretrained("output_dir")
while True:
    text = input("Input:")
    label: list[str] = predictor([text], return_label_type="str")
    print(label[0])

output_dir is a directory that is specified in output_dir in config.