daaja

This repository has implementations of data augmentation for NLP for Japanese:

For Japanese

README_ja.md is written in Japanese.

README_ja

This library about usage and performance is also described in the following article.

• 日本語自然言語処理の Data Augmentation ライブラリ daaja を作りました
• 日本語文書分類・固有表現抽出タスクに対する Data Augmentation の性能検証

Install

pip install daaja

Example

• Quick Example

Augmenters

Augmenters provides various types of data augmentation methods.

Sentence Augmenter

Sentence Augmenter is a data augmentation method for sentences.

| Augmenter | ref | | :-------------------------: | ------------------------------------------ | | RandamDeleteAugmentor | [1] | | RandamInsertAugmentor | [1] | | RandamSwapAugmentor | [1] | | SynonymReplacementAugmentor | [1] | | BackTranslationAugmentor | [3] | | ContextualAugmentor | [4] |

• [1] EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks
• [3] Improving Neural Machine Translation Models with Monolingual Data
• [4] Data Augmentation using Pre-trained Transformer Models

How to use

from daaja.augmentors.sentence import SynonymReplaceAugmentor
augmentor = SynonymReplaceAugmentor()
augmentor.augment("日本語でデータ拡張を行う") #=> 日本語でデータ伸暢を行う

Sequence Labeling Augmenter

Sequence Labeling Augmenter is a data augmentation method for sequence labeling task.

| Augmenter | ref | | :--------------------------------: | ----------------------------------------- | | LabelwiseTokenReplacementAugmentor | [2] | | MentionReplacementAugmentor | [2] | | ShuffleWithinSegmentsAugmentor | [2] | | SynonymReplacementAugmentor | [2] |

• [2] An Analysis of Simple Data Augmentation for Named Entity Recognition

How to use

from daaja.augmentors.sequence_labeling import SynonymReplacementAugmentor

augmentor.augment(["君", "は", "隆弘", "君", "かい"], ["O", "O", "B-PER", "O", "O"])
# => (['は', '君', '隆弘', '君', 'かい'], ['O', 'O', 'B-PER', 'O', 'O'])

Methods

The same method as in the following papers can be tried in methods.

EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks

Command

python -m daaja.methods.eda.run --input input.tsv --output data_augmentor.tsv

The format of input.tsv is as follows:

1	この映画はとてもおもしろい
0	つまらない映画だった

In Python

from daaja.methods.eda.easy_data_augmentor import EasyDataAugmentor
augmentor = EasyDataAugmentor(alpha_sr=0.1, alpha_ri=0.1, alpha_rs=0.1, p_rd=0.1, num_aug=4)
text = "日本語でデータ拡張を行う"
aug_texts = augmentor.augments(text)
print(aug_texts)
# ['日本語でを拡張データ行う', '日本語でデータ押広げるを行う', '日本語でデータ拡張を行う', '日本語で智見拡張を行う', '日本語でデータ拡張を行う']

An Analysis of Simple Data Augmentation for Named Entity Recognition

Command

python -m daaja.methods.ner_sda.run --input input.tsv --output data_augmentor.tsv

The format of input.tsv is as follows:

私	O
は	O
田中	B-PER
と	O
いい	O
ます	O

In Python

from daaja.methods.ner_sda.simple_data_augmentation_for_ner import \
    SimpleDataAugmentationforNER
tokens_list = [
    ["私", "は", "田中", "と", "いい", "ます"],
    ["筑波", "大学", "に", "所属", "して", "ます"],
    ["今日", "から", "筑波", "大学", "に", "通う"],
    ["茨城", "大学"],
]
labels_list = [
    ["O", "O", "B-PER", "O", "O", "O"],
    ["B-ORG", "I-ORG", "O", "O", "O", "O"],
    ["B-DATE", "O", "B-ORG", "I-ORG", "O", "O"],
    ["B-ORG", "I-ORG"],
]
augmentor = SimpleDataAugmentationforNER(tokens_list=tokens_list, labels_list=labels_list,
                                            p_power=1, p_lwtr=1, p_mr=1, p_sis=1, p_sr=1, num_aug=4)
tokens = ["吉田", "さん", "は", "株式", "会社", "A", "に", "出張", "予定", "だ"]
labels = ["B-PER", "O", "O", "B-ORG", "I-ORG", "I-ORG", "O", "O", "O", "O"]
augmented_tokens_list, augmented_labels_list = augmentor.augments(tokens, labels)
print(augmented_tokens_list)
# [['吉田', 'さん', 'は', '株式', '会社', 'A', 'に', '出張', '志す', 'だ'],
#  ['吉田', 'さん', 'は', '株式', '大学', '大学', 'に', '出張', '予定', 'だ'],
#  ['吉田', 'さん', 'は', '株式', '会社', 'A', 'に', '出張', '予定', 'だ'],
#  ['吉田', 'さん', 'は', '筑波', '大学', 'に', '出張', '予定', 'だ'],
#  ['吉田', 'さん', 'は', '株式', '会社', 'A', 'に', '出張', '予定', 'だ']]
print(augmented_labels_list)
# [['B-PER', 'O', 'O', 'B-ORG', 'I-ORG', 'I-ORG', 'O', 'O', 'O', 'O'],
#  ['B-PER', 'O', 'O', 'B-ORG', 'I-ORG', 'I-ORG', 'O', 'O', 'O', 'O'],
#  ['B-PER', 'O', 'O', 'B-ORG', 'I-ORG', 'I-ORG', 'O', 'O', 'O', 'O'],
#  ['B-PER', 'O', 'O', 'B-ORG', 'I-ORG', 'O', 'O', 'O', 'O'],
#  ['B-PER', 'O', 'O', 'B-ORG', 'I-ORG', 'I-ORG', 'O', 'O', 'O', 'O']]

Reference

• https://arxiv.org/abs/1901.11196
• https://arxiv.org/abs/1901.11196
• https://github.com/jasonwei20/eda_nlp
• https://qiita.com/tchih11/items/aef9505d26d1bf06a04c
• https://user-first.ikyu.co.jp/entry/2021/07/27/155513