ABCD: Analogy-Based Controllable Data Augmentation

Shuo Yang*, Yves Lepage

*この研究の対応する著者

研究成果: Conference contribution

抄録

We propose an analogy-based data augmentation approach for sentiment and style transfer named Analogy-Based Controllable Data Augmentation (ABCD). The object of data augmentation is to expand the number of sentences based on a limited amount of available data. We are given two unpaired corpora with different styles. In data augmentation, we retain the original text style while changing words to generate new sentences. We first train a self-attention-based convolutional neural network to compute the distribution of the contribution of each word to style in a given sentence. We call the words with high style contribution style-characteristic words. By substituting content words and style-characteristic words separately, we generate two new sentences. We use an analogy between the original sentence and these two additional sentences to generate another sentence. The results show that our proposed approach decrease perplexity by about 4 points and outperforms baselines on three transfer datasets.

本文言語English
ホスト出版物のタイトルTheory and Practice of Natural Computing - 10th International Conference, TPNC 2021, Proceedings
編集者Claus Aranha, Carlos Martín-Vide, Miguel A. Vega-Rodríguez
出版社Springer Science and Business Media Deutschland GmbH
ページ69-81
ページ数13
ISBN(印刷版)9783030904241
DOI
出版ステータスPublished - 2021
イベント10th International Conference on Theory and Practice of Natural Computing, TPNC 2021 - Virtual, Online
継続期間: 2021 12月 72021 12月 10

出版物シリーズ

名前Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
13082 LNCS
ISSN(印刷版)0302-9743
ISSN(電子版)1611-3349

Conference

Conference10th International Conference on Theory and Practice of Natural Computing, TPNC 2021
CityVirtual, Online
Period21/12/721/12/10

ASJC Scopus subject areas

  • 理論的コンピュータサイエンス
  • コンピュータ サイエンス(全般)

フィンガープリント

「ABCD: Analogy-Based Controllable Data Augmentation」の研究トピックを掘り下げます。これらがまとまってユニークなフィンガープリントを構成します。

引用スタイル