ABCD: Analogy-Based Controllable Data Augmentation

Shuo Yang*, Yves Lepage


研究成果: Conference contribution


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.

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
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


名前Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
13082 LNCS


Conference10th International Conference on Theory and Practice of Natural Computing, TPNC 2021
CityVirtual, Online

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)


「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。