ABCD: Analogy-Based Controllable Data Augmentation

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.