Adversarial Style Transfer for Long Sentences
Wooyong Choi, Su Jeong Choi, Seyoung Park, Sang-Jo Lee · 2019 International Conference on Electronics, Information, and Communication (ICEIC) · 2019
Language style transfer is the task of generating a new sentence with the same content but of a different style. This task is a part of natural language generation and an attractive research topic. There is currently a spike in research on neural networks, and many studies are focused on language style transfer. One approach is a sequence-to-sequence model with multiple decoders. However, this approach cannot avoid the conventional problem of recurrent neural networks: it is difficult to handle long sentences. In this paper, we propose an adversarial style transfer model that generates a style-transferred sentence from a given sentence. To tackle the problem, we adopted a transformer structure to our model because this structure has proven effective for long sequences and requires lower computation costs. For evaluation, the model transferred the sentiments of sentences from a Yelp review dataset. As an experimental result, our proposed model outperforms the baseline and shows robustness for long sentences.