Writing Style Conversion using Neural Machine Translation

Writing style is an important indicator of a writer’s persona. In the age of intelligent chatbots, writing style conversion can enable intimate human-AI interaction, allowing us to bridge the inherent gap between AI agents and human beings. In this paper, we apply sequence to sequence neural machine translation model with global attention mechanism to two writing style conversion tasks, mostly focusing on Shakespearean style conversion task, to explore its capabilities and limitations. In order to acquire parallel corpora of two unique writing styles, we suggest a new method of data acquisition, which leverages the Google Translator engine when only given a single corpus of target writing style. Another decision problem that is crucial to the performance of this model is the embedding matrix of the source and target vocabulary, hyperparameters, global attention mechanism, and many other details of bidirectional Sequence to Sequence model. The bidirectional Seq2Seq model we suggest here outperformed the previous models[1] for style mimicking in BLEU score by greater than 25%. In addition, through human evaluated metrics, we could observe that our bidirectional Seq2Seq model performed better than our simple attentive Seq2Seq model in preserving original meaning and imitating target style.