User-Oriented Paraphrase Generation With Keywords Controlled Network

Abstract

Paraphrase generation can help with both downstream tasks in natural language processing (NLP) and human writing in our daily life. Most of the prevalent neural models focus on the former usages and generate uncontrolled paraphrase while they ignore the subtleties of users' requirement. In addition, the existing tools for users are usually rule-based which is unnatural due to the complexity of the paraphrase nature. To this end, we propose a keyword controlled network (KCN) which can be used as an assistant paraphrase generation tool. The KCN works in an interactive manner and generates different paraphrases given different keywords. The model is based on a Sequence-to-Sequence (Seq2Seq) framework integrated with copy mechanism. Given the source sentence and the keywords, two encoders transform them into vector representations. Then, the representations are fused together and used for the decoder. The decoder with attention mechanism either copies the words from the keywords or generates words from the whole dictionary. In the training stage, as the source sentence and the target sentence are all valid paraphrases, the model is trained to generate each given different keywords, which simulates the behaviors of users. The extensive experiments on three datasets show that our method outperforms baselines in the automatic evaluation (0.06 absolute improvement in BLEU) and the generated paraphrases meet user expectation in the human evaluation.