An Attention-Based Long-Short-Term-Memory Model for Paraphrase Generation.

Neural network based sequence-to-sequence models have shown to be the effective approach for paraphrase generation. In the problem of paraphrase generation, there are some words which should be ignored in the target text generation. The current models do not pay enough attention to this problem. To overcome this limitation, in this paper we propose a new model which is a penalty coefficient attention-based Residual Long-Short-Term-Memory (PCA-RLSTM) neural network for forming an end-to-end paraphrase generation model. Extensive experiments on the two most popular corpora (PPDB and WikiAnswers) show that our proposed model’s performance is better than the state-of-the-art models for paragraph generation problem.