关键词: 复述生成, 编码器-解码器, 自注意力网络, 序列生成, 图生成, 联合学习

Abstract: Abstract: Paraphrase generation is a text data argumentation method based on natural language generation. Concerning the problems of repetitive generation, semantic errors and poor diversity in paraphrase generation methods based on the Seq2Seq framework, a paraphrase generation network based on joint learning of sequence and graph (J-SGPGN) was proposed. First, the encoder of J-SGPGN fused graph encoding and sequence encoding for feature enhancement. Then, the decoder was designed to contain two decoding methods, sequence generation and graph generation, for parallel decoding. Finally, the joint learning method was used to train the model, which aimed to simultaneously improve the accuracy and diversity of generation by taking into account both syntactic supervision and semantic supervision. The experimental results show that the generation accuracy evaluation indicator METEOR of J-SGPGN on the Quora dataset was 6.69% higher than the accuracy optimal baseline model (RNN+GCN). The generative diversity evaluation indicator Self-BLEU was 3.75% higher than the diversity optimal baseline model (BTmPG). It was verified that J-SGPGN could generate paraphrase texts with more accurate semantics and more diverse expressions.

Key words: paraphrase generation, encoder-decoder, self-attention network, sequence generation, graph generation, joint learning