福州大学学报（自然科学版）2025，Vol.53Issue(4)：383-390,8.DOI:10.7631/issn.1000-2243.24225
面向多粒度特征融合与流增强的情感文本生成方法
Affective text generation for multi-granularity feature fusion and flow-enhanced
摘要
Abstract
To address the problem that the variational autoencoder fails to model the latent space of emotion-semantics interaction and the complexity of the true posterior in affective text generation,a multi-granularity feature fusion and flow-enhanced affective text generation(MF3-ATG)is proposed.Firstly,the features of local sentiment and global semantics are extracted by the multi-granularity infor-mation extraction module.Then,the correlation between sentiment semantics is further learned by guided attention on attention.Meanwhile,the posterior distribution is enhanced by using Householder transformation to address the text modeling capability,so as to generate more natural and diverse texts.Experimental results on Yelp and Amazon datasets show that,compared with the baseline models,MF3-ATG achieves an average relative improvement of 2.81％and 2.66％in mean absolute error and distinct metrics.关键词
文本生成/情感可控/变分自编码器/归一化流Key words
text generation/emotion controllable/variational autoencoder/normalizing flow分类
信息技术与安全科学引用本文复制引用
曾曼玲,杨州,蔡铁城,陈超,廖祥文..面向多粒度特征融合与流增强的情感文本生成方法[J].福州大学学报（自然科学版）,2025,53(4):383-390,8.基金项目
国家自然科学基金资助项目(61976054,62476060) （61976054,62476060）
