基于生成对抗网络的类别文本生成

Category Text Generation Based on Generative Adversarial Network

类别文本生成旨在让机器生成人类可理解的文本,并且赋予生成文本特定的类别属性。现有工作主要采用基于生成对抗网络的文本生成框架,往往直接采用卷积神经网络进行文本特征提取,缺乏对文本全局语义的关注;此外,简单地在生成网络中引入注意力无法有效消除解码过程中的噪声。针对上述问题,本文提出一种将文本全局特征与局部特征联合建模的方法,通过将长短时记忆网络提取的全局语义信息与卷积神经网络提取的局部语义信息进行融合,增强生成过程中对文本全局语义信息的关注,并且引入双重注意力,进一步过滤掉序列生成中的无关信息。与基准模型相比,本文提出的方法分别在2个公开的真实数据集(Movie Review和Amazon Review)上取得了至少0.01和0.004的BLEU值的提升,表明了本文方法的有效性。

The research of category text generation is a task to enable machines to generate human-understandable text,and give the text specific attributes.The CNN network has been used to extract the features of the text in most of the existing text generation work based on Generative Adversarial Networks that lacks attention to the global semantics.In addition,the attention mechanism is simply introduced to the generator,which can not effectively eliminate the noise in the decoding process.For this reason,this paper proposes a joint modeling method of local semantics and global semantics.The global semantic information extracted by the LSTM and the local semantic information extracted by the CNN are fused,so that the attention on the semantic information is enhanced during the generation.Moreover,the attention on attention is introduced in the generator to further filter irrelevant information during the sequence generation.Compared with the benchmark models,the method proposed in this paper achieves at least 0.01 and 0.004 improvement in BLEU values on two public real datasets(Movie Review and Amazon Review),demonstrating the effectiveness of the proposed method.