目前情感计算的研究主要关注识别或预测文本情感状态，仍缺乏针对情感生成、领域适应等领域的探索，并且已有文本生成工作也存在缺少情感信息的问题。项目基于Valence-Arousal空间，研究面向多维连续情感空间的文本生成技术。区别于现有类别型方法，所使用的维度型方法可以提供更为准确细致的情感信息。具体包括三个方面的内容：1) 通过注意力机制和动态路由算法对树形长短期记忆模型进行加权，解决权重偏置问题，为后续研究提供句向量表示基础；2) 为了向情感生成提供多领域训练文本，引入对抗学习实现维度型领域适应模型，解决目前研究中的消极迁移问题，实现多领域维度型文本情感分析；3) 利用对抗生成网络和变分自编码器改进序列到序列生成模型，实现面向维度型情感状态的文本生成，解决目前文本生成无情感信息的问题。项目希望形成一套系统性的维度型情感生成方法和原型系统，为后续多种情感分析和生成应用提供理论基础。

Existing researches on affective computing mainly focus on identifying or classifying the emotion states of given texts. There is still a lack of relevant research works on affective generation and domain adaptation, etc. Hence, traditional text generation methods still suffered from the problem of lacking sentiment information. Based on continuous valence-arousal (VA) representation, the project proposes to research on affective generation in continuous space of multi-dimensional space. Compared with existing categorical approaches, such a dimensional approach can provide more fine-grained sentiment information. Main content includes:.(1).By introducing attention mechanism and dynamic routing algorithm, a self-attention tree-LSTM and a capsule tree-LSTM model will be proposed. These models are expected to address the problem of bias weight, and provide the sentence vector for the follow-up studies..(2).For providing training samples in different domains, a dimensional domain adaptation method will be proposed. By using adversarial learning, this method can tackle the negative transfer issue in existing works, and transfer dimensional affective information to multi-domains..(3).By using generative adversarial network and variational auto-encoder, the project seeks to improve Seq2Seq model, and build an affective text generation model in continuous VA space. This model can solve the problems of lacking affective information in current methods of text generation..Overall, this project hopes to build a framework and prototype system of dimensional affective generation, which provides a theoretical basis for the follow-up of various sentiment analysis and generation application.