基于位置感知的情感可控对话生成模型研究

Position-awared Conversation Generation Model with Emotion Contrlled

基于序列到序列的对话生成在实现情感状态转移时大多采用外部情感词嵌入的方式,编码器很难捕获解码器的情感状态,解码器被强制嵌入的外部情感词干扰,造成生成回复情感词堆叠及缺乏情感信息上下文。为解决上述问题,该文提出基于位置感知的情感可控对话生成模型。在编码的过程中,当前输入词向量和位置向量共同参与编码,在不影响当前输入的情况下,上文信息利用分层的编码方式增加额外编码信息。在解码的过程中,利用遮蔽语言的性能,强制模型进行内容理解和学习,编码器和解码器的联合训练能够生成符合语法的情感回复。实验结果表明,位置感知的加入进一步刻画了数据的潜在结构信息,提高了情感可控对话生成的语言质量。

Sequence to sequence generation model mostly uses the way of adding external emotional words when the model transfers the emotional state,which is defected in generation responses with emotional word stacks and lack of emotional information context.To address this issue,this paper proposes an emotion controllable conversation generation model based on position awareness.In the encoding process,the current input word vector and position vector jointly participate in encoding.Without affecting the current input,the preceding context is encoded by an additional leyer.In the decoding process,the masked model is used to force the model to understand and learn the content.The joint training of the encoder and the decoder could generate a grammatical emotion response.The experimental results show that the position awareness further characterizes the potential structural information on the data and improves the model quality.