Research On PAD Emotional State Model-based Dialog Generation

The Seq2Seq model has achieved good results in the open domain dialog generation system.However,since Seq2Seq model does not focus on modeling affective information,the responses generated by the model tend to be emotionally inconsistent with the source sentences.In order to add external affective information into open domain dialog system and to guide it to generate affectively appropriate responses,we introduce PAD affective vectors,which is based on "PAD emotional state model",into Seq2Seq model,achieving a better performance than traditional Seq2Seq model without using it.The"PAD Emotional State Model"is a psychological model used to describe and measure emotional states.It contains three numerical dimensions:Pleasure,Arousal and Dominance.In theory,these three dimensions can be used to represent all emotional status.In this paper,according to a dictionary of PAD affective vectors,we first assign a three-dimensional PAD affective vector,which represents the corresponding affective information,to each word.Then,the PAD affective vector of the word and the word vector are used together as input to the encoder and would be jointly encoded,which augments additional affective information into Seq2Seq model.Secondly,in order that the decoder can selectively "focus on" the content information and affective information of the context in the decoding phase to generate appropriate affective responses,a"Joint Attention Mechanism" is proposed,including a"Content Attention Mechanism”and an"Affective Attention Mechanism"."Affective Attention Mechanism",which is based on the PAD affective vector,can selectively"focus on" and make full use of the affective information contained in PAD affective vectors of the words in source sentence when decoding,thereby guiding the model to generate responses with higher quality in affect.In psychology,empathy is a vital emotion-guiding mechanism.Inspired by this,in order to guide the model to produce"sympathetic"responses,we extend the“Cross Entropy Loss”and propose PAD affective vector-based“Affective Cross Entropy Loss”to replace the"Cross Entropy Loss" and be the loss function of Seq2Seq model.We compared and analyzed the effects of the model through automatic evaluation and human annotations.Experimental results show that our proposed PAD affective vectors-based Seq2Seq models can produce higher quality responses than the traditional Seq2Seq model.On the whole,according to the Perplexity automatic evaluation and human annotations results,S2S+W2AV+JAtt+Aff-XENT model performs the best among all the models we propose.Compared with the baseline model,the model is able to generate responses that are not only more appropriate in affect but also are more fluent,demonstrating the effectiveness of our proposed"Bidirectional LSTM-based Joint Encoding","PAD Affective Vector-based Affective Attention Mechanism"and"Affective Cross Entropy Loss".