1
Multi-events Driven Emotion Dynamic Generation
Using Hawkes Process
Xiang Nan
Liangjiang International College
Chongqing University of Technology
Chongqing , China
xiangnan@cqut.edu.cn
Zhang Mingmin
College of Computer Science &
Technology, Zhejiang Univ.
Hangzhou, China
zmm@cad.zju.edu.cn
Long Jianwu
School of Computer Science &
Engineering, Chongqing University of
Technology, China
jwlong@cqut.edu.cn
Abstract— Multi-events driven emotion generation was an
important research point in the affective computing field.
However, as the events have different types and occurred in
variable times, then computing the emotion state intensity
became a challenge. The existed solutions for this problem did
not take time influences of different event types into
consideration. In order to solve this problem, we provided a
Hawkes process based multi-events driven emotion generation
method. Firstly we appraised every event and generate the
related emotional reaction; secondly, we treated the emotion
generation process with a certain period as a point process and
trained the parameters of Hawkes process by maximum
likelihood estimation with real individual emotional reactions;
thirdly, we used Hawkes process to simulate the accumulated
emotion reactions. The experimental results showed that our
method can generate a multi-events driven emotion more
accurately and efficiently.
Keywords—Affective computing; emotion generation; Hawkes
process; multi-events driven
I. INTRODUCTION
Emotion generation was an important component of
affective computing [1] and was the foundation of emotion
analysis and expression. The purpose of emotion generation
was to create the virtual human with lifelike emotional
reactions. As a result, many researchers had modeled the
interaction environment for virtual human to generate
interaction emotions. However, the interaction environments
of human were complex that most of the modern emotion
generation models only took a message or an event as a
stimulus but neglected the affections of multi-stimulus. Note
that, as a social agent, the emotion generation of virtual human
must take the affections of multi-events into consideration,
otherwise the virtual human seemed to be lacking in
intelligence. As human can encounter vary types of events in a
time period, and all these events as stimulus would trigger
human accumulated emotional reactions. As a consequence,
affection weights of different types events to human emotional
reactions needed to be researched.
The decay of emotion state was another important point
for emotion generation. The current model using cognition
could generate correct emotional reactions to environmental
information, yet most of the models were intended to create
the emotional state of one moment. Although there existed
models that simulated the dynamic process of emotions,
however, these models often set parameters manually. Then
the robust of these models were relatively weak.
In order to solve these problems and made the virtual
emotion more realistic, our system adopted Hawkes process to
simulate the emotion dynamic process. Every emotional
reaction triggered by the events was treated as one point
process and different types of events had different affection
weights. Furthermore, the Hawkes process also can represent
the dynamic decay of emotion reactions. Then we used
emotion reaction process records to train the weight and decay
control parameters to make our method more robust.
The contributions of our works were firstly provided a
method to generate vary event driven emotion dynamic
process based on Hawkes process and secondly, using
parameter training to set the parameters so that the method
robust was improved.
II. BACKGROUND
OCC [2] model defined 22 emotional states which could
meet the basic needs of affective computing. Most of the
current emotion generation models were based on OCC
models such as the EMA [3] model provided by Gratch and
Marsella. They integrated cognition and coping actions into
their model to create emotional reactions to emergencies for
the virtual human. In order to construct virtual human ’ s
cognitive characteristics, Liu [4] combined motivation to the
cognitive appraisal framework. Yang [5] built a
comprehensive computational emotion generation model using
the appraisal that the model could generate action strategy
according to individual’s motivation. However, they did not
focus on the multi-events affections.
For solving this problem, OCC was often combined with
other theory. Egg [6] used OCEAN personality model
construct the long term emotional characteristics of the virtual
human. In addition, OCC model integrated with Fuzzy logic
formed FLAME [7] system. This system contained three
components, the learning component, the emotion component
and the decision component. The learning component was to
connect the environmental object and the emotion states it
triggered. Soleimani proposed a computational model using
guidelines from OCC emotion theory to formulate a system of
fuzzy inferential rules that is capable of tracking the emotional
states of the agent [8,9]. However, the decay of the emotion
triggered by different event types were not considered.
Grounded on OCC model, our model adopted six basic
emotional states to describe the emotional reactions of a
virtual human.
资源评论
