Application and Research on Affection Model Based on Bayesian Network

It needs not only intelligence but also emotion for the computer to realize harmonious human computer interaction, which is one of the research focuses in the field of computer science. This paper proposes a hierarchical approach to represent personality, affection and emotion, using Bayesian Network to affection model and show emotion via virtual human’s facial expression. The affection model was applied to an affective HCI system, proved to be simple, effective and stable. 1 Hierarchical model of the virtual human 1.1 Instruction of the hierarchical model We construct a hierarchical model: The Personality-affection-emotion model. Based on OCEAN model in psychology field, We classify human’s personality into five dimensions:[1] Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism; each factor corresponds to one dimension in the personality space[2], and each dimension is closely-related with facial expression and affection representation. We classify affection into positive and negative [3] adopting the most popular classification of basic emotions: happiness, surprise, fear, sadness, disgust, and anger, in addition, we add a neutral emotion. Corresponding to emotions, we use Ekman’s theory, six basic facial expressions[4][5][6] and another neutral facial expression. 1.2 Extension of the AIML tag We take the chatting robot ALICE as our virtual human, which is based on AIML (Artificial Intelligence Markup Language) Technology. When inputs a question, it will produce a relative answer. There are detailed descriptions about AIML in literature[7]. In order to endow ALICE with emotion, we add an emotion tag to represent her response emotion. There are seven emotion tags corresponding to seven basic emotions mentioned above. For example(5% probability of sad, 95% probability of happy): How are you doing nowadays? Everything is running smoothly. 2 Construction of the affection model based on Bayesian Network Fig. 1. Affection model for each personality As Fig.1 shows, we construct an affection model based on Bayesian network involving two parent nodes and one child node, one corresponding model for each personality factor of the OCEAN model. Of course, user can combine any two or several factors arbitrarily .For example, user can totally constructs such a personality: 20% openness and 80% Neuroticism, the value range of “Current Affection Ac” and “Response Affection Ar” in Fig.1 is either positive or negative. Initial value of Ac depends on different personality. Ar is extracted from the emotion tags of ALICE’s answer. There are different conditional transition probabilities for each personality π to decide next affection. The probability represents affective process. We get the conditional probability for the changing of affection P (An|Ac, Ar) after training to ALICE with different personality. Once the conditional transition probability and prior probability P (ei) are given, a possible affection can be definite according to the following formula: P (An) = BBN(Ac, Ar, π) = P (An|Ac, Ar)P (ei) (1) P (An)decides the affection change. When P (An) > k(a threshold, 0 ≤ k ≤ 1 ), we choose An as the next affection state. Otherwise, hold the previous affection state. The history preserves P (An) for the next computation. 3 Conversion from affection to emotion When the next affection state is certain, it’s necessary to choose qualifying emotion response to control the virtual human’s facial expression. There are three key factors deciding the emotion state: ALICE’s response emotion er , current affection An (output of the affection model) and the previous emotion state, defined as ep . The first key factor can be easily controlled by adding emotion tags in ALICE’s answers when we establish AIML database. The second key-factor is the output of the affection model. As for positive affection and negative affection, we defined emotion state transition probability matrix respectively and experientially. More users testing the system can certainly optimize these values, getting more believable results. Formula (2) shows how to get the next emotion state en : P (en) = ΓAn(Ex(ep), Ex(er))P (er) (2) ΓAn denotes the transition probability matrix of affection An. Ex(er) denotes the corresponding expression. we also set a threshold s, (0 ≤ s ≤ 1) .Only if P (en) ≥ s ,the emotion state changes to en , otherwise , it will remain unchanging.