emotions to numeric PAD emotions, thereby providing an opportunity to embed .... 1: The OCC Model. Table 1: Sample Emotion Terms Grouped on PAD Value.
From Descriptive to Numeric Emotions in Agents Hong Jiang Mathmatics & Computer Science Department Benedict College, Columbia, SC, U.S. Abstract— Representing emotions in agents is important for designing agents that behave like humans or are intended to interact with humans. OCC model, a descriptive emotion model, has been used widely as a standard for emotion synthesis, however it is also criticized for falling short on suggestions for what to do with the emotional states. PAD emotion scale, a numeric emotion measurement of emotions, supplies a highly useful and convenient assessment of consumer emotional reactions to services, products, or combinations of products and services. It closely connects human’s emotional states with the behaviors. If we can find out the relationship between these two models, we can produce a more complete emotion mechanism by using OCC to generate emotional states and PAD to analyze them. Meanwhile, when we consider emotional negotiation, a numeric measurement of emotions is desirable. We need to find a proper way to embed emotions into a utility function. This paper presents a method to map descriptive OCC emotions to numeric PAD emotions, thereby providing an opportunity to embed both the OCC model and the PAD emotion scale into emotional agents to produce a more complete emotion mechanism. The mechanism includes both generating emotions and predicting the effects of emotions on agents’ reasoning processes and behaviors. More specifically, we use WordNet as a lexical ontology and apply ontology matching techniques to produce a mapping from the 22 basic descriptive emotions generated by the OCC model to the numeric PAD emotion space. The result is consistent with the senses in WordNet and the situations in the OCC model. Keywords: Emotional agent, Descriptive Emotion, Numeric Emotion, OCC Model, PAD Emotion Scale
1. Introduction In constructing agents that are intended to either interact with humans or operate on behalf of humans, it is important to include an emotion component. Important aspects are how to represent emotions in agents and how to use the representations as part of the agents’ reasoning and decision making. To date, most of the models for emotions are descriptive, and a large number of these are based on the OCC model. The OCC model is a computational emotion model, developed by Ortony, Clore, and Collins [13], which has established itself as the standard model for emotion synthesis. Many projects have employed the OCC model to
generate emotions for their embodied characters. However, it has been criticized for falling short on suggestions of what to do with the emotional state, and Bartneck [1] points out that the OCC model should be simplified to match the abilities of the embodied emotional character. On the other hand, descriptive emotions are not sufficient for emotional agents. As Jiang, Vidal, and Huhns showed in [4], emotional agents in negotiation need a numeric measurement for emotions, and they used the PAD (PleasureArousal-Dominance) emotion scale [7], [8] to describe emotions numerically. They also presented a mechanism to show how those numeric emotions affect the negotiation process. Other than negotiation, such numeric emotions are useful for a wide range of other applications [10]. They are used to assess consumer reactions to products, services, and shopping environments. Additionally the numeric scales can be used to assess the emotional impact of a workplace, an advertisement, or a medical or psychotropic drug. Mehrabian in [11] also mentions an effort to incorporate PAD in Artificial Intelligence. Becker [2] applies the PAD into a virtual human called Max. Most of the applications of PAD are closely related to the effects of emotions on agents’ reactions, or what to do with the emotional state. In some sense, with emotions in this numeric measurement, it is easier to simulate emotions’ effects on agents’ behavior. Based on the idea in [3], [5] which supplies a generic architecture called EBDI for an emotional agent, it is possible to plug in emotion models as needed by separating practical reasoning techniques from the specific emotion mechanism. Thus, we can simply build a complete emotional agent by embedding the OCC model and PAD into the EBDI architecture. We can use the OCC model to generate emotions and use PAD as a base to model agents’ reactions. The remaining problem is how to map those descriptive emotions generated by the OCC model to numeric PAD emotions. This research paper is the first to address this issue. This paper describes a method to map descriptive emotions to numeric emotions. More specifically, we use WordNet [12], a semantic lexicon for the English language, as a lexical ontology. We then apply ontology matching techniques to achieve a mapping from the 22 basic descriptive emotions generated by the OCC model to a numeric PAD emotion space. The result is consistent with the senses in WordNet and the situation in the OCC model.
2. Related Background 2.1 OCC Model with 22 Basic Emotions The OCC model, as shown in Figure 1, is a computational emotion model developed by Ortony, Clore, and Collins [13]. There are a large number of studies that have employed the OCC model to generate emotions for their embodied characters, and now it has established itself as the standard model for emotion synthesis. Many emotion model designers believe that the OCC model is all that is needed to design emotional agents. However, Bartneck [1] criticizes OCC in observing that it falls short on suggestions for what to do with the emotional state, and points out that the OCC model should be simplified to match the abilities of the embodied emotional character. The OCC model assumes that there are three major aspects of the world, or changes in the world, upon which one can focus. The three aspects are events, agents, and objects. One focuses on events when one is interested in their consequences; one focuses on agents, because of their actions; and one focuses on objects, when one is interested in certain aspects or imputed properties of them. So, emotions are valenced reactions, and any valenced reaction is always a reaction to one of these perspectives about the world. This model specifies 22 basic emotions based on valenced reactions to situations constructed either as being goal relevant events, as acts of an accountable agent, or as attractive or unattractive objects. The basic emotions are grouped as 11 pairs: happy-for and resentment, gloating and pity, hope and fear, satisfaction and fear-confirmed, relief and disappointment, joy and distress, gratification and remorse, gratitude and anger, pride and shame, admiration and reproach, love and hate. This model also offers some variables, such as likelihood of an event or the familiarity of an object to determine the intensity of the emotion types. Briefly, the OCC model is claimed to contain a sufficient level of complexity and details to cover most situations to generate various emotions.
2.2 PAD Emotion Scale The PAD (Pleasure-Arousal-Dominance) emotional state model is a general but precise three-dimensional approach to measuring emotions. Mehrabian [7] reviews versions of the PAD scales with different dimensions, and lists sets of studies that report development and refinement of a final set of the scales. The studies consistently find three nearly orthogonal dimensions: pleasure–displeasure, arousal–nonarousal,
and dominance–submissiveness. Pleasure–displeasure distinguishes the positive–negative affective quality of emotional states, arousal–nonarousal refers to a combination of physical activity and mental alertness, and dominance– submissiveness is defined in terms of control versus lack of control. The analysis shows that these three dimensions provide a parsimonious base for the general assessment of emotional states. Specific terms describing emotions can be visualized as points in a three-dimensional PAD emotion space. Alternatively, when the PAD scale scores are standardized, each emotion term can be described succinctly in terms of its values on the pleasure–displeasure, arousal–nonarousal, and dominance–submissiveness axes. The following sample ratings illustrate definitions of various emotion terms when scores on each PAD scale range from -1 to +1: angry (-.51, .59, .25), bored (-.65, -.62, -.33), curious (.22, .62, -.01), dignified (.55, .22, .61), elated (.50, .42, .23), hungry (-.44, .14, -.21), inhibited (-.54, -.04, -.41), loved (.87, .54, -.18), puzzled (-.41, .48, -.33), sleepy (.20, -.70, -.44), unconcerned (-.13, -.41, .08), violent (-.50, .62, .38). Thus, according to the ratings given for “angry,” it is a highly unpleasant, highly aroused, and moderately dominant emotional state. “Sleepy” consists of a moderately pleasant, extremely unaroused, and moderately submissive state, whereas “bored” is composed of highly unpleasant, highly unaroused, and moderately submissive components. Within the PAD Model, there are eight basic and common varieties of emotion, as defined by all possible combinations of high versus low pleasure (+P and –P), high versus low arousal (+A and –A), and high versus low dominance (+D and –D). For instance, the (–P+A–D) states include feeling aghast, bewildered, distressed, in pain, insecure, or upset; (– P+A+D) states include feeling angry, catty, defiant, insolent, and nasty; and (+P+A+D) states include feeling admired, bold, carefree, excited, mighty, and triumphant. By focusing on individuals, extroverted, arousal seeking, exhibitionistic, nurturing, and affiliative persons are exuberant (i.e., pleasant, arousal, dominant). However, they may differ in terms of the weights of Trait Pleasure (P), Trait Arousability (A), and Trait Dominance (D) associated with each. Dependent persons are pleasant, arousal, and submissive. Anxious or neurotic persons are unpleasant, arousal, and submissive, whereas aggressive persons are unpleasant, arousal, and dominant (e.g., [8]). Mehrabian [9] provided equations showing relationships of specific personality measures to the PAD temperament dimensions. Some sample emotion terms grouped according to Pleasure, Arousal, and Dominance value are given in [7], where the preceding eight groups of emotion terms are derived from ratings of 240 emotion terms with the preliminary state pleasure, state arousal, and state dominance scales. Table 1 summarizes the appendix of [7] and some other samples in documents.
Fig. 1: The OCC Model
Table 1: Sample Emotion Terms Grouped on PAD Value PAD Type (+P+A+D) (+P+A–D) (+P–A+D) (+P–A–D) (–P+A+D) (–P+A–D) (–P–A+D) (–P–A–D)
Sample Emotion Terms Admired, bold, carefree, dignified, elated, excited, masterful, mighty, triumphant Amazed, curious, fascinated, grateful, impressed, loved, respectful At ease, comfortable, relaxed, satisfied, secure, unperturbed Consoled, cruel-admired, docile, domineering-timid, humiliated-lonely, protected, reverent, sleepy,tranquilized Angry, catty, cruel, defiant, hostile, insolent, nasty, unmotivated-distressed, violent Aghast, bewildered, distressed, fear, frustrated, hungry, in pain, insecure, neuroticism, puzzled, anxiety, upset Amazed-daring, awed-domineering, disdainful, humiliated-sad, indifferent, selfish-uninterested,uncaring,unconcerned Bored, despairing, fatigued, inhibited, lonely, sad, sluggish, subdued
3. Descriptive Emotion to Numeric Emotion We must then solve the problem of mapping the above emotions to PAD values. We perform this mapping by following the principle of ontology matching [6] in which vocabularies generally are separated into lists of classes, predicates and instances, and then compared class vs. class,
predicate vs. predicate, etc. However, sometimes it is desirable to compare whole vocabularies without such classification since some authors may represent similar concepts with different types of terms. Thus, we can consider the mapping in some general sense first. Instead of giving specific values of P, A, and D, we can first map the 22 basic emotions generated by the OCC model to the 8 varieties of PAD emotions. If needed, we can figure out the difference in the value of P, A, and D later. Based on this idea, the problem can be simplified as well.
3.1 Method Description Before describing the method to get the mapping, we reconsider all the possible resources we can use: 1) 22 descriptive emotions generated by the OCC model. We can use them as the source words. 2) The lists of sample emotions in PAD. We can use them as the target words. 3) Some online tools, such as WordNet [12]. WordNet is a semantic lexicon for the English language. It groups English words into sets of synonyms called synsets, provides short, general definitions, and records the various semantic relations between these synonym sets. It has been used in automatic text analysis and artificial intelligence applications. Here we can use semantics of the words from WordNet to build a lexical ontology. 4) The specific process of the OCC model and detailed
definition about the PAD scale, since the matching result should be consistent with the situation described in the OCC model and the PAD definition. Our task is to produce a list of matched pairs. Each pair contains two terms: one from the source, the OCC emotion set, and the other from either the target, sample emotions in PAD, or the PAD emotion set. Each term can be multi-word, such as “happy-for”, etc. To find the best matching, we describe first the following five possible approaches: • Whole word matching: In a general sense, if the words are the same by ignoring the case difference, we believe they have the same meaning and we can have a match. This is the first as well as the simplest procedure to be executed. The terms in both ontologies are converted to lowercase and then compared for an exact string match on their names. For example, we have “fear” in 22 OCC emotions, and we also have “fear” in PAD sample emotion terms; thus we directly map “fear” to (–P+A–D). • Word constituent matching: This approach simplifies the matching by getting rid of the part of the word constituent that does not contain the meaning so as to focus on the meaningful part. With this approach, each term is broken into words wherever there is a capital letter, a hyphen, or an underscore. Words such as “a”, “the”, “of”, “in”, “for”, etc. are dropped from multi-word terms. For example, “happy-for” is changed to “happy”. The remaining words for each term are morphologically processed and compared in exact string match to words of each term from the target ontology. Using this procedure, un-obvious matching term pairs such as “anger” and “angry”, “satisfaction” and “satisfied” can be found. • Synset index matching: This approach explores the semantic meanings of the word constituents by using the WordNet [12] synsets to help identify synonyms in matching. A synset is a WordNet term for a sense or a meaning by a group of synonyms. The procedure is similar to the method in word constituent matching in decomposing multi-word terms into their word constituents, except that it does not perform direct matching between the words. For each word in each term in each ontology, if it is in WordNet, then it must belong to one of the synsets and have at least one WordNet synset index number. The procedure associates the WordNet synset index numbers of the constituent words with the term. The synset index numbers are close for synonyms. Thus the two terms that have the largest number of common synsets are recorded and presented. • Type matching: If we can not find a match through the above three approaches, then we can use this one, which explores the ontological category of each word constituent for matching. It is based on the hyponyms
and hypernyms from WordNet synsets. For example, if A is a feeling of joy and B is also a feeling of joy, and we did not find a match of A and B by using the previous methods, then we can use this one to match A and B. Thus, using this procedure, terms that cannot be matched by the previous methods, either string comparison or sense comparison, will be matched if they represent classes or properties of the same type. • Situation matching: By applying the above methods recursively, we can obtain a generic mapping; however, we still need to do some adjustment to fit our specific situation if there is some conflict. For example, we can get a mapping from the 22 OCC emotions to PAD sample emotions based on the above methods, and from the PAD sample emotions we can easily find out which one of the eight PAD variations the matched word belongs to. However, considering the detailed process in the OCC model, if the kind of emotion is a response of an agent to itself, the “Dominance” value should be set to neutral or 0. Because in PAD “Dominance” alludes to some relationship between the subject and the object, in this case, it is all about itself, and not about others. In specific mapping process, by using only one approach, we may not get a good match. Thus, if one approach does not work, we need to try another one or try to apply the above approaches recursively. We may need to combine them to get a match. The detailed matching process is described in Section 3.2.
3.2 Mapping Process and Results The detailed mapping process is defined as: 1) Apply whole word matching to the 22 basic emotions from the OCC model and the set of the sample emotions in PAD. If we find a match for some word, record these matched pairs and delete the word from the set of basic emotions to be matched. 2) Apply word constituent matching to both the set of basic emotions from the OCC model and the set of sample emotions in PAD. If we find a matched pair, record the matched pair. If there are no more matched pairs, update the set of basic emotions and the set of the sample emotions with the word constituents generated during the process, which simplifies the two sets as well. 3) Based on the simplified sets from the above process, we apply synset index matching to find matches. If we find any matched pair, we record it. 4) Otherwise, we apply type matching. If we find any matched pair, we then record it. 5) Delete the matched basic emotions from the basic emotion set. 6) Repeat steps 1 to 5 until the set of basic emotions is empty.
Happy-for
Happy
Happiness
Elation
Satisfied (+P–A+D) Resentment
Hostility
Gloating
Elation
Elated (+P+A+D)
Hostile (–P+A+D) Elated (+P+A+D)
Table 2: Synset Index Numbers for OCC Emotions
Pride Pity Hope
Shame admiration
Admired (+P+A+D)
Pride Relief
Relaxation
Disappointment
Elation
Word Class adjective
resentment gloating pity
noun noun verb noun
hope
verb noun
Relaxed (+P–A–D)
Upset (–P+A–D) Frustration
Joy
OCC Emotions happy
Frustrate (–P+A–D)
Elated (+P+A+D)
verb Pride
Elation
Elated (+P+A+D)
fear
Triumphant (+P+A+D) Shame
Shame
Gratification Remorse
verb
Upset (–P+A–D)
Reproach
Satisfaction Loneliness
noun
Satisfied (+P–A+D)
satisfaction
noun
relief
noun
disappointment joy distress
noun noun verb noun
pride
verb noun
shame
verb noun
Lonely (–P–A–D)
Disappointment Gratitude Hate
Gratefulness Hostility
Grateful (+P+A–D)
Hostile (–P+A+D)
Fig. 2: Mapping by Synset Matching and Type Matching
7) Apply situation matching to refine the matching results. By following the above processes, we have the following mapping by applying whole word matching and word constituent matching: fear ⇒ fear (–P+A–D); satisfaction ⇒ satisfied (+P–A+D); distress ⇒ distressed (–P+A–D); admiration ⇒ admired (+P+A +D); anger ⇒ angry (–P+A+D); love ⇒ loved (+P+A–D). To apply the synset index matching, we first obtain the synset index numbers from WordNet for the 22 OCC emotions, as in the Table 2. Then, we do the synset index matching, followed by type matching. By repeating those steps, we get the mappings shown in Figure 2. Finally, considering the relationship between the original OCC emotions and the mapped emotions, as well as the detailed situation for each of the OCC emotions, we get the following PAD representation, as in the Table 3. If an emotion may have a positive or negative value on some dimension, we set it to be “N” – neutral. For example,
verb admiration
noun
reproach gratification remorse gratitude anger
noun verb noun noun noun noun
love
verb noun verb
hate
noun verb
Synset Index Numbers 01194588, 01088951, 02649875, 01040585 07446390 07429832 00874329, 02147144 07451394, 07205004, 04774185 01804852 07409610, 07438591, 05875007, 10032734, 10896562, 04792935 01809579, 01794298, 00697932 07417148, 07422123, 07418802 01763725, 01763198, 01763564, 01763430, 01761564 07428849, 13800811, 13121678, 01058249 07391167, 14253998, 13119882, 10488747, 01194210, 15073413, 07256304, 01073784, 00351073, 04027856, 01061233 07438140, 00066741 07424946, 05756981 01796740, 01796355 07394350, 14284168, 14134837, 00083473 01780950 07406373, 07429110, 04831402, 07886934, 00746908 01755494 07404456, 14248441, 07205004 02522971, 02483927, 01775138, 01097127 07398628, 07407864, 01203324 06623791, 14250018 00817162 13800811, 01058464 07433612 07402230 07414249, 13850329, 00747687 01768967, 01769975 07440729, 05741438, 09704247, 07386227, 13421623, 00834429 01758160, 01811592, 01758531, 01414190 07443888 01757132
“happy” has one sense that is mapped to “elated” (+P+A+D), and another is mapped to “satisfied” (+P–A+D), then we set the “A” value to be “N” since it can be positive or negative. The “D” value is also adjusted according to the relationship between the subject and the object in the OCC model. Table 3: OCC to PAD Mapping Result OCC Emotions happy-for resentment gloating pity hope fear satisfaction fear-confirmed relief disappointment joy distress pride shame admiration reproach gratification remorse gratitude anger love hate
Pleasure + – + – + – + – + – + – + – + – + – + – + –
Arousal N + + + + + – – – + + + + + + + – – + + + +
Dominance + + + – + – + – + – N N N N + – + – – + – +
3.3 Evaluation Through the above processes, we get the matched pairs from the 22 basic OCC emotions to the sample emotions of PAD. Then based on the sample emotions of the PAD, we map the 22 basic OCC emotions to the 8 varieties of PAD emotions, and the mapping result is in the Table 3. Checking back with our goal, this mapping should be consistent with the common senses of the meaning, and there should not be any conflict with the specific situation of the OCC model or definition of the PAD scale. We discuss these two aspects as follows. • Consistency with common sense: During the mapping process, WordNet is used as a major source. WordNet is interpreted and used as a lexical ontology. The hypernym/hyponym relationships among the synsets is interpreted as a kind of specialization relation between conceptual categories. Since WordNet is a semantic lexicon for the English language, it contains the common meanings for each basic emotion. And, the process gives the best matches based on available resources. Thus the matching result is consistent with the common senses of the meanings for the emotions. • Consistency with specific situations: WordNet is a universal database that has a lot of basic semantics, but it does not consider the specific situation in the OCC model or detailed definition for the PAD
scale. However, our mapping process considers this issue and makes some adjustments. As described in the last approach, situation matching, it checks consistency with the specific situations and removes the potential conflicts. Thus, the final result does not have any conflict with the specific situation in the OCC model or the detailed definition for the PAD emotion scale. Above all, the mapping result achieves our goal successfully. It reflects common sense and is consistent with the specific situation as well.
4. Conclusion and Future Work By using the online tool WordNet as a lexical ontology and applying ontology matching techniques, we achieve a mapping from descriptive emotions to numeric emotions. More specifically, the mapping is from the 22 basic emotions generated by the OCC model to the PAD sample emotion space and then to the 8 varieties of PAD emotions. The result is consistent with the senses in WordNet and the situations in the OCC model. However it is possible to get a better mapping by using different methods. For example: we can apply a statistical method, such as asking 100 or more people to do this mapping subjectively. Thus, one of our future plans is to try different methods to achieve this mapping, and compare the results to get an improved mapping. Other future work is to figure out the detailed value for those basic emotions in P, A, D dimensions and embed the result into some specific application, i.e., to design an emotional agent with a complete emotion mechanism. The complete emotion mechanism would include both generating emotions and predicting the effects of emotions on agents’ reasoning processes and behaviors.
References [1] C. Bartneck. Integrating the occ model of emotions in embodied characters. In Proceedings of the Workshop on Virtual Conversational Characters: Applications, Methods, and Research Challenges, Melbourne., 2002. [2] C. Becker and I. Wachsmuth. Modeling primary and secondary emotions for a believable communication agent. In Proceedings of International Workshop on Emotion and Computing, in conj. with the 29th annual German Conference on Artificial Intelligenz (KI2006), pages 31–34, Bremen, Germany, 2006. [3] H. Jiang and J. M. Vidal. From rational to emotional agents. In Proceedings of the AAAI Workshop on Cognitive Modeling and Agentbased Social Simulation, 2006. [4] H. Jiang, J. M. Vidal, and M. N. Huhns. Incorporating emotions into automated negotiation. In Proceedings of the Agent Construction and Emotions Workshop, 2006. [5] H. Jiang, J. M. Vidal, and M. N. Huhns. EBDI: An architecture for emotional agents. In Proceedings of the Autonomous Agents and Multi-Agent Systems Conference, 2007. [6] J. Li. Lom: A lexicon-based ontology mapping tool. http: //reliant.teknowledge.com/DAML/I3con.pdf, 2007. [7] A. Mehrabian. Framework for a comprehensive description and measurement of emotional states. Genetic, Social, and General Psychology Monographs, 121(3):339–361, August 1995.
[8] A. Mehrabian. Relationships among three general approaches to personality description. Journal of Psychology, 129:565–581, 1995. [9] A. Mehrabian. Pleasure-arousal-dominance: A general framework for describing and measuring individual differences in temperament. Current Psychology, 14:261–292, 1996. [10] A. Mehrabian. General tests of emotion or affect for evaluating consumer reactions to services and products. http://www.kaaj. com/psych/scales/emotion.html, 2005. [11] A. Mehrabian. Incorporating emotions and personality in artificial intelligence software. http://www.kaaj.com/psych/ai.html, 2005. [12] G. Miller, R. Beckwith, C. Fellbaum, D. Gross, and K. Miller. Introduction to wordnet: An on-line lexical database. International Journal of Lexicograph, 3(4):235–244, 1990. [13] A. Ortony, G. L. Clore, and A. Collins, editors. The Cognitive Structure of Emotions. Cambridge University Press, Cambridge, UK, 1988.
