ESCADA: Experimental System for Character Affect Dynamics ...



Ph.D. General Examinations

Technical Area Exam (Erik T. Mueller, Examiner)

Xinyu Hugo Liu

Submitted September, 2004

Exam Topic

Design, implement, and give analysis for an original narrative comprehension system dealing in the realm of affect. Give a discussion of the theory of being put forth and of the grounding for such a theory. Also, discuss the strengths and limitations of the system and relate this to the various theories in the “Techniques for Narrative Comprehension with Imaginative Intelligence” literature (cf. Ph.D. General Examinations Proposal document).

ESCADA: Experimental System for

Character Affect Dynamics Analysis

Abstract

This paper presents ESCADA, a narrative comprehension system specialized to the analysis of how affect is interplayed between story characters. The fundamental metaphor which underlies this theory of affect dynamics is that of exchange. Affect-packages, represented in Mehrabian’s Pleasure-Arousal-Dominance dimensional model, are possessed by and exchange between characters and objects in the story world. One level of cognitive reflexivity is addressed by modeling each character’s “mental world.” A model of dominion and the exchange and possession of physical objects is also implemented. Part of the theory of affect dynamics developed in this paper is the role that objects play as containers of affective-energy; the affective-energy of an object affects the emotional state of the character who possesses it.

In this paper, first, we present motivation for our single-realm approach to story understanding; second we discuss theory of affect dynamics represented by this system with the assistance of pertinent examples indicating the scope of the system’s capabilities; third, we give a technical overview of the implementation; fourth, we present the results of some indicative studies evaluating the system’s performance; and fifth, we connect the ESCADA theory and system to the various theories and systems in the “Techniques for Narrative Comprehension with Imaginative Intelligence” literature (cf. Ph.D. General Examinations Proposal document).

Introduction and Motivation

Traditionally narrative comprehension systems follow either a path of representing the characters and the world in full detail, or of producing very general characterizations of the document. Following the path of full detail, all aspects of the world need to be modeled from story text, necessitating both an enormous amount of common sense knowledge properly formulated, and absolutely precise deep semantic parsing capabilities; unfortunately both prerequisite technologies are not and may not be available for quite some time. However, rather than backing off from character-level story understanding to document-level understanding, we take a possible middle approach which affords more specific understanding about characters and character interactions.

The idea is to understand in a medium amount of detail the world under the lens of a particular realm; that entails interpreting all the detailed events of the story by first projecting the detail into the target realm were some interpretation can be had. Of course, the success of a realm-specific understanding approach requires that the realm be pertinent to a good portion of all the story details.

The realm which is the focus of our project is affect, and this seems to be a particularly longitudinal realm, meaning that all aspects of language and cognition have some meaningful projection in the affect realm. David Gelernter explains the pervasiveness of affect in human experience by suggesting that the role of affect in cognition and memory is as an annotation mechanism which facilitates the indexing of memories and thoughts into broader themes, which he terms “affect linking” (Gelernter, 1994).

In order to project the consequences of story details into the affective realm, some knowledge or linguistic resource would be required. For the ESCADA system, we have mined and hand-annotated three invaluable affective-lexical resources: Roget’s Thesaurus of English Words and Phrases (Roget, 1911), the Affective Norms for English Words (ANEW) Project at the University of Florida (Bradley & Lang, 1999), and Beth Levin’s English Verb Classes and Alternations (1993). Coupled with a broad-coverage surface thematic role frame parser for English, these affective-lexical resources enable the ESCADA system to interpret directly from the language the affective consequences of actions, modifiers, and (through some further reasoning), objects. However, in light of the fact that such resources and mappings are only approximate and are not fool proof, such resources are interpreted by the system in a fail-soft (read: non-logical) way.

Without further ado, the following section presents a theory of character affect dynamics which underlies ESCADA.

A Theory of Character Affect Dynamics

Let’s first consider a simple children’s story and walk through it through the lens of the Character Affect Dynamics (CAD) Theory:

John was eating a delicious ice cream. Mary was secretly jealous of John because she wanted his ice cream. John didn’t know that Mary wanted his ice cream. Suddenly, Mary snatched the ice cream from John. John became very sad and resented Mary.

CAD advocates a squint-your-eyes approach to comprehension, where entities are either Agents (characters) or Objects (especially inanimate things). Agents and Objects can both be invested with, and divested out affective energies, be they positive or negative, and aggressive or submissive. If positive Objects are possessed, they imbue the possessor with positive energies, but if positive Objects are dispossessed, they bring woe to the dispossessor. If an Agonist (acting agent) acts on an Antagonist (usually called “patient” in the thematic role literature), in so doing he transfers an package of affective energy to the Antagonist. Whether and how the Antagonist reacts says something about the Antagonist’s affect-relationship with the Agonist, and about the Antagonist’s personality archetype in general. Each Agent also possesses a mental world, and how much an Agent utilizes his mental world and the discrepancies between his formed mental world and the actual world also help to betray his personality and the situation in general. Before we discuss the theory any further, let us walk through the above example using this theoretical framework.

Commentary appears indented

John was eating a delicious ice cream.

Delicious is positive; the IceCream Object is imbued with positive affect; “to eat” implies possession, and so the Agent John currently possesses the positive IceCream Object

Mary was secretly jealous of John because she wanted his ice cream.

The Agent Mary is experiencing a state of secret jealousy. Our affective lexicon tells us that jealousy is negative, aroused, and aggressive. The fact that the theme of jealousy is modulated by “secretly” means that Mary’s jealousy is a passive act against John; John does not necessarily know about Mary’s feelings toward him. Also, Mary wants to possess the ice cream. The causative link between the first clause and dependent clause are not currently addressed by our theory or implementation.

John didn’t know that Mary wanted his ice cream.

In John’s Mental World, imaginaryMary undesires the ice cream and does not want to possess it. At this point, the mismatch between the Real World and John’s Mental World is a source of story conflict, and our theory can predict that certain salient Real-Mental mismatches lead to conflict. Also, we can make some predictions about the characters’ personalities to explain why John has false knowledge; perhaps we can predict that John is oblivious or naïve or insensitive; we could conversely blame Mary for being calculating (since she has not directly communicated any affect to John).

Suddenly, Mary snatched the ice cream from John.

The word “suddenly” informs us this is an aroused event. Our affective lexicon says that “snatch” is a possession-verb and is imbued with high aggression, high arousal and slightly negative valence (thieves snatch). The act of Mary snatching the ice cream from John means that John is dispossessed of IceCream and Mary now possesses IceCream. Mary benefits from the positive affect energy of IceCream (minus whatever negative energy was imbued upon IceCream in its participation in the negative “snatching” event; every event affects every participating agent or object). John is discouraged by the loss of a Positive Object.

John became very sad and resented Mary.

From Mary’s previous act of aggression against John, John’s reaction is telltale of the element of aggression in his personality. However, instead of retaliating with something aggressive, John is rather submissive; sadness is submissive (in Mehrabian’s PAD model), and so is resentment. The fact that John resents Mary is a purely passive and private act which only affects John’s mental attitude toward Mary and does not transfer any affective energies to Mary.

The type of spatial-metaphorical entity-imbuement and entity-interaction analysis employed in CAD theory has precedence in the Artificial Intelligence literature. Jackendoff’s Trajectory Space (1983) focuses on semantic mappings of text unto spatial representations of objects moving along paths from sources to destinations. Borchardt’s Transition Space (1990) represents a story event sequence by tracking the differential change in the properties of objects. Perhaps the closest work in the literature, however, is Leonard Talmy’s force dynamics (1988).

The conceptual primitives in force dynamics theory are entities, and the typical scenario calls for two entities. Each entity exerts a force on the other and the sum of the forces yields in some consequent. For example, if the forces are equally matched, the result may be stalemate. CAD can be thought of in terms of interactionary forces as well. We can think of a steadfast character as a low-inertia entity, an aggressive character as one who exerts a strong force, and so on and so forth. A further similarity between Force Dynamics and CAD is that both take a cognitive linguistics approach; both argue that inherent in basic language constructions and in the lexicon are the needed semantics. Force dynamics uses syntactical elements (e.g. against, despite, modals) and the semantics of verbs such as “resisted,” “refrained” to animate an underlying spatial representation. CAD utilizes the affective connotations of words, and the spatial connotations of syntax (e.g. transitivity, intransitivity, reflexivity, copula-constructions) and of verb-semantics (e.g. passive versus active verbs) to animate an underlying simulation of the story world.

As we have argued before, we believe there to be great benefits to approaching story understanding in a single-realm, especially when surface language can be used to directly map story events into the underlying representation. Greater granularity can be achieved than document-level characterization of text, without requiring deep semantic parsing and not needing an enormous corpus of common sense knowledge to aid in inferences. All the resources needed for this approach can be found in existing cognitive lexicon projects such as those which were mentioned in the previous section.

Like Talmy’s force dynamics theory, the CAD framework takes a cognitive linguistics approach to language understanding, and to be sure, the understanding which is achieved here is only approximate, heuristic, and connotative – concerned with the affective gist of each linguistic transition – and the style of semantic interpretation is also different from previous approaches. Whereas traditional thematic role frame interpreters and frame-semantics a la FrameNet maintains a fairly diverse and expansive case base of types of situational semantic frames which can be parsed into depending upon the verb-argument structure of a sentence, CAD calls for relatively few types of cases, distinguishing only between dominion transactions (e.g. John gave Mary the kite), affect declarations (e.g. Mary was in a poor mood), passive affective transactions (e.g. John resented Mary), active affective transactions (e.g. Mary stabbed John and stole his ice cream), and contagion transactions (e.g. The delicious ice cream; Mary’s wicked heart).

In the following subsections, we first discuss the representations of affect and personality used in CAD theory, tying these to models of personality from the literature; next, we discuss further grounding for the model in psychology.

Representing Affect. In choosing a representation of affect, we were careful to consider a model that could be generic enough to be attributed to both Agents and Objects, and sensitive enough so that even nuanceful affects could be captured elegantly. Immediately, discrete ontological models such as Paul Ekman’s emotion ontology (1993) of Happy, Sad, Angry, Fearful, Disgusted, Surprised, derived from the study of universal facial expressions. There is a sense that what we want to capture is not emotion, but affect, including states which are subtle or have no linguistic label. It was for this very reason of escaping the discourse of words which motivated Picard to rename the discourse of emotions to that of affect in light of computation (1997).

Our choice was the dimensional PAD model of affect proposed by Albert Mehrabian (1995), which specifies three almost orthogonal dimensions of Pleasure (vs. Displeasure), Arousal (vs. Nonarousal), and Dominance (vs. Submissiveness). In this paper and in our implemented system, the notation we will use is e.g. (P0.25 A0.5 D0.2). P0.0 is displeasure, P1.0 is pleasure; A0.0 is nonarousal, A1.0 is arousal; and D0.0 is submissiveness, D1.0 is dominance.

The PAD model is a unification model into which most of the other models of affect can be mapped. For example, (P0.25 A0.5 D0.25) might correspond to sadness, while (P0.25 A0.7 D0.8) would correspond to anger. The only exception are models which include directionality of affect. For example, “resentment” is an inwardly kept affect whose corresponding outwardly directed affect is “anger.” In CAD, directionality is explicitly represented, with inwardly affect as those which reside in the Agonist’s Mental World, and outwardly affect as those which are transferred as affective-energy packages between Agonist and Antagonist. Interestingly, in our review of existing emotion models, there is yet another dimension in many accounts of human sentiment which to our knowledge is not included in any dimensional model; this is affective focus. Vulgarity is usually unfocused, whereas resentment is usually focused. Incidentally, affective focus can be handled by our CAD theory and system implementation, but its discussion is beyond the scope of this paper.

Representing Personality. In representing personality in the CAD theory, the primary consideration is to choose a representation which most intuitively leverages the scaffolding established by entity-imbuement and energy-transfer and the PAD affect model. If personality is to arise intuitively from the CAD World model, then there should be a complete analogical mapping between the conceptual feature space of the World, and the space of personality.

From the personality literature, we were greatly inspired by three sources: Jung’s theory of psychological types (trans. 1971), which was augmented and re-presented as the familiar Myers-Briggs Type Indicator (MBTI) personality test (Briggs-Myers & McCaulley, 1985); Carol Pearson’s six mythic archetypes (1998); and Karen Horney’s theory of neurosis (1945), which articulates three archetypical behaviors of Compliance, Aggression, and Withdrawal.

The Jungian MBTI postulates four binary dimensions of temperament: Introversion-Extroversion, Intuition-Sensing, Thinking-Feeling, and Judging-Perceiving (which was added in Myers-Briggs). We can relate MBTI to our CAD model as follows. The first dimension describes the directionality of a person’s psychic energies; in CAD, a character who commits passive acts and prefers to hold inwardly directed attitudes may correlate to Introversion, while a character who expresses his inner attitude through extrinsic and social acts is Extroverted. The second dimension describes the character’s chief mode of information from the world. An Intuitive character is better able to guess (intuit) the affective states of others, even when they are inwardly kept. For example, if Mary is secretly sad, but John goes to console her, he is probably Intuiting. Sensing character act primarily on what is superliminally presented to them in explicit fashion. The third dimension is Thinking-Feeling; a Thinker is likely to show more cognitive planning activity and to trust his Mental World model, while a Feeler’s decisions are dictated more by his felt affect. The fourth dimension (not originally included in Jung’s theory of psychological types) is Judging-Perceiving and is more tenuous to computationalize in CAD because it requires the handling of more advanced structures like causatives and hypotheticals, but the gist is that a Judging character is likely to trust and follow the plan laid out in his Mental World more than a Perceiver, who is more likely to revise his conceptions and explore the space of alternatives.

Pearson’s six archetypes theory, a Jungian theory, is premised on the idea that there exists a set of mythic roles which are engrained into the unconscious of all humans past and present; these mythoi are profound and universal. Pearson’s archetypes complement rather than compete with psychological type theories like MBTI because whereas MBTI describes general patterns of people’s temperament, Pearson’s archetypes can be adopted semi-volitionally by people. The six archetypes are Orphan, Wanderer, Warrior, Altruist, Innocent, and Magician. A person may go through all of these in various situations and at various stages in her life. These archetypal roles are also useful ways to understand the affect dynamics of characters in a story. For example, a character who retaliates against an affect attack or fights hard to improve his own mood may be described as a Warrior; a character who gives positive affective energy to others in need may be described as an Altruist; and finally, a character who is carefree yet easily hurt or devastated might be an Innocent. At second glance, Pearson’s archetypes are quite different from MBTI because whereas the MBTI model describes an egocentric phenotype, Pearson’s archetypes describe how an Agonist behaves in the face of an Antagonist, and thus, it is more of an interactionist theory.

Horney’s theory of neurosis brings to light three further personality phenotypes not addressed in MBTI. Horney postulates three broad classes of coping strategies which people use to address their neurotic needs. The first class is Compliance and this includes the phlegmatic personality (sluggish), which in CAD manifests as an unemotional character. The second class is Aggression and describes a character who is Dominant more than Submissive. The third class is Withdrawal and corresponds to a character who avoids conflict rather than battling it. It is interesting the Horney also uses a vocabulary of motion to describe these coping strategies: Compliance, Aggression and Withdrawal are moving-toward, moving-against, and moving-away-from, respectively.

In the above presented personality theories, we’ve articulated possible mappings into the CAD framework which we believe to be intuitive, though admittedly these mappings are also heuristic. Luckily, the science of Personality is also heuristic. While these mappings illustrate that the CAD framework would be versatile enough to accommodate all of these theories, we have selected out aspects of each theory to construct a new model that would resonate best and can be most readily computed from the CAD framework. Below are the proposed dimensions of personality for the CAD.

Inwardly-Outwardly: Similar to Jung’s Introversion-Extroversion and Horney’s Withdrawn-NotWithdrawn. Inwardly-Outwardly is heuristically measured by the ratio of a character’s tendencies to be passive about affect, harboring it within, to the tendencies to be active and extrinsic in communicating affect.

Thinking-Feeling: Straight from Jungian’s psychology type theory. A Thinker is a character who engages in a lot of mental activity, and involves cognitive plans in preparation for his affective acts. A Feeler is a character who focuses primarily on feeling.

Unperceptive-Perceptive: Same as Jungian Sensing-Intuiting. A character is perceptive if he is able to act in response to an Antagonist’s inwardly feelings.

Uncontagiable-Contagiable: How easily moved (in the same direction) a character is by her contact with received affective-energies, and with the possession of affectively-charged objects. Corresponds roughly to Pearson’s NonInnocent- Innocent and MBTI’s Judging-Perceiving.

Submissive-Aggressive: Does a character usually submit to an Antagonist, or does he aggress, initiate attack, defend, and retaliate? Does he indulge more in submissive affects (e.g. sadness, fear) or aggressive affects (e.g. angry)? This corresponds roughly to Pearson’s NonWarrior-Warrior, Horney’s Aggression-NonAggression, and Mehrabian’s Submissiveness-Dominance.

Good-Evil: This notion may be too histrionic and draconian for real life but it is entirely appropriate for story understanding. We can qualify all acts and reacts in CAD as good or evil. For example, doing something bad to someone good is evil, and taking away something bad from someone good is good. Also deceptive behavior, such as acting not in accordance with one’s inwardly held feelings, is evil. Corresponds roughly to Pearson’s Altruist-NonAltruist.

Unelastic-Elastic: Elasticity is proportional to the recoil time of a character in light of distress, measured in narrative time from the time that a character is inflicted with negative affect, to the character’s recovery back to neutral or to positive affect. It does not directly map into any of the three personality theories, although it maps into archetype of obsession.

Admittedly these dimensions are not completely orthogonal, but they are all directly available and computable from a computer simulation of the CAD framework. Nor are these articulated dimensions an exhaustive list of all the personality traits which can be read off the CAD framework; however, they serve as a good point of departure. Having described the representations of affect and personality in CAD, the following subsection discusses the psychological grounding for the theory.

Grounding the CAD theory. We have already discussed the precedence in the Cognitive Linguistics literature for an agent-object-transfer model of affect dynamics; however, we have not yet touched upon the psychological plausibility of conceiving of affect as energy-packages which can be transferred or held, and conceiving of objects as psychic holders of energy which can imbue affect unto their owners; furthermore, we have not discussed the validity of using an affective lexicon and syntactic construction to populate events the affect realm.

In Metaphors We Live By (1980), Lakoff and Johnson put forth a theory that all language and cognition is structured by basic metaphors of containment, space, orientation, and movement. They present linguistic evidence that we conceptualize of people as CONTAINERS and interpersonal communication as flowing along CONDUITS. In regard to emotions, they give the example, “I am in love” to say that some emotions are containers. A person can be inside love. We would say that the “emotions as containers” metaphor only governs a few named and mythed states; love, long romanticized as an institution of the human condition, is appropriate as a public container; every person under the spell of love is in that state. However, one would not say “I am in hate,” or “I am in loathe.” Mostly the “emotions as states” metaphor is valid for positively valenced emotions. There is also a cultural and social bias at play. Love, friendship and other positive affects are public and shared because it is socially couth to do so, but not negative emotions. For example, you can “give her my love” and “offer her my friendship” but not “give her my hate” or “offer her my resentment.” However, we are quick to point out that this is most applicable for named emotions, not necessarily for unnamed affective states. Affective states, because they are not articulated as emotions, remain my more personal, and we describe them as attributes or entities contained within the self. “I have this fear in me;” “There is a feeling in my heart;” “I am stricken by grief.” These examples demonstrate that it is common for people to conceive of themselves as containers, and to describe affect as that with which they are infected or imbued.

In describing communication, Lakoff and Johnson suggest that the CONDUIT metaphor is the most prominent. I have something I wish to communicate and it is packaged up as an entity, and sent to you, and you are to unpack it. If you are oblivious, you will not see the package coming and will miss it. If you are dense, you will not be able to unpack it properly. “He was giving her a hint.” Because the conduit metaphor underlies communication, we would speculate that in part, every directed communication act can be conceptualized as a conduiting of something, regardless of whether or not that metaphor is actively at play in structuring the linguistic utterance; it will nevertheless structure our deep subconscious conceptualization of each act.

The power of the CONTAINER and CONDUIT metaphors in narratives may be even greater than in real life. Stories are a world governed by imagination and the abstract, and thusly we speculate that CONTAINER and CONDUIT being devices friendly to abstraction and mental constructivism, are devices which the reader utilizes to construct the story mentally. Each character is constructed using our human faculties for Theory of Mind, a concept from the Cognitive Science literature, using rational tools such as Dan Dennett’s Intentional Stance (1987); on the flipside, the aesthetic profile of the story also constructs in the reader’s unconscious another dimension of understanding for the characters. If indeed reading is an act of constructing characters and populating their states and psyches, then perhaps it follows that the reader will want to extract every bit of information from the text, paying attention to the nuance of words. Verbs, it so happens, contains a lot of nuanceful cues and implications along different dimensions, and Cognitive Linguists know this well. Marvin Minsky also knows this, and he proposed in Society Of Mind (1986), using Roger Schank’s conceptual dependency transframe construction, that a linguistic act like “John gave Mary a kite” should be understood as a collection of transframes, one residing in each realm: affective, physical, dominion, etc. The implication of this is that every act can be projected into the interpretive spaces of a variety of realms. It is then consistent with Minsky and Lakoff and Johnson’s theories that each act is in part interpreted by the reader as a conduiting of an affective package from Agonist to Antagonist, consciously or unconsciously.

Ortony, Clore and Collins’s theory of emotions (1988) argues that emotions result from cognitive appraisal, though to be sure, this appraisal is not necessarily conscious or attentive. They postulate that emotions only make sense when directed at an appropriate focus, and they name three types: events, agents, and objects. This supports CAD theory’s conceptualization of an affect as associated with the self, with objects, or with other agents via the self’s Mental World construction. This begs the question, “But what of those emotions regarding events? There is no event construction in CAD theory!”

This is a valid criticism, because the CAD is a simplification theory meant to achieve a medium level of understanding about characters, but not full story understanding. Certain types of events are represented in CAD, for example, events between two characters. “John hurt Mary” is an event, and is represented by Mary’s receipt of negative affect-energy from John. However, in the formulation of the CAD framework, it is not possible for John to attribute his response to the event, unless the event is named. Because if the event is named, the event is treated as an object, and objects can be imbued with affect.

However, because CAD is not meant to handle the complexities of events, it cannot relate any event to character’s goals, as more complete story understanding demands. Schank and Abelson, for example, interpret events based on whether or not they facilitate goals such as Achievement, Entertainment, Instrumental, and Crisis (1977). As a simplification framework, CAD does not keep track of character goals, with one example: possession. CAD has a dominion model which tracks who possesses what objects, and how affectively valuable an object is. For example, if a story reported that “Mary wanted a kite,” then the Kite Object is invested with positive affect, because it is wanted. Then if the story reports, “John bought a kite for Mary,” that event is deemed a transaction of positive affect to Mary because the Kite Object has positive valence. This can be viewed as the satisfaction of an Achievement goal, but the reality of how CAD represents this is not as a goal at all. Another way that events are handled in CAD is that if an event is named, e.g. “the accident,” then suddenly it can be represented as an Object and imbued with affect. With some adjustment of the CAD framework, a character’s involvement with an event can be represented as a character “owning” an Event Object. This may be a natural enough way to handle events, especially traumatic ones, where there is a sense that a character simply “can’t rid himself from its memories;” he is unable to dispossess that event!

Another source of grounding for CAD theory is possible if we are able to switch from a Western cultural view to an Eastern one. Western culture does not “see” affect as anything concrete, but Eastern culture does. In the Eastern concept of “Chi,” which means breath and energy, objects and people are thought to possess energy, and it is thought that psychic energies can be transferred from agent to agent and object to agent. Such is the concept with underlies ancient psychologies like “Feng-Shui” (wind-water) and Jungian psychoanalysis, which views the human psyche as a container of energy, and the environment as a field of objects whose energies influence the psyche (cf. Symbols of Transformation, 1912) As Jung’s psychoanalysis has penetrated Western culture, we see some examples of the belief that affective-energies are kept and conveyed in interpersonal communication, e.g. “I’m getting some bad vibes from her.” The belief that objects are imbued with affective-energies and that possessing objects leads to the inheritance of their energies has been brought closer to the mainstream and to science by Csikszentmihalyi and Rochberg-Halton who developed a theory of possession around this in The Meaning of Things (1981).

Having completed our current discussion of the theory underlying Character Affect Dynamics, the next section discusses a system implementation of the theory.

The ESCADA System

ESCADA is an Experimental System for Character Affect Dynamics Analysis. It implements a good portion of the CAD theory presented above.

The system is implemented as 1,400 lines of Python source code, built on top of the MontyLingua NLP engine (Liu, 2004a) which is itself 7,000 lines of Python source. A 500KB lexical knowledge base was compiled for this project, from three sources: Roget’s Thesaurus of English Words and Phrases (Roget, 1911), the Affective Norms for English Words (ANEW) Project at the University of Florida (Bradley & Lang, 1999), and Beth Levin’s English Verb Classes and Alternations (1993). In addition, we hand-annotated the lexical classes present in these sources using PAD as the annotation structure, and the annotation file is 8KB.

The system is input a narrative text; that text is segmented into sentences, a deictic stack is implemented to track characters and objects and resolve anaphora. Text is tokenized, part-of-speech tagged, lemmatized, chunked, linked, and extracted into syntactic frames of the form: Verb-Subject-Object-Object* (VSOO). The semantic interpretation from these syntactic frames are produced by a number of competing understanding demons, who opportunistically recognize different types of affect-conveyance and affect-imbuement situations and compete for their interpretations to be accepted by a demon manager, who alternately allows and inhibits different classes of demons, such as cognitive-reflexive demons, conveyance demons, dominion demons, and context demons. The architecture for this is inspired by Oliver Selfridge’s Pandemonium feature detection system (1958), although admittedly the demons in ESCADA do not currently adapt themselves over experience, and coordination is rudimentary in the demon manager; however we anticipate that as the number of recognition demons increases, the system can scale under this architecture.

Each demon is capable of recognizing a different syntacto-semantic sentential case. The currently implemented demons are given below, with names expanded for clarity, accompanied with an explanation of their purpose and nuances, and the lambda expression they produce which when applied to the world, updates it.

AGONIST EXPERIENCING MOOD (e.g. John felt depressed)

Characterizes an agonist’s current affective self-concept.

(λworld: (world.find_agent(AGONIST).set_feel_about(SELF,PAD))

AGONIST PASSIVE-ACT ANTAGONIST (e.g. John sulked; Mary resented him, John secretly cursed Mary)

Passive acts are those which form the agonist’s mental attitudes, but are not overtly communicated to the antagonist. Iff the antagonist is Perceptive, she may intuit it.

(λworld: (world.find_agent(AGONIST).set_feel_about(ANTAGONIST|SELF|OBJECT,PAD))

AGONIST ACTIVE-ACT ANTAGONIST (e.g. John cursed Mary; Mary shouted obscenities at John)

Act or extrinsic acts are those which get overtly communicated as affective-energy packages from the Agonist to Antagonist. The primary distinguishment of passive-versus active acts are made by examining the theme (verb). Roget and Levin’s verb classes both distinguish between mental, inwardly acts and social, extrinsic, public acts. In addition, modifiers such as adverbials which convey stealth are used as cues.

(λworld: (world.find_agent(AGONIST).feel_at(ANTAGONIST,PAD)) ||

(λworld: (world.find_agent(READER).set_feel_about(AGONIST,PAD))

AGONIST MENTAL-ACT CLAUSE (e.g. John knew that …)

Mental acts are those which populate the Agonist’s Mental World. The setup of the Mental World is the same as the Physical World, except that an Agonist may hold a different version in its head. The CLAUSE can contain any utterance parseable by another demon (e.g. John knew that Mary resented him).

(λworld: (world.find_agent(AGONIST).mental_world.(…))

AGONIST IMBUMENT (e.g. John had a horrible time sleeping; John life was one rollercoaster ride after another)

This demon is an extremely general construction. It allows any sentence construction because it does not consider syntax. The idea here is of affect as a contagion. Any description whose affect can be calculated is then imbued unto the agonist. Of course, the system is sensitive to negations, which is handled by a theme-inversion mechanism. The demon also tries to avoid confusion by refusing to infect if more than one agent is present.

(λworld: (world.find_agent(AGONIST).set_feel_about(SELF,PAD))

OBJECT IMBUMENT (e.g. Mary had a delicious ice cream; The accident was particularly traumatic for John)

This demon imbues objects rather than agonists, is also general, but works with the same confusion-avoiding caveat mechanisms. Note here that an “accident” event can be treated as an object particularly because it is a named entity.

(λworld: (world.find_object(OBJECT).imbue_value(PAD))

OBJECT DOMINION TRANSFER (e.g. Mary had a delicious ice cream, and then John swiped it from her.)

An object can be possessed by an Agonist, or dispossessed. Dispossession can be by an Antagonist (e.g. John stole Mary’s heart), or can be existential (e.g. The ice cream melted).

(λworld: (world.find_object(OBJECT).possessed|dispossessed_by(AGONIST))

STORY CONTEXT (e.g. There was a stench in the air)

This demon attends to the affect in every sentence, including those of no pertinence to any character or object. This is used to imbue the story World itself with a sequence of affects.

(λworld: (world.imbue_value(PAD))

The demon manager inhibits or accepts the interpretations of the demons, and accepted interpretations are applied to the simulation World, having the effect of continuously updating the world at each story step. The simulation World is inhabited only by Agents and Objects. Agents contain the following data and capabilities:

SIMULATION AGENT:

history of attitudes

current attitude

possessions

feelings about alters

mental world – same as physical world, except its agents have no mental world

agents

objects

feel-about(pad)

mental-feel-about(pad)

mental-feel-at(pad)

notified-of-incoming-affect(sender,pad)

possess(object) / dispossess(object)

pretty-print()

clone_agent() – this is used to produce an imagination scenario

Objects contain the following data and capabilities:

SIMULATION AGENT:

history of imbued values

owner

possessed-by(new-owner) / dispossessed-by(old-owner)

current-value()

To better illustrate the workings of the implementation, we walk through a trace of a story. The following story is a paraphrase and adaptation from the first-grade children’s story, “Up and Away,” published by Houghton Mifflin (McKee et al., 1966). It was the original story which Charniak used verbatim in his 1972 story comprehension program (Charniak, 1972). The paraphrase eliminates character first-person utterances, which is out of the scope of the implementation; significantly simplifies the rhetorical structure of the text around the conventions of declarative form (although aspectuals, moods, tenses, and dependent clauses are still expressed); and allows the narrator to interject some subjective commentary into the recounting of the details; the ending was also changed.

First, here is the story in full:

Jack had some gorgeous paints given to him by his dad. By comparison, Janet owned some lousy dirty pencils. Janet was jealous. She planned to underhandedly swindle Jack. So she lied to him. Jack was led to believe that the paints made his pictures look stupid. He was also made to think that the pencils were more pleasing and could draw more splendid pictures. Janet offered to take the paints from Jack. Jack received the pencils from Janet. Jack was very happy. Jack thanked Janet. Janet was regretful and ashamed.

Next, we will walk through the sentences of the story and present the lambda interpretations derived from each (with slight float rounding and pretty printing). We do not present a dump of the world after each sentence (only once at the end) but we encourage the reader to obtain the freely available source to this project and run the given example story in the system for personal verification.

System output and commentaries are indented to here. Recall that an affect is given in the form [p, a, d], where 0.5 is the neutral fulcrum value.

Jack had some gorgeous paints given to him by his dad.

lambda world: world.find_object("paints")

.imbue_value([0.55, 0.53, 0.72])

lambda world: world.find_object("paints").possessed_by("Jack")

lambda world: world.imbue_mood([0.55, 0.53, 0.72])

Positive energy imbued onto paints, possessed by Jack. At this point, if the system was asked about Jack’s feelings, the heuristic rule of “Affective energy of possessions rub off on possessor” could make some prediction that Jack is feeling positive.

By comparison, Janet owned some lousy dirty pencils.

lambda world: world.find_agent("Janet")

.set_feel_about('self',[0.40, 0.48, 0.59])

lambda world: world.find_object("pencils")

.imbue_value([0.40, 0.48, 0.59])

lambda world: world.find_object("pencils").possessed_by("Janet")

lambda world: world.imbue_mood([0.40, 0.48, 0.59])

Similarly, pencils are imbued with negative energy, and the owner is Janet. In this case, the contextual contagion demon caused the first lambda interpretation which purports that Janet feels negatively toward herself because she possesses something “lousy” and “dirty.”

Janet was jealous.

lambda world: world.find_agent("Janet")

.set_feel_about('self',[0.3, 0.8, None])

lambda world: world.imbue_mood([0.3, 0.8, None])

This is a direct characterization of Janet’s affective state. The value of the PAD’s dominance dimension is “None,” because there is no information as to its value. When the Janet simulation Agent code object, a history of her states is being kept, allowing us to questions about changes or developments in her state as the story progresses. Similarly, the affective history of the world’s mood is being tracked, as a record of the affective dynamics of the development of the narrative.

She planned to underhandedly swindle Jack.

lambda world: world.find_agent("janet")

.set_feel_about('jack',[0.2, None, None])

lambda world: world.imbue_mood([0.3, None, 0.8])

The “swindle” theme has a definite affective projection: it is the transfer of negative affective energy. However, because theme is preceded by “plan to” and because “underhandedly” modifies the theme, the interpretation is that this ordinarily extrinsic act is in this case passive. Thus, it only affects Janet’s mental affective attitude toward Jack. Breaking at this point, some further story understanding feats could be had. Even though ESCADA is overly reductive in its interpretation of the situation, not knowing about Janet’s new goal, if the system were asked to explain why Janet was negative toward Jack, jealousy could be reconstructed as a possible motivation, since Jack owns positive affect and Janet owns negative affect. The system could also predict depending on what is known currently about Janet’s personality type, whether she is likely to perpetrate an overt negative act against Jack. Actually, depending on her later action, the system could learn about Janet’s personality type since it is possible to create an explanation case base associating various personality types with the type of reaction they are likely to take in a given affective situation. For the most part, the current implementation does not support such fine-grained abductive explanation for character behaviors which could betray the character personality types, but this would not be a difficult augmentation.

So she lied to him. Jack was led to believe that the paints made his pictures look stupid. He was also made to think that the pencils were more pleasing and could draw more splendid pictures.

lambda world: world.imbue_mood([0.34, 0.62, 0.54])

lambda world: world.find_agent("Jack")

.set_feel_about('self',[0.47, 0.43, 0.51])

lambda world: world.find_agent("Jack")

.mental_world.find_object("paints")

.imbue_value([0.39, 0.43, 0.42])

lambda world: world.find_agent("Jack")

.mental_world.imbue_mood([0.39, 0.43, 0.42])

lambda world: world.imbue_mood([0.47, 0.43, 0.51])

lambda world: world.find_agent("Jack")

.set_feel_about('self',[0.59, 0.48, 0.69])

lambda world: world.find_agent("Jack")

.mental_world.find_object("pencils")

.imbue_value([0.65, 0.48, 0.66])

lambda world: world.find_agent("Jack")

.mental_world.imbue_mood([0.65, 0.48, 0.66])

lambda world: world.imbue_mood([0.59, 0.48, 0.69])

The theme of lying could not resolve to a substantive consequence in the affective realm, so only global mood was affected. In the second sentence, paints were devalued in Jack’s mental world, but not in the real world. In the third sentence, those shabby pencils appreciated in value in his mental world, but not in the real one. Here, even in the affective realm, we can see that plot tension is created. Jack’s mental world is opposed to what exists in the real world. Although the simplicity of the representation does not allow us to associate Janet with causing this, it could be guessed from Janet’s negative attitude toward Jack – that is sufficient motive. With just a few heuristics (or that personality archetypes explanation base), the system could also presumably understand that Jack is gullible or naïve or unperceptive for having such distorted perceptions. You’ll also notice some lambda interpretation which affect Jack’s mood as well -- this is simply the contextual contagion keeping consistent and fluent Jack’s mental world and his affective state.

Janet offered to take the paints from Jack. Jack received the pencils from Janet.

lambda world: world.find_agent("Janet")

.set_feel_about('self',[0.54, 0.59, 0.67])

lambda world: world.find_object("paints")

.imbue_value([0.54, 0.59, 0.67])

lambda world: world.find_object("paints").possessed_by("Janet")

lambda world: world.imbue_mood([0.54, 0.59, 0.67])

lambda world: world.find_agent("Jack")

.set_feel_about('self',[0.59, 0.27, 0.64])

lambda world: world.find_object("pencils")

.imbue_value([0.59, 0.27, 0.64])

lambda world: world.find_object("pencils").possessed_by("Jack")

lambda world: world.imbue_mood([0.59, 0.27, 0.64])

The theme “offered to take” is given here to illustrate the understanding limitations of the current approach. Of course the semantics of an “offer” situation might require an acceptance to be issued before the dominion transfer occurs, but the current theory is perhaps overly naïve to these details. In these two sentences, Jack and Janet swap the paints and the pencils. There are perhaps some other lambda interpretations which the reader disagrees with; those are the result of contextual contagion and imbuement demons. The fact that “offering” as a theme reflects positively on the offerer causes Janet to be imbued with positive mood for “offering” even though it was done under false pretexts (she is tricking Jack). Overall though, we are afforded some more story understanding lucidity. The fact that Jack lost something good due to the misinformation of his own mental world demonstrates his ignorance. Janet’s willingness to proceed with the transaction despite her lacking the excuse of holding false beliefs betrays her deception.

Jack was very happy. Jack thanked Janet.

lambda world: world.find_agent("Jack")

.set_feel_about('self',[0.7, 0.69, 0.73])

lambda world: world.find_agent("jack")

.set_feel_about('self',[0.9, 0.69, 0.7])

lambda world: world.imbue_mood([0.7, 0.69, 0.73])

lambda world: world.find_agent("jack")

.feel_at('janet',[0.8, None, None])

lambda world: world.imbue_mood([0.55, None, 0.73])

Here, the “thank” theme is interpreted as an extrinsic event; thus Janet explicitly receives an affective package from Jack. As with the previous two sentences, Jack’s unperceptiveness is revealed. Jack’s happiness can lead us to abductively explain that Jack’s perceptions of the affective energies of the paints and the pencils was based on his own mental world’s assessment, not based on the reality the narrator has painted. The continuing violation of the system heuristic that “Affective energy of possessions (based on real-world valuations) rub off on possessor” could constitute an anomaly, as Ram used the concept in the AQUA question-driven story understanding system to motivate further understanding and explanation-making (Ram, 1994).

Janet was regretful and ashamed.

lambda world: world.find_agent("Janet")

.set_feel_about('self',[0.21, 0.59, 0.3])

lambda world: world.find_agent("janet")

.set_feel_about('self',[0.16, 0.59, 0.3])

lambda world: world.imbue_mood([0.21, 0.59, 0.3])

Janet’s negative self-concept at this story moment contains so much information when correlated with the history of the story. There are many “critical understanding agents” which could be implemented to draw various sorts of conclusions from the history of the character affect dynamics of this story; for example, a critic should note that Janet went from dominant to submissive, while Jack, despite Janet’s negativity toward him, and despite his ignorance of her deception, began happy and ended happy. If the critic could look for these types of patterns, it could report its results in the discourse of something like Lehnert’s plot units (Lehnert, 1984); this story might be something like an “Ineffective Aggression” affective plot unit or something along those lines. There are a great many potential simple and complex plot units to be had, given that the CAD theory furnishes +, - valences, “m” mental states, events (extrinsic and passive), and goal achievement or failure might possibly be inferable from the shifts in dominance posture (dominant ( submissive might signal goal failure). Other goals, such as the attainment of positive affect, are implicit in the CAD framework.

Finally, a pretty-print dump of the world’s state at the end of the story is shown below:

####### THE REAL WORLD: ###############

WORLD TIME STEP: 0

OVERALL MOOD: [P0.46 A0.55 D0.63]

MOOD PROGRESSION: [P0.55 A0.53 D0.72]; [P0.4 A0.48 D0.59]; [P0.3 A0.8 D?]; [P0.3 A? D0.8]; [P0.34 A0.62 D0.54]; [P0.47 A0.43 D0.51]; [P0.59 A0.48 D0.69]; [P0.54 A0.59 D0.67]; [P0.59 A0.27 D0.64]; [P0.7 A0.69 D0.73]; [P0.55 A? D0.73]; [P0.21 A0.59 D0.3]

AGENTS:

CHARACTER: jack

FEELS: [P0.9 A0.69 D0.7]

FEELING HISTORY: [P0.47 A0.43 D0.51]; [P0.59 A0.48 D0.69]; [P0.59 A0.27 D0.64]; [P0.7 A0.69 D0.73]; [P0.9 A0.69 D0.7]

FEELS [P0.8 A? D?] ABOUT janet

HISTORY:

SENT janet [P0.8 A? D?]

OWNS:

pencils WORTH [P0.5 A0.38 D0.62]

POSSESSION HISTORY:

POSSESSED paints WORTH [P0.55 A0.53 D0.72],

DISPOSSESSED paints WORTH [P0.55 A0.56 D0.7],

POSSESSED pencils WORTH [P0.5 A0.38 D0.62]

oOoooOoo JACK's MENTAL WORLD: oOoooOoo

WORLD TIME STEP: 0

OVERALL MOOD: [P0.52 A0.45 D0.54]

MOOD PROGRESSION: [P0.39 A0.43 D0.42]; [P0.65 A0.48 D0.66]

MENTAL AGENTS:

IMAGINED OBJECTS:

IMAGINED OBJECT: paints

OWNED BY:

IMBUED VALUES: [P0.39 A0.43 D0.42]

IMAGINED OBJECT: pencils

OWNED BY:

IMBUED VALUES: [P0.65 A0.48 D0.66]

oOoooOooOOooOOoOoOOOoOOoOoOOooOoooOoo

CHARACTER: janet

FEELS: [P0.16 A0.59 D0.3]

FEELING HISTORY: [P0.4 A0.48 D0.59]; [P0.3 A0.8 D?]; [P0.54 A0.59 D0.67]; [P0.21 A0.59 D0.3]; [P0.16 A0.59 D0.3]

FEELS [P0.2 A? D?] ABOUT jack

HISTORY:

RECEIVED [P0.8 A? D?] FROM jack

OWNS:

paints WORTH [P0.55 A0.56 D0.7]

POSSESSION HISTORY:

POSSESSED pencils WORTH [P0.4 A0.48 D0.59],

POSSESSED paints WORTH [P0.55 A0.56 D0.7],

DISPOSSESSED pencils WORTH [P0.5 A0.38 D0.62]

oOoooOoo JANET's MENTAL WORLD: oOoooOoo

WORLD TIME STEP: 0

OVERALL MOOD: [P? A? D?]

MOOD PROGRESSION:

MENTAL AGENTS:

IMAGINED OBJECTS:

oOoooOooOOooOOoOoOOOoOOoOoOOooOoooOoo

OBJECTS:

OBJECT: paints

OWNED BY: janet

IMBUED VALUES: [P0.55 A0.53 D0.72]; [P0.54 A0.59 D0.67]

OBJECT: pencils

OWNED BY: jack

IMBUED VALUES: [P0.4 A0.48 D0.59]; [P0.59 A0.27 D0.64]

#######################################

Also, a most meager personality gister has thus far been implemented, and it is extremely barebones compared to what is possible had the personality archetypes case base (had it existed it would be implemented like Schankian explanation patterns (1986); patterns could be indexed by the personality archetypes which they fit). However, for completeness, the gisted character analysis is given below.

OVERALL STORY MOOD: [0.46166666666666667, 0.54800000000000004, 0.62909090909090915]

STAT REPORT FOR: jack

AVG EGO MOOD: [P0.65 A0.51 D0.65]

AVG ALTERS MOOD: [P0.8 A? D?]

AVG INCOMING: [P? A? D?]

AVG OUTGOING: [P0.8 A? D?]

MENTAL ACTIVITY INDEX: 2

INTRO/EXTROVERSION RATIO: 1.0

STAT REPORT FOR: janet

AVG EGO MOOD: [P0.32 A0.61 D0.47]

AVG ALTERS MOOD: [P0.2 A? D?]

AVG INCOMING: [P0.8 A? D?]

AVG OUTGOING: [P? A? D?]

MENTAL ACTIVITY INDEX: 0

INTRO/EXTROVERSION RATIO: 0

What can be understood here is that Jack is in general, a positive person, who feels positively and acts positively toward others. Janet in general feels negatively, and is negative toward others as well. If deviation had been given in addition to mean, we might see that Janet’s happiness and dominance fluctuates much more than Jack, and is inherently the more tragic and more interesting character (there are many directions to stretch these conclusions). Had many of the further analyses in the aforementioned been implemented, all of the CAD theory’s prescriptions for personality dimensions might be analyzable; but for now, let the aforementioned sentence-by-sentence analysis be suggestive of the system and theory’s potentialities.

In the following section, we explore another more general mode of operation and understanding, and discuss the results of an early indicative evaluation of ESCADA’s performance at a real-world understanding task.

Indicative Evaluation over Free Text

In the previous section’s example, we demonstrated that the CAD theory and implementation is capable of a fair deal of medium-difficulty and medium-specificity understanding tasks when the expression of the story obeys certain pragmatic, stylistic, and syntactic bounds. Of course, the presence of these bounds are akin to the laws of a microworld. All deep story understanding systems in the literature are ultimately only valid within their respective microworld domains (folk stories are a favorite microworld).

CAD and its ESCADA system hope to distinguish itself from previous approaches by also operating on and understanding unrestricted natural language texts; albeit limiting understanding to a comparatively coarse granularity and to the domain of affect and its emergent correlate, personality. This is supported by ESCADA’s flexible parsing and demon-recognition approach, and by the presence of a rich and expansive affective lexicon specifically assembled for this project and for use over general text. Of course, this is a dual mode of operation where all the above-discussed inferences which were dependent on the closed-world assumption (such as action-reaction, since the reaction may not have been detected in the free text) are unusable, but other more general personality understanding techniques and techniques for understanding inter-personal interactions, are still valid, especially when the input text corpus is of a sufficient size so that the results of this kind of opportunistic sensing of personality carries statistical significance.

In order to gain an initial assessment for the quality and robustness of ESCADA’s personality assessment, we performed an indicative evaluation. The goal of this evaluation was to permit ESCADA to analyze the personalities of some notorious characters from fiction and real life, allowing the authors and the readers to gain a better intuition for the true capabilities and limitations of the current system implementation. The corpus used was summaries and digests of great fictional works, taken from ; this source was most appropriate because it is a non-concise third-person narrative paraphrase, a desirable format addressable by the current system implementation. For each fictional work, we report the statistics on personality means (we regret not having fully implemented the translation of these crude statistics into the detailed personality facets we presented in the theory section of this paper) for each of the prominent characters, once for the corpus of the first half of the story, and once again for the corpus of the second half of the story. The reason for this segregation is that either half becomes a baseline for evaluating the other; because both samples come from the same story, corpus idiosyncrasy is controlled for. In addition, the two-halves model allows for the observation that characters often experience change through a text, and calculating a personality mean over the whole of the story text cannot possibly betray this change. Analyzing a story in halves is also simpler than plotting the affective trajectory of each character against an abscissa of narrative time. This alternative visualization which demands some robustness improvements to first be installed, will be tabled for now. For each result, we offer some brief explanatory intuition.

For Romeo and Juliet, the whole corpus was 20,000 words, and the average number of lambdas generated per sentence in the corpus (a measure of recall) is 2.61. The results follow below in Table 1.

Table 1. Two halves analysis of the Spark Notes chapter by chapter recounting of Romeo and Juliet.

|1st |STAT REPORT FOR: Romeo |STAT REPORT FOR: Juliet |

|Half |AVG EGO MOOD: [P0.37 A0.58 D0.42] |AVG EGO MOOD: [P0.56 A0.63 D0.56] |

| |AVG ALTERS MOOD: [P0.2 A0.9 D0.9] |AVG ALTERS MOOD: [P0.5 A0.9 D0.9] |

| |AVG INCOMING: [P0.65 A? D?] |AVG INCOMING: [P0.3 A0.9 D0.65] |

| |AVG OUTGOING: [P0.24 A0.73 D0.86] |AVG OUTGOING: [P0.47 A0.9 D0.65] |

| |MENTAL ACTIVITY INDEX: 0 |MENTAL ACTIVITY INDEX: 4 |

| |INTRO/EXTROVERSION RATIO: 0.25 |INTRO/EXTROVERSION RATIO: 0.67 |

|2nd |STAT REPORT FOR: Romeo |STAT REPORT FOR: Juliet |

|Half |AVG EGO MOOD: [P0.46 A0.63 D0.55] |AVG EGO MOOD: [P0.59 A0.72 D0.59] |

| |AVG ALTERS MOOD: [P0.46 A? D?] |AVG ALTERS MOOD: [P0.2 A? D0.8] |

| |AVG INCOMING: [P0.3 A? D0.8] |AVG INCOMING: [P0.3 A? D?] |

| |AVG OUTGOING: [P0.36 A? D0.8] |AVG OUTGOING: [P0.35 A? D0.8] |

| |MENTAL ACTIVITY INDEX: 1 |MENTAL ACTIVITY INDEX: 3 |

| |INTRO/EXTROVERSION RATIO: 0.5 |INTRO/EXTROVERSION RATIO: 0.67 |

A caveat on the above results: the avg ego mood is the most statistically significant (best recall), followed by outgoing, then alter’s mood, then incoming. The mental activity index is for now, simply a rote count of the number of times a mental activity demon has produced a lambda interpretation. As you can see, recall for this is to be desired, but the results are nevertheless illustrative, since cross-character comparison gives us good baseline.

What this analysis of Romeo and Juliet suggests is that Juliet is consistently more positive and more affectively aroused than Romeo, showing only slight intensification of her character at the end. She thinks more than Romeo and so is perhaps more introspective. Romeo begins quite negative, and expresses negativity toward others (recall that he was in a fight), but his valence ameliorates in the second half. What we can read into these character portrayals is perhaps that Juliet is a positive steadfast character, while Romeo is troubled, but undergoes some positive reform. The current evaluation does not reveal that the ending is tragic, but the sequence of global mood progressions does indeed show much affective activity and extreme emotions toward the end of the story, and then the resolution is characterized by a pacification of global story mood arousal.

For the next corpus, we chose The Unbearable Lightness of Being by Milan Kundera. We used the 8,500 word SparkNotes digest of the story. We would have preferred to use the actual next in this case because it is a third-person narrative but an all-electronic copy could not be obtained. Also, because the novel’s narrative is completely out of sequence, we did not partition by halves. Table 2 holds the results of the system’s run.

Table 2. Character and pair-of-character analysis for the Spark Notes section by section summary of The Unbearable Lightness of Being by Milan Kundera .

|STAT REPORT FOR: tomas |STAT REPORT FOR: tereza |

|AVG EGO MOOD: [P0.48 A0.66 D0.52] |AVG EGO MOOD: [P0.56 A0.61 D0.56] |

|AVG ALTERS MOOD: [P0.67 A? D0.8] |AVG ALTERS MOOD: [P0.8 A0.8 D?] |

|AVG INCOMING: [P0.8 A? D?] |AVG INCOMING: [P0.32 A0.64 D0.78] |

|AVG OUTGOING: [P0.48 A? D0.8] |AVG OUTGOING: [P0.8 A0.8 D?] |

|MENTAL ACTIVITY INDEX: 2 |MENTAL ACTIVITY INDEX: 5 |

|INTRO/EXTROVERSION RATIO: 0.6 |INTRO/EXTROVERSION RATIO: 1.0 |

|STAT REPORT FOR: sabina |STAT REPORT FOR: franz |

|AVG EGO MOOD: [P0.48 A0.58 D0.55] |AVG EGO MOOD: [P0.57 A0.61 D0.53] |

|AVG ALTERS MOOD: [P? A? D?] |AVG ALTERS MOOD: [P0.73 A0.7 D?] |

|AVG INCOMING: [P0.65 A0.8 D?] |AVG INCOMING: [P0.8 A? D?] |

|AVG OUTGOING: [P? A? D?] |AVG OUTGOING: [P? A? D?] |

|MENTAL ACTIVITY INDEX: 0 |MENTAL ACTIVITY INDEX: 0 |

|INTRO/EXTROVERSION RATIO: 0 |INTRO/EXTROVERSION RATIO: 0 |

|0 |INTRO/EXTROVERSION RATIO: |

|STAT REPORT FOR: tomas and tereza |STAT REPORT FOR: franz and sabina |

|AVG EGO MOOD: [P0.55 A0.74 D0.56] |AVG EGO MOOD: [P0.41 A0.62 D0.52] |

|AVG ALTERS MOOD: [P? A? D?] |AVG ALTERS MOOD: [P? A? D?] |

|AVG INCOMING: [P0.8 A? D?] |AVG INCOMING: [P? A? D?] |

|AVG OUTGOING: [P? A? D?] |AVG OUTGOING: [P? A? D?] |

|MENTAL ACTIVITY INDEX: 0 |MENTAL ACTIVITY INDEX: 0 |

|INTRO/EXTROVERSION RATIO: 0 |INTRO/EXTROVERSION RATIO: 0 |

|STAT REPORT FOR: tomas and sabina | |

|AVG EGO MOOD: [P0.61 A0.62 D0.64] | |

|AVG ALTERS MOOD: [P? A? D?] | |

|AVG INCOMING: [P? A? D?] | |

|AVG OUTGOING: [P? A? D?] | |

|MENTAL ACTIVITY INDEX: 0 | |

|INTRO/EXTROVERSION RATIO: 0 | |

The above analysis of the novel’s four main characters and three love pairings is generally correct. The system correctly characterized Franz as more positive than Sabina, and their love pairing as negative; this is consistent with Kundera’s portrayal of Franz as a optimistic romantic and one afflicted with kitsch; it is also consistent with Sabina’s unbearable lightness, a condition of melancholy, and the fact that Franz and Sabina’s love match was a poor one. The system also correctly predicts that Tomas and Sabina’s (his lover) relationship is more positive than his relationship with his wife Tereza; and that Tereza is always thinking (about her relationship with Tomas) and possesses inner strength (since she is the recipient of negativity but nonetheless is positive, acts positively, and feels positively toward others).

Of course this type of evaluation lacks rigorous control, but as an indicative study, it was very informative and gives reason for optimism over the ESCADA implementation and ultimately CAD theory. There are myriad improvements to be had over the design of an evaluation for this type of system, but these must be tabled for future work.

Having discussed the CAD theory, ESCADA implementation, walked through a detailed example, and presented some pilot evaluation, the next section will conclude the paper with healthy reflection on the potentialities of CAD theory, and the relationship of CAD theory to the other story understanding systems presented in the “Techniques for Narrative Comprehension with Imaginative Intelligence” literature (cf. Ph.D. General Examinations Proposal document).

Discussion

In this section, we stretch the story of the CAD theory in various directions, exploring more fully its potentialities and limitations, and making connections to other theories in the narrative comprehension literature.

Representational issues. The representation that the ESCADA implementation has chosen is to extract lambda interpretations out of text, which then in turn updates a simulation world model. A flavor of the frame problem in AI rears its ugly head here. Although working AI systems prove that they are not incapacitated by the frame problem, it does still introduce inconsistencies that must usually be addressed with heuristic assumptions, and ESCADA is no exception. The simulation world representation is lucid (all aspects of the world are exposed and have clear semantics), but the lambda updates to it are only fragmentary and incomplete because the language of the narrative is incomplete; (for example, knowledge such as common sense is suppressed). For instance, consider that “Mary resents John.” The affective certainty here is that Mary’s internal attitude toward John is negative; however, does Mary’s resentment of John cause any change in the state of John? Is John perceptive enough to have his states affected by this act? Does Mary’s resentment of John cause a change in her own mood and in her self-concept? It seems that so many factors are at play that one may never be absolutely sure what is modified. The causality of each act is tangled up in the contexts of each character’s personality, relationship to each other, to the environment, and so on.

The coping strategy to the frame problem employed in the STRIPS system (Fikes and Nilsson, 1971), made the assumption that ‘all that is not explicitly changed by an action remains unchanged.’ There is also an extended STRIPS assumption which makes a address of consistency: ‘states not explicitly changed by an action but which in the post-action sit in inconsistency are changed.’ This solution, while theoretical successful within situation calculus, is not practical to applications like ESCADA which face real-world complexities, such as actions which are incomplete specifications, and with numerous context-dependent entailments. ‘Assuming that states not explicitly changed remain unchanged’ endangers the system by making it possibly inconsistent. In this system, inconsistency can initiate large cascades of misunderstanding. For example, suppose that Mary resents John and then John kills Mary. If John was perceptive, he may have known that Mary resented him, and his killing her would be retaliation. If John did not know about resentment toward him, then that unprovoked act could lead one to characterize him as evil.

In ESCADA’s current implementation, the notion of contextual contagion and imbuement is the coping strategy for the frame problem. Because in our system, inconsistency is so sinister, we hedge our bets. Whereas the STRIPS assumption would prescribe that Mary’s resentment toward John only dictates her attitude toward him and not toward herself, we assume the opposite – that each action nudges all other plausibly related states in the same vector direction. The implemented contextual contagion demons cause every act to also affect the AGONIST in the act; so Mary, in the act of resenting, is made negative herself. We should think of contextual contagion as a form of state relaxation across the simulation.

Not present in the current implementation, we would also propose to implement in future work, a multiple hypothesis model of action. That is to say, unknown variables (e.g. character’s personality) which may affect the entailments of an action are instantiated as multiple hypotheses. Then, each hypothesis is reinforced or de-reinforced depending on whether or not later actions in the story support or contradict each hypothesis. By the conclusion of the story, certain hypotheses will prevail over others. In this manner, the unspecified parameters of character personality may be learned by reinforcing various hypotheses about their values (cf. hypothesis reinforcement in Ram’s AQUA (1994)).

Increasing understanding breadth via explanation. In the detailed story example walkthrough discussed in the implementation section of this paper, we suggested that the ESCADA system could be fitted with more explanatory power, allowing it to complete a broader range of narrative comprehension tasks. Even if not externally prompted for explanation, the system should be able to detect an interesting story feature or anomaly, and try to explain it.

The approach taken by Ashwin Ram’s AQUA system is most relevant (1994) here. In AQUA, explanation is tackled with a case-based memory. The memory contains cases which Ram terms “explanation patterns, ” which are either rote memorized explanations (called “Explanatory Cases”); or “Abstract Explanation Schemas,” fossilized chains of abduction from a consequent (modeled as directed acyclic-graph i.e. decision tree). Each case contains some prerequisites for the execution of that case’s script, and in this sense, each case can be thought of as a recognition demon (since a case is really just a procedure with a condition for execution).

Here is an example of an anomaly whose explanation would bear fruit of understanding: “John’s ice cream is delicious. John dislikes the ice cream.” This violates the expectation established by our assumption that the possession of positively charged objects imbues the possessor with positive feeling. A positively charged object should be desirable. The detection of this anomaly could be accomplished by programming a demon to check for the truth of axiomatic invariants over the simulation world state. In this case, possible explanations could be that the ice cream in John’s mental world is negative; or John is a masochist. Just as is done in AQUA, these hypothesis could persist through the tenure of the story, being occasionally reinforced or disavowed.

Of course, in order to address the aesthetics of narratives, especially fictional texts, we must also accept as valid explanations some of the following: “irony,” “tragedy,” “tension.” In and of themselves, these are aesthetic aspects of story which are perfectly valid explanations; some inconsistencies are actually intended!

Discovering Plot Units. Lehnert’s Plot Units (1982) sought to define abstract prototypical forms for story structure, an ontology of story pieces; to understand the plot unit structuring of a narrative is an important gestalt reasoning task that can then contextualize and inform more fine-grained understandings of story. The elementals out of which Plot Units are composed happen to resemble understandings mineable out of the CAD framework. For example, Lehnert distinguishes between mental acts, positive events, and negative events – all directly readable from the ESCADA simulation world. The challenge in detecting plot units is actually one of attending to certain key events, and not attending to insignificant events. Because of the verbosity of a text, a plot unit could unfold over a chapter or even on the scale of the whole book; however, if the system is unable to disregard insignificant events scattered between significant events, that is to say, if it lacks a good heuristic of relevancy and saliency, then it will certainly miss the important points in the story which actually justify a plot unit. Because of the CAD theory’s cognitive linguistic approach, a possible problem for plot unit detection is that every transitive sentence can be viewed as an event, even though it is an insignificant event, or simply a non-event syntactically construable as an event; this leads to more noise which must be filtered through. However, it may be possible to apply the chunking of smaller events into larger events in order to detect plot units which live at the scale of the whole story. In the Romeo and Juliet example in the previous section, a story was segmented into two halves and analyzed as chunks; from the trajectories of affect across halves and between characters, we may be able to recognize some interesting plot units (though certainly lacking the specificity of Lehnert’s repertoire), like inspiration, or, tragedy, or betrayal.

Metaphorical representation. The knowledge representation employed by CAD theory is fundamentally metaphorical rather than literal, and in that sense, CAD is a first-of-its-kind narrative understanding theory. Not metaphorical is the literary or poetic sense, but rather, the projection of arbitrary thematic role themes (expressed as verbs) onto the affective plane is a metaphorical mapping. The use of the syntacto-semantic vehicle of verb transitivity as a CONDUIT enables the metaphorical interpretation of affective transfer; the impact of the CONDUIT language metaphor on thought is hypothesized in (Lakoff & Johnson, 1980). Its metaphorical approach distinguishes CAD from story understanding systems which deal with literal acts and engage purely in rational reasoning.

In the discourse that Gelernter presents in The Muse in the Machine (1994), rational reasoning and interpretation, which has been the mainstay paradigm of AI story understanding systems, is high-focus thought, whereas metaphorical or analogical reasoning is medium-focus thought. Medium-focus thought is an important component of the totality of human thought, but has historically been under-explored in AI. Medium-focus thought is quite closer to intuitive thinking because of the relative importance of affect as a bridge between memories and ideas; affect itself is completely personal, living within the person-subject, and is often difficult to articulate; in contrast, rationality and language are not inherent to the person-subject because it is in part, defined socially and culturally. Other than Gelernter, many in the Philosophy literature have long held that intuition, not rationality, is the person-subject’s chief method of real understanding (e.g., cf. An Introduction to Metaphysics (Bergson, 1903)).

We would like to think of CAD theory’s metaphorical blur-the-eyes approach to story understanding as at the very least, a good complement to any rational reasoning theory. Additionally, one of the great affordances of CAD and other metaphorical analyzers is that their representations are simpler than with frame-semantics, scripts, goal-schemas, et alia, and thus can use more inexact resources like lexical libraries and connotation engines rather than requiring the careful handcrafting of entirely detailed and precise semantic frames.

Works Cited

Henri Bergson: 1903, “An Introduction to Metaphysics.” Kessinger Publishing 1998.

G. C. Borchardt: 1990, “Transition space”, AI Memo 1238, Artificial Intelligence Laboratory, Massachusetts Institute Of Technology, Cambridge, MA

M.M. Bradley, & P.J. Lang: 1999, Affective norms for English words (ANEW): Instruction manual and affective ratings. Technical Report C-1, The Center for Research in Psychophysiology, University of Florida.

Isabel Briggs-Myers and Mary H. McCaulley: 1985, Manual: A Guide to the Development and Use of the Myers Briggs Type Indicator, by. Consulting Psychologists Press.

Eugene Charniak: 1972, Toward a model of children's story comprehension (Technical Report AITR-266). Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, MA.

Mihaly Csikszentmihalyi, Eugene Rochberg-Halton: 1981, The Meaning of Things: Domestic Symbols and the Self, Cambridge University Press, UK.

Daniel C. Dennett: 1987, The Intentional Stance Cambridge MA, MIT Press

Paul Ekman: 1993, Facial expression of emotion. American Psychologist, 48, 384-392.

R. E. Fikes & N.J. Nilsson: 1971, STRIPS: A new approach to the application of theorem proving to problem solving, Artificial Intelligence 2, 189–208.

David Gelernter: 1994, The Muse in the Machine: Computerizing the Poetry of Human Thought. Free Press.

Karen Horney: 1945, Our Inner Conflicts: The Constructive Theory

of Neurosis. New York: WW Norton.

Ray Jackendoff: 1983, “Semantics of Spatial Expressions,” Chapter 9 in Semantics and Cognition. Cambridge, MA: MIT Press.

C.G. Jung, and H.G. Baynes (translator): 1971, Psychological Types, C.G. Jung, H.G. Baynes (translator). Bollingen Series, Princeton U.P.

Carl Gustav Jung: 1977, Symbols of Transformation. Princeton NJ: Princeton UP (Original work published 1912)

George Lakoff, Mark Johnson: 1980, Metaphors We Live by. University of Chicago Press.

W.G. Lehnert: 1982, Plot units: A narrative summarization strategy. In W. G. Lehnert & M. H. Ringle (Eds.). Strategies for natural language processing. Hillsdale, NJ: Lawrence Erlbaum Associates.

Beth Levin: 1993, English Verb Classes And Alternations: A Preliminary Investigation. The University of Chicago Press.

P. McKee et alii: 1966, Up and Away. Boston: Houghton Mifflin Co.

Albert Mehrabian: 1995, for a comprehensive system of measures of emotional states: The PAD Model. Available from Albert Mehrabian, 1130 Alta Mesa Road, Monterey, CA, USA 93940.

Marvin Minsky: 1986, The Society of Mind. New York: Simon and Schuster.

A. Ortony, G.L. Clore, A. Collins: 1988, The cognitive structure of emotions, New York: Cambridge University Press.

Carol S. Pearson: 1998, The Hero Within: Six Archetypes We Live By, New York, NY: HarperSanFrancisco

Rosalind W. Picard: 1997, Affective Computing, The MIT Press.

Ashwin Ram: 1994, “AQUA: Questions that drive the explanation process.” In Roger C. Schank, Alex Kass, & Christopher K. Riesbeck (Eds.), Inside case-based explanation (pp. 207-261). Hillsdale, NJ: Erlbaum.

Peter Mark Roget: 1911, Roget’s Thesaurus of English Words and Phrases. Retrieved from .

R.C. Schank & R.P. Abelson: 1977, Scripts, plans, goals, and understanding. Hillsdale, NJ: Lawrence Erlbaum.

R.C. Schank: 1986, Explanation Patterns: Understanding Mechanically and Creatively. Hillsdale, NJ: Erlbaum Assoc.

O. G. Selfridge: 1958, Pandemonium: A paradigm for learning. In Mechanisation of Thought Processes: Proceedings of a Symposium Held at the National Physical Laboratory, London: HMSO, November.

Leonard Talmy: 1988, Force Dynamics in Language and Cognition. Cognitive Science 12: 49-100.

................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download