Computing Point-of-View: Modeling and Simulating Judgments of Taste



Download 438.05 Kb.
Page2/8
Date18.10.2016
Size438.05 Kb.
#1905
1   2   3   4   5   6   7   8
§
#1. The space of cultural taste can be modeled as a richly connected fabric of consumerist interests, mediated by subcultural identities, taste cliques, and taste neighborhoods. Individuals can be located on the fabric as patterns of affection. Taste Fabric (Liu & Maes 2004a; Liu, Maes & Davenport 2005) is an implemented system that essays a modeling of cultural taste viewpoint. As illustrated in Figure 2-2, the space of possible cultural tastes is conceived as a richly connected semantic fabric whose nodes are consumerist interests—such as music (artist, album, genre), books (author, title, genre), films, sports, foods, etc. The pairwise taste-affinity or similarity between each pair of interests is specified numerically. The topology of taste fabric was acquired by computing the latent semantic connectedness of interest keywords found across the texts of 100,000 social network profiles. An individual’s location in cultural taste space is conceived intertextually as the region in the fabric demarcated by a grab bag of favorite interests, such as that which could be acquired by reading each person’s social network profile or homepage.
T


Figure 2-2. A semantic fabric representation of viewpoint in cultural taste space.
his analysis of cultural taste in terms of consumer interests quite resembles the data that Bourdieu had worked with. The notable difference between the approaches is that while Bourdieu sought to implicate capitals as the fundamental dimensions of taste-space, no basic dimensions are assumed or discovered by Taste Fabric. Instead, taste distance calculation relies on the tightly interwoven nature of consumerist interests. The density of interconnections is made possible by the scale of mining 100,000 profiles in which 20-50 keywords are extracted from each. However, the topology of the fabric is far from smooth and uniform—local structuration is supplied by semantic mediators—such as special nodes representing subcultural identities (e.g. ‘hipster’, ‘urbanite’, ‘socialite’ as depicted in Figure 2-2); cliques of tightly knit interests; and larger regions of interests called ‘taste neighborhoods’.
#2. The space of gustation can also be modeled using a semantic fabric—as an interweaving of recipes, ingredients, and cooking procedures, mediated by cuisines and basic flavors. The Synesthetic Cookbook (Liu, Hockenberry & Selker 2005) is a viewpoint system implemented for the gustatory space. The gustatory semantic fabric constitutes the space of foods and flavors by interconnecting 60,000 recipes, 5000 ingredient keywords, 1000 sensorial keywords (e.g. ‘spicy’, ‘chewy’, ‘silky’, ‘colorful’), 400 cooking procedures, and 400 nutritional keywords with numerical affinity scores. Additionally, cuisine types (e.g. ‘Chinese’, ‘dessert’) along with basic flavor keywords (e.g. ‘sour’, ‘spicy’) structure the gustatory space by acting as semantic ‘hubs’, each reaching out to potentially thousands of ‘spoke’ nodes.
Relations between foods and sensorial keywords were mined from the Thought for Food corpus—thousands of sentences embodying cooking common sense. Resemblance relations between recipes and ingredients were mined by parsing and correlating amongst the 60,000-recipe database, and the several food-related encyclopedias and knowledge bases collected together as the ‘cultural corpus’. Atop the gustatory fabric, users of the system can define their own ‘tastebud’, i.e. gustatory location, as a pattern of foods that they like, love, dislike, and hate. Because user-defined tastebuds go beyond keyword specification to entail a contextual region in the gustatory fabric, taste judgments can be simulated quite robustly and meaningfully for almost any food-related stimulus.




Figure 2-3. Jung’s four-dimensional representation of aesthetic perception viewpoint.

§
#3. Aesthetic perception can be modeled as a psychological space whose dimensions are Jung’s four fundamental psychological functions—sense, intuit, think, feel. Aesthetic perception viewpoint is an individual’s preferred disposition for gathering and processing sensory data. While cultural taste and gustatory taste are clearly mediated by linguistic materials such as interest keywords and food keywords, it is less clear how much the way we perceive the world is shaped by the external culture, and how much is due to innate or internal factors. The best theoretical account of perceptual dispositions come from Carl Jung’s theory of type [], and this is the model that is adopted and computed here as a viewpoint space. In his theory of type, Jung proposed four fundamental psychological functions—sense, intuit, think, feel—to account for all the differences in human perception. The example illustrated in Figure 2-1 explains this Jungian dimensional space—Mary and Jack assumed different aesthetic perception viewpoints unto the meaning of ‘sunset’. Sunset evokes feelings (e.g. ‘romance’, ‘warmth’) and intuitions (e.g. ‘embrace’) in Mary, while for Jack sunset evokes ratiocinations (e.g. ‘off work’, ‘dinner’) and sensations1 (e.g. ‘dark’). If Mary and Jack’s judgments are indicative of their longer-term dispositions for perception, then it might be said that Mary is a romantic while Jack is a realist.


So, the aesthetic perception space is a dimensional space, having the four axes—sense, intuit, think, feel. That an individual is located in a coordinate region of this space means she is predisposed to interpreting sensory data in a particular way. For example, a romantic is located in high feeling, high intuition, low sensing, low thinking. An implemented viewpoint acquisition system called Experimental System for Character Affect Dynamics Analysis (ESCADA) [] performs psychoanalytic reading over an individual’s self-expressive texts, such as a weblog, and infers the individual’s position in aesthetic perception space. Because this is a psychological space, possible linguistic materials that the viewpoint may react to are not contained within it, per se. Rather, the space may be thought of as a prism which, whenever passing over a symbol, refracts it into its Jungian spaces of possible interpretations. The Aesthetiscope (Liu & Maes 2006) has machinery that can refract symbols into their tensors of interpretation. This is illustrated in Figure 2-3, where the whole space is passing over the symbol ‘sunset’, and the annotated axes suggest how a tensor of possible interpretations may fill up the space.2
§
#4. The space of attitudes can be modeled using semantic sheets, as the set of all affective orientations toward all possible topics. Individuals can be located in this space as one particular pattern of affection over the topics sheet. Behaving like individuals, a culturally average pattern of affection can also be recorded as a sheet so individual attitudes can be compared against cultural norms. Semantic sheets are a more disconnected representation than the continuous dimensional space representation and the densely connected semantic fabric representation. They are as such because the space of attitudes is even more diverse than the space of consumerist interests which mediates cultural taste. There is plenty of organization and consistency in semantic sheets, albeit localized. As shown in Figure 2-4, the sheet of topics bear implicit organizations due to the inter-relatedness of concepts at the linguistic level, and the folksonomic organization of concepts at the cultural and social. Figure 2-4 also suggests ideology as a semantic mediator; for example, two sets of ideological alignments are shown between the person’s locations and some cultural norm.




Figure 2-4. A semantic sheet representation of viewpoint in attitude space.



#5. The space of humor can likewise be modeled using semantic sheets, constituted by archetypal patterns of psychic tensions for numerous niche humor cultures. There are many accounts of humor in literatures computational and humanistic. The experiment pushed forth here proceeds from on Freud’s (1905) account of tendentious jokes, and is founded in early psychoanalysis’s conceptualization of the affective unconscious as an expanding and contracting hydraulic bag, which alternatingly stores and gives catharsis to psychic energies and tensions. Freud distinguished between innocent and tendentious jokes. Innocent jokes elicit just a smile or chuckle and are not emotionally heated, while tendentious jokes have a sexual or aggressive character, and can draw out aggressed, howling laughter. Tendentious jokes are a boon to the psychic economy of the listener, says Freud. They give catharsis to psychic tension which has pent up around particular topics. Freud theorizes that people of similar cultural background, such as members of a lifestyle or ethnic group—due to shared upbringing and experiences with family, gender, sex, and money—will also share a pattern of social inhibition and psychic tension. Cultural humor, then, effects aggressive laughter as a way to give catharsis to tensions created by social inhibition. Sexual jokes are prevalent, in part because the almost all societies inhibit sexuality to some degree. Based on Freud’s theory of the tendentious joke, the notion of archetypal tension is conceptualized as patterns of psychic tensions shared by a niche cultural grouping of persons. By mining the archetypal tension of each niche culture, and by classifying individuals as members of those cultures, jokes can be paired with their best audiences.
Catharses, an implemented viewpoint system for humor, shares a representation and mechanism with the attitude viewpoint system. A sheet of all possible topics is overlaid with an individual’s affective positions over those topics. Instead of a three-tuple affect score, each topic is associated with a unary score tracking the psychic tension level surrounding that topic. Harkening to Freud, psychic tension is measured as the state of negative valence, high arousal, and dominance. A corpus of 10,000 jokes is categorized into humor niches such as blonde jokes, foreigner jokes, sexual jokes, Bush jokes, Clinton jokes. For each niche, a group of ten persons with weblog diaries were manually identified as ‘exemplars’ of the niche culture. Catharses modeled the viewpoints of each exemplar, and the viewpoint of the niche culture as a whole—resulting in an archetypal pattern of psychic tension for each niche. By identifying the niche cultures that a person best resonates with, the most suitable jokes can be recommended. Furthermore, as an individual’s tensions shift daily, as observable in recent weblog entries, Catharses becomes a therapeutic daily companion.
§




Figure 2-5. A semantic diversity matrix

Three different knowledge representations were introduced to support the modeling of five viewpoint spaces—semantic fabrics, dimensional spaces, and semantic sheets. Yet why three and not simply one? Because sometimes the viewpoint space to be modeled has straightforward dimensionality (Figure 2-3) while other times a space can appear quite disorganized (Figure 2-4). The choice of representation is ultimately an engineering consideration, but I believe that the three representations developed through this thesis are principal.


The lifeblood of modeling aesthetic and cultural matters are rich associations and underlying consistencies. Thus, there is a clear preference for certain knowledge representations. Dimensional spaces such as aesthetic perception space (Figure 2-3) are most preferred, as meaning is most organized, and Cartesian distance is easily measured. Next best are semantic fabrics, which are densely connected networks with topological features like cliques and stars. Though semantic fabrics are super-connected representations, they lacking total ordering, and possess only patchwork consistency. Distance is non-Cartesian here, but it can still be measured simply by spreading activation (Collins & Loftus 1975). The mining of the latent space of cultural tastes from social network profiles leverages semantic fabrics because while the mutual information between consumerist interests can be calculated, it is believed that the dimensionality of this space are too complex to be able to name principle dimensions. In the poorest case, neither dimensions nor connectedness are known, as is the situation for the attitude viewpoint space. The space of all possible topics is made locally consistent by semantic mediators like ideology and genre divisions, but there is no clear global consistency. The semantic sheet representation makes the best of this situation—it is in essence a well laid out annotation model.
Yet all knowledge representations are equivalent in the sense that, though maladroit, each viewpoint space could be represented by any of the three introduced knowledge representations. Inspired by Marvin Minsky’s “causal diversity matrix” (Minsky 1992) in which Minsky plots the space of intelligence problems along the axes of number of causes and number of effects, Figure 2-5 presents a “semantic diversity matrix” which summarizes the representational tradeoffs just discussed. Note that a third axes could also be named—semioticity. With the additional dimension, “dimensional spaces” could be distinguished as either have principled and theoritized axes such as Jung’s psychological functions, or as being an unprincipled, data-emergent “quality space” (Gärdenfors & Holmqvist 1994).
2.3 The judgmental apparatus

The seed grows out of the soil. If a point of view is one’s location within various latent semantic spaces, then perspective is that apparatus interior to each person formed by uptaking the exterior semantic neighborhood available to the given location. Obeying constraints such as the principle of economy, and the principle of capital, this interior apparatus has an integral organization, thus it is not an arbitrary rote memory; the apparatus affords taste judgments in systematic fashion, thus it contains canonical vehicles of interpretation. The previous section exposed static representations for viewpoint space and location. This section suggests how a judgmental apparatus can be built from those pieces and animated to simulate judgments of taste.


§
Constructing the judgmental apparatus out of the materials of a location means molding those available materials into a coherent system with distinct spirit and sense. Coherency and consistency of perspective can be seen as necessary in several respects, argued here. First, human memory and cognitive function are restricted by the principle of economy. Given that physical memory is resource constrained, mental structures optimize for maximum utility—e.g. dream-work can be regarded as a garbage collection process crunching over recent experiences, absorbing some into consolidated memory, rejecting extraneous others; creativity means re-appropriating knowledge tied to one context, into other unexpected contexts, thus the knowledge which tends to persist are those with overloaded utility, versatile enough to solve problems under a variety of contexts. Second, competency in the social world necessitates the need to communicate oneself, and consistencies of viewpoint facilitate its communication. Like viewpoint, social identity is not the whole of an individual but rather just a collection of coherent roles played by the individual—this can be seen in the Etruscan etymon of the word ‘person’, phersu, meaning ‘mask’. If a viewpoint is self-consistent, then every judgment will imply all other judgments from that viewpoint, allowing others to more quickly grasp the gist of someone. Third, individuals’ possession of social, cultural, and political capitals, means that their viewpoints will necessarily inherit from social, cultural, and political archetypes and norms. Bourdieu’s argument for the primacy of the class habitus indicates this convergence. Fourth, consistency and coherence is an aesthetic desideratum. Individuals’ consumptive choices tend to cohere around their design for a life-style—clusters of goods having common implications to lifestyle are known as ‘Diderot Unities’ (McCracken 1988) and ‘consumption constellations’ (Solomon & Assael 1987). Cognitive dissonance (Festinger 1957), the unpleasant experience of mental self-conflict, leads individuals to resolve those quandaries by maintaining ‘truth’ and consistency.
Building the judgmental apparatus means post-processing individuals’ acquired location models with network operations called semantic relaxations—these serve the dual purposes of 1) maintaining ‘truth’ and consistency in the system of judgments (Doyle 1979), and 2) dilating scope and robustness—using judgmental precedents known explicitly in the location model to guide assumption-making over unprecedented judgments. Specific invocations of semantic relaxation for the five viewpoint systems are now discussed.
In the spaces of attitudes and humor, known affections are relaxed along the lines of implicit topic hierarchies. Consider that an individual’s affection toward the topic of ‘water conservation’ is known to be pleasurable-arousing-dominant (+1.0,+1.0,+1.0), and consider that the same individual’s affections for the topics ‘environmental protection’, and ‘recycling’ are unknown to the individual’s acquired location model. To transform the model into the judgmental apparatus, the known affection for ‘water conservation’ is semantically relaxed over implied topic hierarchies using spreading activation. That ‘environmental protection’ is a super-topic of ‘water conservation’ is known through hypernym knowledge bases such as WordNet [], ConceptNet [] and folksonomies such as DMOZ [], and so affection flows ‘upstream’ from topic to super-topic. Along the way, an upstream discount of 0.75 is applied, resulting in ‘environmental protection’ (+0.75,+0.75,+0.75). Assuming that ‘recycling’ is the only sub-topic of ‘water conversation’, with a downstream discount of 0.5, the relaxation result would be ‘recycling’ (+0.5,+0.5,+0.5). Topics with multiple sub-topics impart their affection downstream with an additional ‘fan-out’ discount which is inversely proportional to its number of sub-topics. The mechanism of spreading activation continues propagating affection from topic up to supertopic to super-supertopic, and from topic to subtopic to sub-subtopic, each time applying the appropriate discount, until the post-discount affection is negligibly small. Note that the top-level topics are discourses. Spreading activation outward from all known affections, some topics will inherit affection from multiple other topics. In spreading activation, multiple affections are collapsed into a single affection by averaging. Had the multiple affections been contradictory, those affections would have cancelled out into neutral affection; had they been corroborating, those affections would have led to a more pronounced affection score. The hope is that spreading activation propagates precedent into the unprecedented, but if precedents should ever bear contradictory prescriptions for an unprecedented, that unprecedented should stay unknown.
In the spaces of cultural taste, a person’s location model is relaxed from an explicit set of items into an ethos, which subsumes them. A person’s location in cultural taste begins as an explicit set of consumerist interests—a set that can be mined from a homepage or a social network profile. The explicit set, though, can be thought of as a pointillistic impression of a more profound but ineffable ethos (character) which pervades and unifies members of the explicit set. Defining the ethos, then, is the performance of an aesthetic relaxation, and is achieved by spreading activation. Suppose that A, B, and C are three nodes which in the Taste Fabric have strong mutual affinities, forming a clique. Supposing then that only A and B are members of the explicit set of items, the result of relaxation is to implicate C also into the ethos of the individual. Relaxation is mediated by subcultural hubs, taste-cliques, and taste neighborhoods. If item A has a strong affinity to subculture X, and X has strong affinity to items A, B, C, and D, then when activation spreads into X, all of strong members of X are immediately pulled into the gathering ethos. An individual’s ethos is far more likely to stumble into a few subcultural hubs and pick up all their member spokes than to be constituted by nodes which are not interconnected themselves by semantic mediators. The emergence of subcultural hubs as forces of coherence in the machine-learned Taste Fabric could be construed by some to support Bourdieu’s claim that cultural capital is a major determinant of taste, since subcultural nodes function as indexicals of goods sharing a cultural capital value.
A person’s known attitudes can be supplemented by the attitudes of her imprimers. Marvin Minsky (forthcoming) introduced the notion of ‘imprimer’ as a mentor or confidant that one forms an affected attachment with, and engages in a viewpoint mimesis of. For example, parents imprime their children, an advisor imprimes a student, and even fictional characters can imprime. Minsky’s litmus test for an imprimer is one who can stir self-conscious emotions such as pride and embarrassment within oneself. In modeling attitude viewpoint from a person’s self-expressive texts, imprimer relations are also assessed by Minsky’s method, and the texts of the imprimers are likewise modeled as a semantic sheet. In constructing the judgmental apparatus, imprimer’s sheets become incorporated into the apparatus, acting as mental critics who can produce supplemental judgments in cases where the individual’s own model offers no precedent. The hypothesis is that our natural tendency for mimesis would cause us to outsource judgments we were not sure of to our mental representations of our imprimers’ viewpoints. An evaluation testing this hypothesis will be presented in Chapter 4.
§
Once the judgmental apparatus is constructed for each viewpoint realm by performing semantic relaxations over an individual’s location in those spaces, the apparatus can simulate judgments of taste, given some textual or semic input, which is referred to here as fodder. The simulation process has two steps—first, unknown symbols in the fodder are resolved by mapping them into the known semantic space; second, memory-based reasoning and aesthetic distance calculation are invoked to appraise the fodder. Reflecting upon the gestalt of what is modeled and simulated, we see that it captures a stereotype of an individual’s perspective, useful, though still lacking much of persons’ dynamism and dialogical nature.
The input to judgment simulation is fodder—meaning any textual or semic input, such as a concept, a sentence, or a document. Given the thesis’ semiological approach to point-of-view, all aspects of computation are phrased in terms of linguistic signal processing. Culture and the individual are acquired textual systems, and the inputs to taste judgment are textual, so the output of taste judgment is a numerical score reflecting the aesthetic resonance between person and fodder. Some examples of suitable inputs follow. The WWTT attitudes system can react to propositions like “robots will have consciousness” with pleasure level, arousal level, and dominance-submission level. The Aesthetiscope perspectival art robot [] takes fodder such as the word ‘sunset’ or a poem such as Pablo Neruda’s “A Song of Despair,” and outputs an artwork which depicts the fodder through a person’s perceptual viewpoint. Ambient Semantics [] is a matchmaking system which takes as fodder the text-based profile of another person, and outputs the numerical degree of taste-affinity between the self and that other person.
Step 1—the meanings of novel symbols are resolved within the known semantic space via conceptual analogy. Each viewpoint space, purporting to represent all semantic possibilities for a particular realm, is constituted out of symbols belonging to some finite lexicon. Call this the discourse of the realm, or the known semantic space. If a fodder contains symbols (e.g. word, concept, event) that are not in the realm’s discourse model, those unknown symbols must first be mapped to their best-guess counterparts within the discourse model in order to continue the appraisal process. In the cognition literature, the task of resolving symbols from outside a system to within a system is known as conceptual system alignment. Two approaches to the task are known to the literature. First, neural network learning can be engaged to create approximate correspondences between two graphs (Goldstone & Rogosky 2002). Second, structural analogy can be engaged to find correspondences between two concepts based on shared intensions (Gentner 1983; Fauconnier & Turner 2002). The thesis follows the latter approach, and appropriates the ConceptNet common sense reasoning system [] to resolve unknown symbols. For example, suppose that the fodder ‘war’ is fed as input to a viewpoint model of an environmentalist in WWTT. Suppose that the environmentalist has only commented on topics relating directly to the environment, and so the symbol ‘war’ is unprecedented. Feeding ‘war’ into ConceptNet’s get_analogous_concepts() function results in a rank-ordered list of analogous concepts, among which are some environmental symbols. Some actual output edited for legibility is shown below.

[war is like storm] the concepts share:


==PropertyOf==> bad
==PropertyOf==> violent
==PropertyOf==> dangerous
==CapableOf==> destroy property

[war is like pollution] the concepts share:


==PropertyOf==> evil
==CapableOf==> kill
==CapableOfReceivingAction==> cause
==CapableOfReceivingAction==> stop

Based on the shared attributes, functions, and effects between ‘war’, ‘storm’, and ‘pollution’, the known symbols ‘storm’ and ‘pollution’ can replace the unknown symbol ‘war’ in the fodder, by applying a 0.5 uncertainty discount to the substitute concepts. ConceptNet is capable of resolving not only words, but also second-order concepts like ‘eat burger’, ‘open door’, etc. Its capabilities are covered more fully in later chapters.


Finally, it must be noted that the application of analogy in the simulation of taste judgments is heuristic. It is assumed that the environmentalist’s known attitudes toward ‘storm’ and ‘pollution’ may be lifted out of their original discourse context, and be opportunistically re-appropriated to appraise the concept of ‘war’. This heuristic act is quite related to the notion of sublation (Aufhebung) in literary theory, which Hegel [] foregrounded as the tendency to break off older ideas and contain aspects of them into newer ideas. The sublation phenomenon is seen to support this thesis’s assumption of aesthetic consistency within a single judgmental apparatus.
Step 2—memory-based reasoning and aesthetic distance calculation are applied to the fodder to produce taste judgments in the form of numerical affinity scores. Montgomery’s ( ) psychological study reported that the valence of perspectival judgment was proportional to the psychological distance between the experimental subject and object. Building upon this finding, in the cultural taste realm, a taste judgment is positive if the fodder is proximal to the individual’s ethotic cloud of nodes, and is neutral or negative if the fodder is distant. The aesthetic distance between individual’s ethos and the fodder is calculated as the percentage overlap between the ethos of individual, and the fodder’s ethos, i.e. fodder node plus its surrounding nodal context. In the realms based on the semantic sheet representation (i.e. attitudes realm and humor realm), taste judgments are produced by applying memory-based reasoning [] to the fodder. After a fodder is parsed, a bag of recognizable topics are extracted, and each topic evokes its corresponding affect score as given in the individual’s semantic sheet, or imprimers’ semantic sheets (Figure 2-4). The appraisal of the fodder overall is produced as the linear combination of all evoked affect scores, weighted by the importance of each topic to the fodder.
Limitations—this computational strategy captures a stereotype of an individual’s perspective, but it still lacks much of persons’ dynamism and dialogical nature. Appraising fodder using the precedent of prior recorded judgments can only constitute a stereotyped representation. First, doubt can be cast on the universality of the aesthetic consistency hypothesis. Ideology and other forces may impose exceptions to aesthetic consistency. For example, ‘tree’ and ‘rock’ share the super-topic of ‘nature’, so aesthetic consistency presumes that positive attitudes about ‘tree’ can induce symmetrically positive attitudes about the sister concept of ‘rock’. However, aesthetic consistency is violated in the instance of ‘dogs’ and ‘cats’, which share the super-topic of ‘pets’. It is far from clear that a sympathetic attitude toward ‘dogs’ can predict a sympathetic attitude toward ‘cats’. Looking at empirical data in WWTT, ‘dogs’ and ‘cats’ actually tend to form an aesthetic opposition—dog lovers tend toward a distaste for cats, and vice versa. Pet preference seems to be a politicized and ideological space exempt from aesthetic consistency—perhaps because pets are so often invoked in the present culture to signify the personality of their owners. Dog lovers are presumed to be social, whereas cat lovers are presumed to be asocial.
Second, even if prior attitudes could be applied judiciously to incoming fodder, a person’s reaction is more sophisticated than simply reiterating their views. Most views are complex enough such that a person may adopt both positive and negative attitudes about a topic under different social or political contexts. A new situation’s fodder may provoke a particularly clever, even self-contradictory reaction, as an individual hopes to achieve sarcasm, irony, or self-overcoming. With his notion of doxa—a person’s self-conscious location in culture—Bourdieu meant to suggest that perspective could be a reflexive representation that is itself a material subject to appropriation. In [], Bourdieu explains that theorists, for example, are keenly aware of their location in the space of criticism, and thus their acts of judgment often deviate from their locations because they are constantly playing games with their doxa. This sort of self-reflexivity which alters judgments, reflects the dialogical nature of human dynamism (Bakhtin 1935). Unfortunately the sophistication of this dynamism cannot be simulated, and is far beyond the computational scope of this thesis work. The complexity of the problematic—whose heart is human creativity—might fairly be considered AI-complete.
Thus far in this chapter, a theoretical framework for representing and acquiring viewpoint in its various realms, and a method for simulating taste judgments via the judgmental apparatus have been presented. However, stepping away from theory, the prospect that individuals’ taste perspectives could be reified, shared, and simulated suggests new approaches for learning, self-reflection, matchmaking, and deep recommendation. The final section of this chapter overviews the six built perspectival artifacts, and distills from these examples some principle design desiderata.
2.4 Designing perspectival artifacts

If a person’s perspective could be captured and reified through a computational artifact, what sorts of interactions should such artifacts support, and what could such technology mean for applications? In this section, six implemented perspectival artifacts—an art robot, a taste-based matchmaker, an identity mirror, a panel of virtual mentors, a tastebud-oriented cookbook, and a jocular companion—are overviewed (to be fully presented in Chapter 5). Then, three key interaction design principles for perspectival artifacts are distilled—1) continuous observation and feedback, 2) just-in-time and just-in-context, and 3) tinkerability.



§
#1. The aesthetiscope is a robotic artist capable of assuming someone’s perceptual perspective and creating artwork through that lens. Aesthetiscope engages in a creative reading of an inspirational text such as a word, poem, or song lyric—based on a particular creative reading, the art robot creates color grid artwork a la Ellsworth Kelly and the early twentieth century abstract impressionists. The peculiar nature of the creative reading is driven by a person’s perceptual viewpoint model. This reading guides the manner and quality of the generated artwork. Aesthetiscope continuously observes items in the user’s attentional context such as what poetry the user is reading or what songs are queued in the playlist, then dynamically changes the color grid artwork to ‘pair’ with the fodder, just as wines are selected to pair with meal courses. Another perspectival game that can be played is for two individuals both standing in front of the same artwork to find their shared viewpoint by averaging their models, or to violate the other’s aesthetic rendition by using one model to corrupt the other model.
#2. Ambient Semantics is a social matchmaking system, which facilitates serendipitous social introductions based on shared tastes. The Ambient Semantics system monitors wireless wristbands worn by users, and supports their interactions with objects (e.g. books) and other people (e.g. via a handshake) with just-in-time information displayed on a nearby wall projection. Each user is backed by their social network profile, and a corpus of their texts such as their blog or homepage, and from that, their cultural taste viewpoint is acquired. To support social introduction with strangers, the system visualizes the shared context of two people, including common friends, and their shared tastes and interests. Whereas exact interest keyword matching would miss the opportunity to connect a ‘rock climber’ with a ‘skydiver’, the intersection of those two cultural taste viewpoint models reveal the shared identities of ‘adventurers’ and ‘extreme sports lovers’.
#3. Identity Mirror shows you your dynamic textual reflection against the changing cultural fabric. This mirror lets you see who you are, not what you look like. Powered by cultural taste viewpoint, the mirror visualizes a performer’s ethos as a swarm of keywords laid over an abstracted image of the performer as captured by a live digital video camera. The performer can use dance to negotiate her identity—for example, walking to and fro the mirror affects the granularity of the keywords being shown, describing a far away performer with subcultural keywords, and an up-close performer with descriptors like song names, books, food dishes, etc. Reflecting the dynamicity of cultural taste viewpoint, the reflection updates itself continuously as it monitors live news feeds which represent the pulse of the cultural zeitgeist. For example, when Oscars season strikes, the reflection portrays the more glamorous side of the performer.
#4. What Would They Think? is a panel of virtual mentors capable of providing just-in-time affective feedback. Virtual mentors simulate the attitude perspectives of mentors or friends. Each virtual mentor is depicted by a black-and-white photographic icon, and the whole panel of mentors can be placed anywhere on a computer desktop. As a computer user surfs websites, writes research papers, or composes emails and instant messages, WWTT is continuously observing the user’s textual context, and producing affective reactions to that context. The three affective dimensions of each reaction are mapped to this visual metaphor: the degree of pleasure and agreement tints the mentor icon progressively greener while displeasure and disagreement tint the icon red; arousal is rendered by the luminance of the icon; dominance manifests as sharpness of image, whereas submission begets blurriness. If a user wishes to further investigate a produced reaction, double-clicking an icon will return a list of relevant quotes from that mentor which supports the reaction.
#5. The Synesthetic Cookbook is a social cookbook whose avatars learn the tastebuds of family members and simulate their reactions as you browse recipes. This visual recipe browser allows users to articulate their cravings by imagining the tastes of food. For example, typing “old, beautiful, urgent, alive, homey, aromatic, zen” yields a recipe for “bohemian stew.” Foraging for food is made social by avatars whose tastebuds are programmed for the gustatory viewpoint of family members. So as a mother browses for dinner suggestions, avatars for her husband and kids constantly emote their reactions to recipes alongside her.
#6. Catharses is a jocular companion that releases stress playfully by seizing upon everyday gripes as opportunities for humor. Catharses is a desktop agent that monitors a person’s everyday narrative activities such as instant message conversations, email writing, browsing, and the composition of daily weblog entries. Based on a prior model of the person’s humor viewpoint, plus the context of everyday happenstance, Catharses delivers jokes apropos to the user’s ever-shifting moods.

§
Three design principles for communicating perspective were distilled from the creation of these perspectival artifacts—1) continuous observation and feedback, 2) just-in-time and just-in-context, and 3) tinkerability. These are discussed below.


Principle #1: continuous observation and feedback. It may be said that just as light has no resting mass, point-of-view is not intelligible in stasis. To fully appreciate and intuit a viewpoint, it must be animated and allowed to react to a broad many things. In the interaction design literature, Bill Gaver (1991) has foregrounded the idea that an artifact is easy-to-use when its affordances are perceptible. According to Gaver, perceptual psychologist J.J. Gibson (1979) first coined the term ‘affordance’ to describe the ability of people to intuit an object’s potential for actions from perceptual cues and feedback. For example, thin vertical door handles afford pulling, while wide horizontal doorplates afford pushing. The case of viewpoint, however, affords more complex actions. According to Gaver, systems of complex actions require more active perception, and many exploratory engagements with the system. Complex objects can often be dissected into nested affordances revealed over time. The affordances of a perspective might likewise be conceptualized as consisting of semantically nested affordances—that is, aesthetic consistencies latent in the perspective can suggest divisions of affordance-space. By Gaver’s mode of analysis, then, a useful way to expose the affordances of a viewpoint could be for a perspectival artifact to continuously observe a user’s textual context, and proactively offer feedback. Because information push treats all user browsing and writing activity as implicit fodder, it avoids the time cost of users manually formulating queries. Thus, exploration of affordances can be had with less effort, and animating a perspective resembles an exploratory walk through a perspective, which increases its surface area of perceptibles.
Principle #2: just-in-time, just-in-context. Perspectival artifacts continuously observe the user’s context and offer feedback. The nature of this feedback should be just-in-time—meaning that feedback given should pertain to the user’s present textual context, such as the text on a webpage the user is scanning over with the mouse, or the sentence that the user just completed. The design of just-in-time information retrieval (JIT-IR) agents was advocated for by Bradley Rhodes [], and was exemplified by his Remembrance Agent, a proactive memory aid that used the physical context of the user to trigger relevant notes. JIT-IR interaction is most appropriate when interacting with a perspectival artifact because the artifact’s affective reactions only make sense when can be readily bound to the user’s present context. But more than just-in-time, the nature of feedback should also be just-in-context—meaning that the reaction should as much as possible pertain to the gestalt of the user’s present context, rather than considering only particular elements in the context. This is important because reacting to the gestalt imparts a nuanced understanding about how judgments about different topics might synergize and interact with one another when the different topics are elements of the same fodder. It is the position of this thesis, after all, that taste is intertextual, so it demands to be explored as such. Finally, because users are always creating textual contexts out of novel combinations of different ideas, just-in-context reactions should also tend toward novelty, often producing provocative or surprising reactions motivating further investigation.
Principle #3: tinkerability. While proactive feedback gives users of perspectival artifacts many perceptual entrees into a perspective, there needs also to be a way for users to tinker with the perspective itself. For example, WWTT allows a perspective to be changed in a text editor, and allows a user to dig deeper into any reaction by asking a virtual mentor to justify a reaction with a corpus of quotes from memory. The Aesthetiscope allows the user to tweak the perceptual perspective by moving horizontal sliders to change location in the Jungian perception space, and then immediately visualizing reactions. The tastebuds of the avatars of the Synesthetic Cookbook can be reprogrammed with keywords on-the-fly by clicking on the mouth to reveal the tastebuds. Tinkerability also ensures that users learn about the limitations of computational viewpoint modeling—for example, seeing the quotes from a virtual mentor which predicted a reaction allows users to verify the accuracy of the perspective.
3 Psychoanalytic reading

The point-of-view models architected in Chapter 2 can be automatically acquired by analyzing the self-expressive texts of the individual, and the texts of the culture that circumscribes her. The method of textual analysis introduced here will be termed psychoanalytic reading. Application of the method to the analysis of cultural texts will be termed culture mining. In this chapter, 3.1 introduces the genre of self-expressive texts; 3.2 gives presents a semiotic schema of self-expressive texts; 3.3 discusses the mechanics of psychoanalytic reading; 3.4 illuminates culture mining as an extension of psychoanalytic reading. 3.5 and 3.6 present two building-block technologies central to psychoanalytic reading and viewpoint simulation—commonsense reasoning and textual affect sensing. But first, a motivation for psychoanalytic reading is developed, accompanied by a working definition.



§

Psychoanalytic reading is defined as a computational method of excavating various viewpoint models of a writer from her self-expressive texts by reading for implied associations and judgments, and correlating these implications into a system of subjectivity. The following subsections concretize this working definition by discussing the subjective capacity of self-expressive texts, and the principles and mechanics of computing psychoanalytic reading. But first, a background of and motivation for psychoanalytic reading is developed.
Psychoanalytic reading extends recent work in computational reading. Where early AI systems assumed a monolithic model of understanding, position their work as “story understanding” (Winograd 1971; Charniak 1972; Dyer 1983), more recent work reflects a more nuanced cognitive view, which is newly unencumbered by the anxiety of having to discover the text’s true and objective meaning. Moorman and Ram’s (1994) ISAAC reader could focus, attend, and willfully suspend disbelief, AQUA (Ram 1994) could interleave and motivate reading with the asking of questions. Srinivas Narayanan’s KARMA system (1997) reads texts metaphorically, using Petri-nets to understand physical metaphors in text, e.g. “Japan’s economy stumbled.” ISAAC, AQUA, and KARMA all rely on underlying situation models (Zwaan & Radavansky, 1998), a construct meant to operationalize rationally coherent comprehension.
Up to now, the situation models proposed and explored in the computational reading literature have focused on capturing the explicit and intended messages communicated by a text. However, subjectivity and viewpoint are not so much the foreground message of the text as they are the background parameters of all messages. Often called the mood, tone, style, or manner of a text, these subjective parameters leave traces in all aspects of the text. The modeling of these subjective parameters from the text has been addressed in the semiotics literature, especially in the divisions of narratology and stylistics. A goal of this thesis is to import the semiotic approach to reading for subjectivity into the computational literature.
In order to excavate the writer’s viewpoint from a reading of her texts, heavy emphasis is placed on the unconscious judgments and symbolic associations implied by each utterance. These associations and judgments can be organized into coherent models of the writer, namely, the viewpoint models introduced in Chapter 2. Because reading for viewpoint is the art of sense-making about judgments and associations, and their significance to the writer, it is illuminated here as a psychoanalytic project, par excellence. It should be noted that psychoanalytic reading and its techniques are an essential part of the semiotic study. In her introductory text, The Subject of Semiotics, Kaja Silverman [] foregrounded the centrality of subjectivity and psychoanalysis to semiotics. Silverman’s pronouncement that “signification occurs only through discourse, … discourse requires a subject, and … the subject itself is an effect of discourse” (ibid., vii) complements Bourdieu’s notion that an individual’s tastes are constructed from and reflective of location in cultural fields—here ‘field’ and ‘discourse’ are mirror concepts.
3.1 The genre of self-expressive texts

Self-expressive texts are strongly editorial texts in which writers thematize whichever topics they approach through the unifying lenses of viewpoint—including attitudes, dispositions, and tastes. Examples of self-expressive texts suitable are weblog diaries, commentary-rich papers, personal emails, instant message conversations, personal homepages, and social network profiles. The self-expressive genre of texts is identified as the most productive textual input for psychoanalytic reading for two prominent reasons—they are first-person, and they contain redundant evidence.


First, self-expressive texts facilitate the ease of natural language processing tasks necessary for computational reading by being first-person. Third-person texts are constituted by potentially multiple narrative voices, characters, and perspectives which alternate, sometimes even without explicit segmentation. Segmenting and disambiguating between narrative voices is a difficult problem in narrative comprehension—and in fact its contextual requirements seem to require full story understanding. Early attempts at story understanding systems such as Charniak’s (1972) were tenuous, in part because the children’s stories that were chosen as the textual corpus were riddled with narrative shifts between multiple characters. Self-expressive texts, in contrast, are first-person, and can be more successfully assumed to contain, for the most part, the attitudes, dispositions, and tastes of the writer’s own perspective; hence avoiding a hairy segmentation and disambiguation task.
Second, redundant expression is a boon to reading for viewpoint, and self-expressive texts are fraught with redundant evidence of the writer’s judgments about the world. Biographical texts would seem in some respects to be superior sources of viewpoint because they contain crisp and summary propositions about a person’s life experiences and attitudes. However, while the pithiness and focus of biographical texts are appreciable by human readers because they portray a person with archetypal scope, they may actually be a detriment to a machine reader. The reason is that machine readers are shallow readers, and cannot appreciate that some statements are more summary and definitive than others. Since biographical texts are reductive and summary, they contain less redundant information about a person’s viewpoint—ironically, while precision is appreciated by fatigue-sensitive human readers, shallow machine readers such as the psychoanalytic readers implemented for this thesis require redundancies and lengthy textual corpora in order to triangulate upon a correct model of a person’s viewpoint. Self-expressive texts such as a weblog diary or a commentary-rich paper do not explicitly state the writer’s attitude about a particular topic just once, but rather, the attitude toward that topic tints the subtext of several statements, thus creating multiple and redundant evidence. In addition, whereas biographical texts would tend to omit detailed instances of a person’s judgment in favor of more general statements, self-expressive texts are not subject to this editorial pressure. That self-expressive texts find the writer opining over a very broad set of subject matters is a boon to the acquisition of rich models of viewpoint.
In addition to its advantages, self-expressive texts are herewith subject to limitations. Two limitations, the publicity bias, and the performance bias are explained.
Self-expressive texts are limited by a publicity bias because complete candor about one’s views is not always possible due to fear of audience. The sources of self-expressive texts enumerated in the above all anticipate an audience of more than just the writer herself. Composers of weblog diaries and social network profiles may anticipate that friends and co-workers could stumble upon the text. Danah Boyd [], in her ethnography of social network profiles, points out that a social network profile that is at once subject to the audience of people from many different life contexts—such as friends, family, lovers, ex-lovers, and co-workers—is necessarily steered away from complete candor. In fact, it is more likely that a person composing a social network profile will only reveal aspects of self which are universally palatable to friends, lovers, and the boss. Notwithstanding an authorial restraint instilled by anticipation of publicity, a supposition of psychoanalytic reading is that so long as attitudes, judgments, and tastes are insinuated by a text or present in the intratextual unconscious, they might be excavated still.
Performance bias is the idea that self-expressive texts may contain not a single underlying viewpoint, but rather, may manifest viewpoints associated with different personae, according to the social context represented by different audiences. Erving Goffman’s [] thesis in The Presentation of Self in Everyday Life was that social interaction could be likened to a dramatic performance—we are all capable of wear different social masks, and the mask we choose to wear is negotiated by how we fit into varying social contexts such as –the workplace, with friends, with a lover, with family. This limitation, may nonetheless be found acceptable, because while a self-expressive text composed for one’s research and a text composed for one’s friends reveal different aspects of oneself, the two texts will tend to express judgments over different topics—one set related to research, the other set related to social life. According to the aesthetic consistency hypothesis for viewpoint, and supported by the principle of cognitive economy, it would seem reasonable to believe that one’s various masks complement each other as variations on a shared theme rather than wholly contradict one another. It is not a typical case that a self-expressive research paper and a self-expressive weblog diary about everyday activities will express, en masse, directly contradictory attitudes about a shared set of topics. In such cases where contradictory judgments about a particular topic are detected across one individual’s self-expressive texts, the psychoanalytic reader simply concludes insufficient information about that topic.
3.2 Structurations and schemas

How does a self-expressive text communicate the writer’s viewpoint? In the aforementioned, it was suggested that the writer’s attitudes, dispositions, and tastes need not be stated explicitly in proposition form, e.g. “I love classic rock,” but rather, all pieces of a text are tinted with some excavatable trace of viewpoint, e.g. “my friends surprised me with tickets to the Rolling Stones concert.” The strategy of psychoanalytic reading then, is to identify all the topics present across utterances, plus the likely emotive ‘tint’ which underlies each topic in each utterance. By inferring the stable emotive ‘tint’ bound to each topic, and by organizing topics along implied semantic dimensions, the machine reader will have acquired a model of viewpoint, assuming that the sampling of self-expressive texts constituting a personal corpus is sufficiently broad and representative.


The manifestation of viewpoint through textual expressions is best explained through the isotopy model of textual semantics initiated by A.J. Greimas [] and applied to authorial modeling in literary stylistics, cf. Todorov (1968) for a more complete introduction. The model illuminates how an overarching thematic such as viewpoint can be seen as the intratextual unconscious which unifies word-level expressions.
The basic model of isotopy is concerned with the relationships between semes, lexemes, classemes, and isotopies. A lexeme is the formal name given to a linguistic concept – be it a word or noun phrase. Words by themselves are ambiguous; accordingly, each lexeme contains numerous potential word senses, or semes, which are the smallest and most specific units of meaning. When the writer constructs an utterance, she means each lexeme to stand for one particular seme, unless she is engaged in word play. A reader, operating on the pragmatic assumption that a text has a single coherent meaning, will disambiguate all the lexemes into particular semes such that all semes are contextually consistent with one another. Greimas termed this process of convergent reading monosemization. Adding an intermediate structure to the analysis, each seme is capable of a number of contextual entailments, known as classemes. For example the seme of ‘barks’ in the sentence “the dog barks” can be associated with classemes such as [+caninity] [+animalhood], etc. (NB the bracket and plus sign convention for expressing classemes). During monosemization, lexemes are disambiguated into semes such that a system of complementary and mutually consistent classemes are selected. One whole system of classemes is called an isotopy.
Of course, an infinite number of classemes, or contextual themes, are possible. Building upon Greimas’s basic model, literary stylistics has subdivided the space of possible classemes into different realms of concern, known as registers. The most important register with respect to the modeling of authorial viewpoint is identified in stylistics as the emotive register – the stable affective themes which organize semes in the text. Todorov, in his introduction to literary stylistics, insists that “the interpenetration of all these registers in the concrete text cannot be exaggerated” [p26], and explains the pervasiveness of emotive classemes in texts thusly []:
The linguistic forms such “traces” take are numerous, [including] indications as to the attitude of the speaker and/or of the hearer with regard to the discourse or its object (which constitutes only semes, aspects of the meaning of other words). It is by just this means that the total speech-act penetrates all verbal statements: each sentence involves an indication as to its speaker’s dispositions. Someone who says “This book is beautiful” makes a value judgment and thereby introduces himself between the utterance and its referent; but someone who says “This tree is tall” offer a judgment of the same kind, though less obvious.
Todorov raises the interesting suggestion that the utterance “this tree is tall” might advance some an emotive classeme, despite the fact that the lexeme ‘tall’ is not explicitly emotive as ‘beautiful’ is. To be clear, in fact, the utterance could forward various and even contradictory emotive classemes, depending upon its intention and linguistic context. Appearing in the context, “those trees are stumpy, but this tree is tall,” the emotive classeme might be [+reverence], while the context, “I was looking for a short tree, but this tree is tall”, might advance the classeme [+disappointment]. This example illustrates a key understanding in semiotics called the speech-act—ever utterance is underlied by its pragmatic communicative intention, called the illocutionary force, which certainly casts a long shadow onto the emotive register. While the illocutionary force is ambiguous in an isolated phrase, the collocation of many ambiguous utterances causes the convergent interpretation, or, monosemization.
§
Building upon the isotopy model and the related concept of the emotive register, we now consider a structuration for self-expressive texts.
A first assertion about the self-expressive genre is that it is a text that indirectly communicates an image of the writer’s emotive self by constantly juxtaposing the writer against various subject matters. Roman Jakobson’s [] semiotic model of communication distinguishes six functions for textual communication and observes that due to the coherency effect of monosemization, a text always converges upon a primary function. The model conceptualizes six loci in a communication scheme—1) a ‘sender’ pushes 2) a ‘message’ across 3) a ‘channel’ to 4) some ‘receiver’. The message is underlied by 5) some ‘context’, and encrypted with 6) some ‘code’. Jakobson observes that texts foreground one of these loci as primary, and that locus implicates the text’s communicative function. ‘Sender’ implicates the emotive function, ‘message’ implicates the poetic function, ‘channel’ the phatic or channel verification function, ‘receiver’ the conative or imperative function, ‘context’ a referential function, and ‘code’ a meta-lingual or code negotiation function. Operating within Jakobson’s model, while a self-expressive text’s primary locus is the sender and primary function is emotive, the context locus’s referential function is also actively engaged. The broadness of the portrait that is painted about the sender is due to the fact that the sender is passing through a series of diverse contexts or topics, each time forming the message as the resonance between the context and sender. For example, a weblog diary has the diary’s author traveling through the many contexts of everyday happenstance; a social network profile has the author traveling through the various categories of life-style and consumer interests (e.g. books, films, music, food, sports).
The goal of psychoanalytic reading is to collect together the total image of the writer in the emotive register. But herein arises the question, which emotive register? Surely attitudes constitute a different granularity and genre of emotive register than does cultural tastes or psychological dispositions. We are thus led to a second assertion—different sources of self-expressive text betray an image of the writer under different emotive registers, and in viewpoint computation, each viewpoint realm corresponds to a different sort of emotive register.
§
Finally, based on the above discussed structurations, several schemas for self-expressive texts can be presented—one for each of the five viewpoint realms considered in the thesis. These schemas are specific to the knowledge representation of viewpoint spaces and will serve as the bias of the psychoanalytic reader’s mechanism.
Attitudes. Texts such as weblog diaries and commentary-rich papers can be summarized by a list of dominant semes called topic-classemes which are converged upon by the lexemes in the text. Each lexeme is underlied by an inexact emotive valence—its illocutionary force—which can be assessed by the technology of textual affect sensing, to be presented in Section 3.6. Topic-classemes operate as semantic mediators about which stable intratextual meanings cathect. At the most general granularity, the emotive register for texts is comprised of attitude-classemes. An attitude-classeme is comprised of the pair, (topic-classeme, valence), where valence is the stable emotive cathexis about a topic-classeme. To avoid the difficulty of committing to an ontology of emotive valence, Albert Mehrabian’s [] PAD affective unification framework is used to represent emotive valence. Because that model is a three-dimensional Cartesian space, each lexeme’s underlying emotive valence can be plotted as a vector in PAD. An attitude-classeme’s stable valence can thusly be computed using vector mathematics as the first-order moment, or arithmetic mean, of emotive valence vectors underlying the lexemes belonging to the particular topic-classeme. The pattern formed by all the attitude-classemes suggested by a text constitutes the isotopy.
Humor. The humor schema is identical to the attitudes schema, with the exception that a single measure of psychic tension replaces the 3-tuple PAD score in the attitude-classemes of the humor isotopy. The tension that manifests in self-expressive texts is tracked as displeasure, high arousal, and dominance. High arousal plus dominance constitutes the aggression that is emitted by underlying tension. Aggression may either be directed at a person or topic, or it be directed at one self, but it is not necessary to make this disambiguation for purposes here.
Cultural taste. A different sort of emotive register can be described for cultural taste— one that does not utilize the full affect space as with attitudes, but focuses only on positive qualities. Recall that the cultural taste viewpoint realm is represented as a semantic fabric populated by consumerist interests— e.g. books, films, authors, music genres—and aggregated around subcultural identities—e.g. “extreme sports lover,” “intellectual,” “goth.” Because the cultural taste space is constituted by a plenitude of items whose consumption and experience incur time and capital expense, it is less likely that, as with attitudes, every person will have formed judgments about the book “Madam Bovary” or the band “Peaches.” Most persons have formed judgments about all everyday topics in attitude viewpoint space, but the same is not true over all everyday consumerist interests. In fact, it can be hard to judge consumerist interests which have not been experienced first-hand and lies outside the practice of one’s localized life-style. For these reasons, consumer materials are not associated with PAD valences as for attitude-classemes; rather, only positively valent consumer materials are considered, for the pragmatic reason that negative tastes are not often found in corpora such as social network profiles, and homepages. In lieu of a PAD valence, each consumerist interest and subcultural identity identified in the text is associated with an arousal score, which indicates the significance of an interest. Thus, the emotive register is constituted as the set of positively valent interest-semes and subculture-classemes, stable across texts3. The concrete textual materials from which semes and classemes are inferred are interest-lexemes and subculture-lexemes. The lexeme form of an interest are its surface keywords, while the seme form of an interest is the normalized form. For example, the interest-seme ‘rock music’ is to varying extents implied by, inter alia, the interest-lexemes ‘rock music’, ‘zeppelin’, ‘led zeppelin’, and ‘alternative rock’. The stability of interest-semes and subculture-classemes inferred across the corpus of self-expressive texts is defined in mathematical terms as the first-order moment of the arousal scores associated with lexemes that constitute each particular seme. The set of all stable interest-semes and subculture-classemes constitutes one isotopy.
Gustation. The schema for gustatory viewpoint follows the schema of cultural taste, as they share the same knowledge representation. Instead of interest-semes, the gustatory materials include recipe-semes, ingredient-semes, cooking-procedure-semes. Instead of subculture-classemes, there are cuisine-classemes and basic-flavor-classemes.
Perception. Perception’s schema is the most unlike the other four schemas because a writer’s perceptual dispositions affect all textual production, but impress only the faintest of traces upon each individual textual expression. Rather, a clear portrait of dispositions emerges only at the highest levels of a text—overall, is a writer driven by ratiocination or sentimentality? Perceptual disposition belongs truly to a text’s unconscious, and the attempt at its psychoanalytic excavation made by this thesis can only be regarded as a first stab. According to the viewpoint model of perception based on Jung’s four basic psychological functions, a writer can be disposed toward four poles, organized into two oppositions—‘sensing’--‘intuiting’ and ‘thinking’--‘feeling’. Perception lexemes are both textual entities and relationships between textual entities including ego-lexemes which are references to the writer herself, which is underlied by some PAD valence; alters-lexemes which are references to any other person or noun phrase textual entity, also underlied by a PAD valence; incoming-lexemes, which are transfers of emotive energy from alters into ego, and vice versa as outgoing-lexemes. Classemes are more clear—sensing-, intuiting-, thinking-, and feeling- classemes are associated with some stable emotive valence in PAD-space. However, unlike in previous cases, the stable emotive valence of classemes cannot simply be computed as a first-order moment of lexemes, since is it quite complex how lexeme statistics best map to classemes. In order to learn an optimal mapping, machine learning was used to train a classifier which maps from lexemes to classemes. This is discussed thoroughly in Chapter 4.
It should be noted that an alternative structuration and schema for perceptual disposition which is not addressed in this thesis, could be developed upon the basis of personality theory. In Jung’s type psychology, for example, there is a ‘mother archetype’, which could be regarded as a semantic inter-mediator between lexemes and classemes. ‘Mother archetype could be insinuated by mothering lexemes like ‘care’, ‘embrace’, and ‘nurture’, while ‘mother’ could be heuristically mapped into perception space as [+intuiting], [+feeling].
3.3 Mechanics

An algorithm for psychoanalytic reading is now presented. The goal of implemented psychoanalytic readers is to digest a corpus of self-expressive texts into the schemas of the respective viewpoint models. The algorithm has four phases—1) natural language normalization; 2) topic extraction; 3) textual affect sensing; and 4) statistical estimation.


Step #1—natural language normalization. Examples of suitable self-expressive texts were given earlier—weblog diaries, commentary-rich papers, personal emails, instant message conversations, personal homepages, and social network profiles. These are all assumed to be ‘in-the-wild’ English natural language texts—so-called ‘open’ texts. The first step of processing is to prepare texts for computation by natural language normalization, which includes genre-unspecific tasks like tokenization, semantic recognition, part-of-speech tagging, lemmatization, anaphora resolution, phrase chunking, phrase linking, and syntactic frame extraction. Social network profiles have a peculiar format and are thusly associated with a special case normalization process—which is discussed in Chapter 4.
The MontyLingua natural language processor [] is used as the subsystem to perform all normalization tasks—so only a very brief review is given here. Tokenization. An open text is split into word and punctuation tokens. Punctuation marks are assumed distinct tokens, modulo marks used in abbreviations. Contractions such as “can’t”, “shouldn’t” are resolved into “cannot” and “should not”. Semantic recognition. Named-entities must be recognized in preprocessing because they confuse the tagger and chunker. Examples of entities and their normalized forms for named-entities (e.g. “John W. Stewart” ==> $NAME_JOHN_W_STEWART$), temporal expressions (“last Tuesday” ==> $DATETIME_1142022711$), ontological interest items (e.g. “Nietzsche” ==> $BOOKAUTHOR_FRIEDRICH_NIETZSCHE$).

Download 438.05 Kb.

Share with your friends:
1   2   3   4   5   6   7   8



The database is protected by copyright ©ininet.org 2024
send message
    Main page
basic concepts
same time he
computational modeling
structure that