Computing Point-of-View: Modeling and Simulating Judgments of Taste
|
|  Yüklə 8.55 Mb.
|
2 Techniques & TechnologiesThis chapter presents the techniques and technologies that are central to person modeling from everyday texts. Sections 3.1 and 3.2 present reading for affective themes, and culture mining—situating these techniques in their respective related works. Sections 3.3 and 3.4 articulate two technologies that are core to this thesis—commonsense reasoning, and textual affect analysis—they are also compared to related works. 3.1 Reading for affective themesReading for Affective Themes (RATE) is developed in this thesis as the means of acquiring models of—a person’s tastes, attitudes, ways of perceiving, taste for food, and sense of humor—from his/her everyday texts, such as weblog diaries, social network profiles, homepages, emails, etc. RATE is an associative reader that employs lightweight natural language processing techniques, a knowledge-based topic spotting skimmer, and a knowledge-based textual affect skimmer. Corresponding to each realm is a reading schema specific to that realm’s person model. A reading schema dictates how the outputs of skimmers should be translated into a realm’s model. For example, the reading schema for the realm of attitudes is most direct—a person’s everyday texts are skimmed for topics and affects, and each topic is associated with the affective contexts that underlied its instances in the text. Using statistical estimation over a person’s entire corpus, stable affect values can be bound to the topics14—these constitute the affective themes gisted from the person’s text that then become input to the model generalization phase. The rest of this section gives systematic presentation to the RATE technique—1) related work in computational reading is discussed, and the RATE technique is compared to this literature; 2) the genre of first-person, everyday texts from which this modeling work proceeds, is characterized; 3) schemas specific to reading for each of the five aesthetical realms are presented; 4) an algorithm for RATE is narrated. § Related work in computational reading. Reading for affective themes builds on related work in the interrelated literatures of computational reading, story understanding, text summarization, and structuralist models of reading. RATE’s textual affect analysis aspect is expansive enough to warrant its own survey of related work—which will be given in Section 3.4 to accompany discussion of the textual affect analysis technology. Formal and knowledge-based story understanding. Understanding stories is a classic modeling problem in artificial intelligence with a long history and vast literature. Some early story understanding systems (Charniak 1972; Charniak 1977; Cullingford 1978; Wilensky 1978; Carbonell 1979; Dyer 1983) modeled stories formally using elaborate hand-crafted knowledge in the form of predicate logic rules, frames, scripts, plans and goals, which could achieve in-depth understanding, but were too brittle to handle most stories found ”in-the-wild.” To improve robustness, a more recent incarnation of knowledge-based modeling employed a very large collection of commonsense knowledge to support inference making over stories (Mueller 1998). Memory-based readers. Most story understanding systems assumed that understanding was objective, but some systems adopted a more cognitive posture. IPP (Lebowitz 1980) and CYRUS (Kolodner 1984) understood stories using episodic memories, and IPP was capable of generalizing episodes. These memory-based approaches were bolstered by Haberlandt et al.’s (1980) experimental finding that human readers naturally projected episodic structure onto stories. Shifting from objective understanding to subjective reading, Moorman and Ram’s (1994) ISAAC was a creative reader that could focus, attend, and willfully suspend disbelief. AQUA (Ram 1994) could interleave and motivate reading with asking of questions. ISAAC and AQUA represent segues from story understanding into computational reading. Rather than regarding stories as containing truths to be recognized, computational readers skim text, constructed situation models (Zwaan & Radavansky 1998) to explain the text, and revise and consolidate these models to maintain consistency. Information-extraction readers. Some other systems were also backed by frames and scripts, but used them to instead skim texts, leading to a more opportunistic, information retrieval approach to text understanding. FRUMP (DeJong 1979), for example, used semantic frames and case patterns to skim summaries from new stories. Plot units (Lehnert 1982) was another plan for narrative summarization. More recently, FERRET (Mauldin 1991) improved upon FRUMP’s pattern-based approach. SpinDoctor (Sack 1994) used patterns sorted by ideology (e.g. (?criminal murdered ?victim) and adopted actor-role schema from Greimasian to summarize the ideology implied by news stories. Probabilistic and statistical readers. Another class of story understanding systems incorporated connectionist networks to support story comprehension (Charniak 1986; Dolan 1989; Lange & Dyer 1989; Miikkulainen & Dyer 1991; St. John 1992; Langston et al. 1995). The probabilistic aspect of these systems softened symbolic constraints and this in turn improved the handling of contextual issues in story processing. Recent work (Halpin 2003) makes use of both fully symbolic and also fully statistical approaches such as Latent Semantic Analysis to analyze story plot. The role of affect and perspective in stories. The importance of affect as a parameter and organizer of cognition in humans and machines has been argued for (Simon 1967; Sloman & Croucher 1981). Tan (1994) conceived a model of story processing based on the premise of stories as super-episodes containing many emotion episodes. In story understanding systems, formal symbolic accounts of affect in relation to modeling characters and plot were presented in (Lehnert 1982; Dyer 1983; Elliott & Ortony 1992). Formal models of the related notion of narrative perspective were presented in (Wiebe & Rapaport 1988; Wiebe 1994). Reading for affective themes (RATE). Reading for affective themes mixes the approach of information-extraction readers, with the approach of knowledge-based story understanding. At the large scale of reading, RATE takes an information extraction and statistical approach to estimating statistically salient topics and affects, but at the small scale of reading, it uses knowledge-based component technologies. Rather than attempting in-depth understanding achieved by formal story models, RATE privileges recall and robustness over depth. Thus, it gathers textual evidence for affective themes at the sentence-level, stores the evidence in a memory database, and uses statistical methods to estimate average values and salience for affective themes using the memory base. RATE is also associative. For example, in the realm of attitudes, an attitude is a topic plus a statistically summarized affect value. A topic’s stable affect value is measured by statistical estimation over all the affect values that co-occurred with instances of the topic in the text. This associative method for attributing affect to topics can be compared to work on learning the meaning of words from their surrounding context (Berwick 1989; Cardie 1993), and Yarowsky’s (1992) statistical method for word-sense disambiguation. A difference is that RATE applies statistical disambiguation not to words but instead to thematic abstractions (topic and affect), which were themselves inferred from the text. To adapt RATE so that it can populate person models for all five realms, five reading schemas are used, which prescribe how the outputs of the text skimmers are to populate the various realms’ models. The idea of reading based on different schemas can be compared to situation models for reading (Zwaan & Radavansky 1998), and to actor-role analysis (Sack 1994). At the small scale of reading, a RATE reader skims text dually with a topic-spotting skimmer, and with a textual affect skimmer. These skimmers take a knowledge-based approach that may be compared with (Mueller 1998). The topic spotter belongs to ConceptNet (Liu & Singh 2004), which makes use of a large knowledge base of common sense relations to annotate passages of text with inferences, and then triangulates on topics that emerge in the space of inferences. The textual affect analyzer also uses a component, Emotus Ponens (Liu, Lieberman & Selker 2003) that makes commonsense inferences about a text’s affective qualities based on its event structures. Having compared RATE with related approaches in computational reading, we should keep in mind that RATE’s approach is not being considered for arbitrary story texts. RATE is considered only for the genre of first-person, everyday texts from which the taste and attitude modeling work proceeds. Next, we characterize the properties, assumptions, and limitations of this genre of text. § The genre of everyday texts. Consider weblog diaries that allude to the same repertoire of friends and situations day after day, entry after entry—each time giving a slightly different take, but always seeming to recapitulate a theme. Everyday texts are the target of our person modeling. Examples of suitable everyday texts are weblog diaries, commentary-rich papers, personal emails, instant messenger conversations, personal homepages, and social network profiles. We define everyday texts as the genre of first-person texts that we author informally and as part of day-to-day life. As such, these texts are semi-private (they are certainly not formal) and they are self-expressive—they should have a strong editorial quality because they were written with free expression of the author’s opinions about all the things that he may have encountered in everyday happenstance. Everyday texts are also self-expressive because we hypothesize that by modeling tastes and attitudes expressed in them, we have actually modeled their author. Some arguments are made for the suitability of everyday texts as a target corpus for person modeling. First, everyday 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—the complexities of which are exposed in (Wiebe 1994). Early attempts at story understanding systems such as Charniak’s (1972) were tenuous as scenarios, in part because the children’s stories that were chosen as the textual corpus were riddled with narrative shifts between multiple characters. Everyday texts, in contrast, are first-person, and can be more successfully assumed to contain, for the most part, the attitudes, dispositions, and tastes constituting the writer’s own perspective; hence avoiding a hairy segmentation and disambiguation task. Second, redundant expression is a boon to reading for affective themes. Everyday texts are fraught with redundant evidence of the writer’s judgments about the world. Biographical texts would seem in some respects to be a superior source for person modeling because they contain crisp and summary propositions about a person’s life experiences and attitudes. Ironically, while precision is appreciated by fatigue-sensitive human readers, shallow machine readers such as the RATE readers implemented for this thesis require redundancies and lengthy textual corpora in order to triangulate upon a correct model of a person. Everyday 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. For example, consider weblog diaries that allude to the same repertoire of friends and situations day after day, entry after entry—each time giving a slightly different take, but always seeming to recapitulate a theme. In addition, whereas biographical texts often omit detailed instances of a person’s judgment in favor of more general statements, everyday texts are not subject to this editorial pressure. That everyday texts find the writer opining over a very broad set of subject matters is a boon to the acquisition of rich person models. In contrast to its advantages, everyday texts are herewith subject to two limitations—the publicity bias and the performance bias—as explained below. Everyday texts are limited by a publicity bias because complete candor about one’s views is not always possible due to anxiety of audience and anticipatory self-censoring. The sources of everyday texts enumerated in the above all anticipate an audience of more than just the writer herself. Authors of weblog diaries and social network profiles may anticipate that friends and co-workers could stumble upon the text. Boyd (2004), 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 that are universally palatable to friends, lovers, and the boss; thus person models acquired from texts affected by these biases will capture who a person portrays herself as, not necessary the “real” person that they really are. Notwithstanding authorial restraint instilled by anticipation of publicity, a supposition of RATE is that so long as attitudes, tastes, and ways of perceiving are insinuated by a text or present in the intratextual unconscious, there is hope for their indirect excavation. Performance bias is the idea that everyday texts may contain not a single underlying perspective, but rather, may manifest perspectives associated with different personae, according to the social context represented by different audiences. Goffman’s (1959) 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 wearing 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 an everyday text composed for one’s research and a text composed for one’s friends reveal different aspects of oneself, the two texts tend to express judgments over different topics—one set related to research, the other set related to social life. In such cases where contradictory affects about the same topic are detected across one individual’s everyday texts, RATE’s statistical estimation process would simply cause the contradictory affects to cancel out. § Reading schemas. Reading for affective themes employs two skimmers. A topic-spotting skimmer returns a scored list of topics for textual passages of arbitrary length. A textual affect skimmer returns a PAD affect value for textual passages of arbitrary length. However, the raw outputs of these two skimmers still need to be mapped into the five realm models, and their inter-relationship is not yet specified. To operationalize the mapping, five reading schemas are now presented. But first, a uniform vocabulary for describing schemas is desired, and a vernacular that is appropriate to build on is found in Greimas’s (1966) isotopy model of textual interpretation. The isotopy model gives a standard account of how coherent themes can arise out of a collection of textual fragments, which are themselves ambiguous. The model is also consistent with schemes for word sense disambiguation (cf. Yarowsky 1992). A brief review of the isotopy model follows. A simple vocabulary consists of the terms—lexeme, seme, classeme, and isotopy. A lexeme is a fundamental unit of a lexicon—e.g. “Frank Sinatra” is a lexeme in our lexicon of cultural taste. For the reader, the meaning of a lexeme is unconfirmed or ambiguous—that is, each lexeme could have multiple senses of meaning, or semes. But not so for the writer, who had a seme in mind when uttering a lexeme. To discover the seme that was intended, Greimas assumed that a text’s meaning should converge; thus by co-occurring in a text, lexemes mutually disambiguate one another. Greimas termed the strategy of reading for convergence, monosemization. Adding an intermediate structure to the analysis, each seme is capable of a number of higher-level 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. In RATE, lexemes correspond to unconfirmed and ambiguous concepts; semes correspond to significant (confirmed) and disambiguated concepts; classemes correspond to higher-level abstractions such as discourse-level topics. Because the basic isotopy model gives us a useful way to describe how themes are built up, we describe reading schemas by sub-typing its vocabulary. Schemas for our five realms are now presented. Attitudes. The goal of reading for attitudes is to summarize a weblog diary into a list of its topics, each tinted with a stable affect. When the weblog is first parsed, many word-level (‘party’), phrase-level (‘bachelor party’), and event-level (‘throw party’) topic-lexemes will be recognized in the topic spotter’s lexicon (ConceptNet plus topic folksonomy). The topic-spotting skimmer intersects the ambiguous meanings of these topic-lexemes, converging upon a list of topic-classemes. At the same time, each topic-classeme points back to the topic-lexemes that were its supporting evidence. The textual proximity of each topic-lexeme is associated with some unconfirmed affect-lexeme as scored by the affect skimmer, and given as a PAD value. Using statistical estimation, the affect-lexemes that underlied each topic-classeme’s supporting evidence is summarized into an affect-classeme. Finally, the topic-classeme and its affect-classeme are bound into a duple, and this is called an attitude-classeme. The scored list of a text’s attitude-classemes constitutes an isotopy which populates the attitudes person model. Humor. The humor schema is explained as a transformation of the attitudes schema because both have semantic sheets as their target representation. Topic-lexemes, and topic-classemes are retained from attitudes schema. Instead of measuring affect in terms of PAD values, each 3-tuple affect-lexeme value outputted by the affect skimmer is transformed into a unary tension-lexeme value. Tension is scored between 0.0 and 1.0, proportional to the degree to which the incoming PAD resembles a prototype for psychic tension—displeasure, high arousal, and dominance. High arousal plus dominance represents Freud’s notion of tendentiousness. Tension-lexemes are then summarized statistically into tension-classemes and bound to the topic-classemes to produce humor-classemes. The scored list of a text’s humor-classemes constitutes an isotopy which populates the humor person model. Cultural taste. Although the thesis summary only described how cultural tastes were extracted from social network profiles, this reading schema allows cultural tastes to also be extracted from any everyday text such as a weblog diary or homepage. Recall that since the model of cultural taste is represented on a semantic fabric, it does not attach a PAD value for each interest or identity descriptor; hence, PAD will again be transformed into a unary significance-lexeme value that screens for significance of interest—based on PAD’s degree of similarity to the prototype: pleasure, arousal, dominance. A text is parsed and filtered through the cultural taste lexicon, resulting in interest-lexemes (e.g. books, films, authors, music genres) and identity-lexemes (e.g. ‘extreme sports lover’, ‘intellectual’, ‘goth’). Then, the significance-lexemes underlying all identical interest-lexemes are statistically summarized and bound to the interest descriptor to constitute an interest-seme. For example, we find that the television show ‘Fresh Prince of Bel Air’ occurred four times, and all had high significance, thus, ‘French Prince of Bel Air’ becomes an interest-seme, for we have confirmed its significance to the writer. Likewise, the significance-lexemes underlying identical identity-lexemes are statistically summarized, and we call what results is an identity-classeme.15 Finally, the scored list of interest-semes and identity-classemes constitutes the isotopy and populates the cultural taste person model. The stable significance scores associated with each interest and identity determine the initial activation energy of nodes on the taste fabric. Acquiring the cultural taste model from social network profiles did not benefit from this level of nuance. Food. The food schema is explained as a transformation of the cultural taste schema because both have semantic fabrics as their target representation. Significance-lexemes carry over from the cultural taste schema. Instead of interest-lexemes and identity-lexemes, the raw text is parsed into recipe-lexemes, ingredient-lexemes, cooking-procedure-lexemes, sensation-lexemes according to the lexicon of the food fabric—and these are associated with their significances and summarized into their corresponding semes. Furthermore, because an extensive metadata system annotated recipes with their cuisine type and basic flavor types, the topic-spotting skimmer then builds up confirmed recipe-semes, ingredient-semes, etc. into cuisine-classemes and basic-flavor-classemes, for an additional level of unification. Finally, all the resulting semes and classemes constitute the isotopy. Perception. The perception schema is the most unlike the other four schemas. Whereas cultural taste schema tracked a very large lexicon of interests and identities, perception schema only tracks—ego-lexemes (occurrences of ‘I’, ‘me’, or ‘my *’), alters-lexemes (any syntactic subject or object that is not an ego-lexeme), action-lexemes (verbs), incoming-lexemes (any sentence frame whose agent is an alters-lexeme and whose patient is an ego-lexeme), outgoing-lexemes (any sentence frame whose agent is an ego-lexeme and patient is an alters-lexeme), and mental-lexemes (invocations of thought states, e.g. ‘I thought that’). The affect skimmer output PAD affects for each of these lexemes. The PAD affect of an incoming or outgoing lexeme is based on the PAD of the agent, action, and patient—this calculation is complex and is covered in Chapter 4. Finally, classemes are produced from these lexemes, as the statistical summary of their associated PAD values. Two exceptions are that action-lexemes produce the ratio-classeme, which gives the statistical ratio of introverted to extroverted actions; and mental-lexemes produce the mental-classeme, which estimates the density of thought state propositions in the text. These classemes constitute a perception isotopy, and are later fed into a learned classifier that yields their location in the Jungian dimensional space. Five reading schemas were presented to prescribe how the output of text skimmers should be used to populate each realm’s model. Using these schemas, we achieve a generality for acquisition—for example, we note that cultural taste can be acquired not only from social network profiles, but using its schema it can be acquired from weblog diaries or homepages. Next, we narrate a four-step algorithm for RATE that uses these schemas. § An algorithm for RATE is now presented. The goal of implemented RATE readers is to digest a corpus of everyday texts into the schemas of the five realms. This algorithm is applicable to weblog diaries, homepages, and other free-form everyday texts; this discussion will make less sense for social network profiles, which are already semi-structured, though it is not incompatible with those texts. The algorithm has four phases—1) natural language normalization; 2) topic skimming; 3) affect skimming; and 4) statistical summarization. Step 1—natural language normalization. The first step of processing is to prepare texts for computation by natural language normalization, which includes tasks like tokenization, semantic recognition, part-of-speech tagging, lemmatization, anaphora resolution, phrase chunking, phrase linking, and syntactic frame extraction. The MontyLingua natural language processor (Liu 2002) is used as the subsystem to perform these normalization tasks—so only a very brief review is given here.
“John/NNP likes/VBZ to/TO play/VB board/NN games/NNS” ==> (NX John NX) (VX likes to play VX) (NX board games NX)
(NX John NX) (VX robbed VX) (NX the bank NX) with (NX the money) ==> (John (robbed
(the bank (with (the money)) Note that “the money” was linked to “the bank” and was not instead implicated as the second argument to the verb “robbed”. This mechanism makes the common sense assumption that “the money” was not the instrument used by John to perform the robbery, though such a scenario is certainly possible. More challenging cases of linking arise in the encounter of surface movement phenomena such as subject-verb inversion (e.g. “have you the money?”), topicalization (e.g. “to the bank I go”), and passive voice (e.g. the subject is nested in an agentive “by” phrase in the utterance “The bank was robbed by John”).
SENTENCE FRAME [verb] “robbed” [subject] “John” [direct object] “the bank” “with the money” A sentence frame may contain any number of direct and indirect objects. A sentence frame is constructed for each clause in the text. A dependent clause has a frame which is linked to the frame of the clause upon which it depends. Step 2—topic skimming. After open texts are normalized into phrases, events, and lexical items, the topic-spotting skimmer is run, according to the five specific reading schemas. Topic extraction is performed using ConceptNet common sense reasoning toolkit, but augmented with various folksonomies in order to gist topics for realms with custom lexicons like cultural taste realm. The ConceptNet topic extraction mechanism is presented in Section 3.3. Step 3—affect skimming. A mechanism for textual affect analysis scans over text and annotates it—at the various granularities of phrase, sentence, paragraph, and document—with its affect valence score, given in the Pleasure-Arousal-Dominance format of Mehrabian (1995b). The reading schema prescribes the particular granularity that is used. Some schemas also prescribe a transformation of PAD into measures of tension or significance. The affect skimmer is constituted from knowledge of both—1) lexical affect—a database of sentiment words (e.g. ‘cry’, ‘sad’) annotated with affective valence plus a database of non-sentiment words (e.g. ‘homework’, ‘recess’) annotated with their typical affective connotation as measured in psychological focus groups; and 2) event-centric affect—a database of commonsense knowledge affording typical affects for everyday concepts and events (e.g. ‘be(person,fired)’). By combining lexical and eventual treatments, affect can be sensed as a more robust combination of both surface language and deep meaning. Section 3.4 presents details for textual affect analysis technology. Step 4—statistical summarization. Having transformed a corpus of raw everyday texts into normalized concepts, topics, and affects; affects are grouped by the concepts and topics they underlie, and are statistically summarized, according to the needs of each reading schema, using heuristically determined statistics. For example, to gist stable attitudes for the attitudes realm, a reinforcement statistic (details in Chapter 4) is used in order to simulate a reflexive memory. The reinforcement statistic has rather low tolerance for topics having contradictory PAD values—in which case a neutral PAD is learned. A more forgiving statistic, the first-order moment, is used to stabilize affects in the schemas of cultural taste, humor, food, and perception. Recall that the goal of monosemization is to converge upon stable themes. These statistics, by definition, converge. One shortcoming of statistical summarization is that it cannot handle certain texts, such as jokes and sarcastic passages, whose action is to provoke ruptures of isotopy, as Greimas called it. For example, a joke narrative leads a reader to converge upon a false isotopy, only to have that isotopy overturned by a punch-line, which forces systematic re-semization of the text, resulting in the actual isotopy. Statistical convergence certainly would draw wrong conclusions from such texts, as a statistical summarizer is unable to recognize the rupture, and unable to selectively overturn things that it has already read. So, it is assumed that sarcasm and jokes within the chosen everyday texts are limited to isolated episodes within the textual corpus. The statistical estimation mechanism can tolerate a certain amount of these episodes as ‘noise’ in the text. In this section, the ‘reading for affective themes’ (RATE) technique was presented. RATE was first articulated against a backdrop of previous work in computational reading—finding resonance with previous approaches in statistical reading, knowledge-based reading, and memory-based reading. Second, the genre of everyday texts to which RATE applies was characterized as first-person and self-expressive. Third, we described five reading schemas which specify how the outputs of RATE’S topic skimmer and affect skimmer should be mapped into person models for the five realms. Finally, an algorithm was given. Whereas RATE can acquire a person’s textual traces, creating a generalized person model demands that we interpret a person’s traces as articulations against a backdrop of cultural patterns. The next section introduces a technique that models the backdrop of cultural patterns. |