Computing Point-of-View: Modeling and Simulating Judgments of Taste
|
|  Yüklə 8.55 Mb.
|
2.2 Person modeling from everyday textsThe thesis approach to ‘person modeling’ is now described and situated in the user modeling literature. Collaborative filtering and other behavior-based approaches have the virtue of being driven by actual data, but behavioral traces taken to represent a user (e.g. transaction history, browsing history) are typically application-specific—thus these traces constitute context-dependent data portraits of a person. In fact, the name ‘user modeling’ itself reflects a common motivation for such systems—tailoring and personalizing application behavior around its users. GRUNDY (Rich 1979) actually considered persons more generally, since what demographic stereotypes capture are intuitions about cultural patterns of preference. However, such patterns were entered in a priori and not learned from actual cultural data. This thesis extends the data-driven approach of user modeling to what I term person modeling, by considering evidence of persons in the social everyday rather than considering histories of user actions, which are inevitably specific to their application’s domain. To facilitate computation, this domain of the social everyday is further subdivided into realms of concern. The experimental systems described by this thesis explore models of persons within five aesthetical realms—cultural tastes, attitudes, ways of perceiving, taste for food, and sense of humor. As much as possible, the representations of persons in each of these realms is inspired by and consistent with theories of people and cultures known in the humanistic literatures such as psychology, cultural theory, and semiotics. The approach to person modeling espoused here is data-driven, but whence evidence of persons in the social everyday? One plentiful source is the corpus of his/her everyday texts—weblog diaries, social network profiles, homepages, instant messenger conversations, etc. By applying natural language processing techniques to read this text, textual traces are extracted. The parsed and normalized traces are then amenable to machine-learning generalization just like behavioral traces. Reading a person’s texts yields explicit textual traces. But in order to create a predictive person model, these textual traces must be generalized. GRUNDY performed generalization via a priori stereotypes, while DOPPELGANGER generalized users’ interest on the basis of community-maintained models, which underlied each user. Our approach to generalization is to locate a person’s textual traces onto the nodes of a cultural topology that prescribes affinities and relationships between nodes. This cultural topology is not a priori like GRUNDY, but is data-driven like DOPPELGANGER. However, whereas DOPPELGANGER’S community models were neatly represented and directly computable, cultural topology sometimes needs to be mined from natural language text; this represents an additional challenge. The technique of culture mining is described in Chapter 3 as a way to extract cultural topology from cultural corpora, such as 100,000 social network profiles, and political texts of Democrats and Republicans. Once a cultural topology is acquired as a graph structure, generalization proceeds via spreading activation (Collins & Loftus 1975). In the realm of attitudes, analogy (Gentner 1983) and imprimers (Minsky forthcoming) are additional techniques for generalization. Spreading activation, analogy, and imprimers are admittedly heuristic techniques, requiring the manual setting of various parameters like discount for spreading, discount for fan-out, etc. Although spreading activation along cultural topology seems very different from collaborative filtering’s vector-similarity method for generalization, they are not so dissimilar. Mining cultural topology is an information-theoretic process of calculating affinities between cultural items that are true across a population—this procedure can be compared with the clustering effect achieved with item-item collaborative filtering. A final resonance in the user modeling literature is with discourse-based modeling. First, extracting textual traces from a person’s everyday texts can be compared with modeling a user’s discourse history through a conversation with a natural language dialog system, as both tasks concern gisting evidence from text. Second, a person’s textual trace resembles her personal discourse; a culture’s topology resembles wirings in the cultural discourse. Cassell & Bickmore’s (2003) embodied conversational agents located a user’s discourse model onto a spreading activation discourse planner. While the motive of spreading activation there was to generate temporal moves, more generally the act of plotting a user’s topics onto a network of pre-wired topics and spreading activation outward is an act of topological generalization. This same technique is used in the present approach to expand a person’s discourse along the connections of the learned cultural topology. Once a generalized model of a person’s tastes, attitudes, ways of perceiving, etc. is produced, the model is used in applications to simulate a person’s reactions to arbitrary textual input. This ‘reactive’ posture on simulation deviates from the more standard ‘prescriptive’ posture in user modeling, which is that a user model predicts preferences, and these predictions are simply offered up as recommendations. In some sense, asking a recommender system to react to some thing of your choosing could be as strange as bringing some medication up to a pharmacist, and asking her if she thinks you should take that medication. The motivation for using a general model of a person to simulate reactions is to enable a class of what I call perspective-based applications—such as virtual mentors and pundits that constantly react to a user’s context. Strategies for generating reactions include spreading activation and memory-based reasoning. Having situated the thesis’s approach to person modeling in the user modeling literature, the next three sections will detail the three phases of the described approach—acquire, generalize, and apply—focusing on how phases are operationalized for each of the five aesthetical realms that were modeled. |