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

səhifə	3/28
tarix	26.06.2016
ölçüsü	8.55 Mb.

1 2 3 4 5 6 7 8 9 ... 28

1.2 Roadmap and contributions

This thesis is fueled by a series of experimental systems that were built to model persons from their everyday texts, within five aesthetical realms—their cultural tastes, their attitudes, their ways of perceiving, their taste for food, and their sense of humor. Supported by evaluations and implemented applications, three broadly stated contributions of the thesis are:

Through several parallel lines of investigation (via five aesthetic realms), it was demonstrated that rich ‘person models’ can be built by reading a person’s everyday texts—her so-called ‘textual traces’ (cf. ‘behavioral traces’)—and can be generalized according to the underlying topology of the cultural space via spreading activation, analogy, and imprimer supplementation.
Textual affect analysis was developed to implement an associative reading for affective themes. This technique complements previously used ‘topic spotting’ techniques for attitude mining. By focusing on first-person everyday texts like weblog diaries and social network profiles, the difficult task of viewpoint attribution was obviated, and it could be assumed that the text’s emergent affective themes were the writer’s own attitudes, tastes, etc. The strengths and limitations of this reading approach were made evident in three model evaluations.
Six perspective-based applications were implemented, demonstrating a range of applications entailed by the thesis’s approach to person modeling—they constitute tools for self-reflection, person learning, and deep customization. The utility of at least one of these applications in a person-learning task was strongly supported by user study.

The rest of the thesis is structured as follows.

Chapter 2 presents the thesis’s approach to modeling a person’s tastes in the aesthetical realms, structuring discussion into three prongs of ‘acquire’, ‘generalize’, and ‘apply’. Related work in User Modeling prefaces introduction of the thesis’s approach. Related work in Intelligent User Interfaces accompanies discussion for ‘apply’.
Chapter 3 presents techniques and technologies for person modeling from everyday texts. Two techniques—‘reading for affective themes’, and ‘culture mining’ are discussed. ‘Reading for affective themes’ is situated in related work on computational reading. ‘Culture mining’ is situated in related work in musical similarity, social network analysis, and text mining. Two key technologies—‘commonsense reasoning’ and ‘textual affect analysis’—are presented and situated in their respective related works.
Chapter 4 narrates implementation-level details for each of the five built acquisition systems—cultural tastes, attitudes, ways of perceiving, taste for food, and sense of humor. Evaluations for three of these systems—cultural tastes, attitudes, and ways of perceiving—are presented in full.
Chapter 5 presents six perspective-based applications, which are enabled by person modeling. 1) an art bot which creates art suited to a person’s ways of perceiving the world; 2) a kiosk that facilitates serendipitous social introductions based on shared tastes; 3) a textual mirror to support self-reflection; 4) virtual mentors and pundits which give just-in-time feedback; 5) a synesthetic cookbook which simulates the tastebuds of family mentors; and 6) a jocular companion to parlay everyday woes into opportunities for humor. Evaluations for two of these applications are reported.
Chapter 6 reflects upon the philosophical underpinnings of the computational theory and methodology presented in this thesis, offering a theory-inclined digestion of the presented material, and betraying a direction for future work.
The thesis concludes in Chapter 7.

1 Approach

This chapter details an approach to modeling persons as they are in the social everyday, and parlaying these models into a range of perspective-based applications. Examining evidence of personal self-expression in everyday texts enables modeling in this most general of domains, and distinguishes the character of the present investigation from related works in the user modeling literature.

Section 2.1 discusses seminal and related works in the user modeling literature. Section 2.2 states this thesis’s approach, finding situation for the present work in the user modeling literature. Sections 2.3-2.5 entertains separate discussions and related works for the three prongs of the approach—acquire, generalize, apply.^⁸

2.1 Related work in user modeling

The field of user modeling is mature and over 25 years old. Because user modeling research is usually motivated by improvement of application performance, some closely related literatures which have themselves become developed are recommender systems, user-adaptive systems, and adaptive hypermedia. All the while, the user modeling field has also maintained a methodological wing focusing on generic techniques, and ‘user modeling shells’—backbone systems which can be populated with domain-specific data and rules. Numerous proper surveys of user modeling can be found elsewhere—Kobsa (2001) reviews user modeling shell systems through 2000; Brusilovsky (2001) betrays the related field of adaptive hypermedia; a Communications of the ACM special issue (Resnick & Varian 1997) surveyed the related field of recommender systems. What follows is certainly an improper and incomplete survey—particular systems and themes are focused upon in order to invite comparison to specific themes in the literature.
From the viewpoint of the present work, user modeling can be divided into two approaches—category-based modeling, and behavior-based modeling. A third approach—discourse-based modeling—is also touched upon, though the area is still a relatively recent development resulting from cross-over research in corpus-based computational linguistics.
Category-based modeling is of interest because knowing the category of someone affords generalization of a user’s model based on properties of the category. Though it was one of the first works in user modeling, Rich’s (1979) book recommender system, GRUNDY, is still a par excellence example of category-based modeling. GRUNDY consisted of a user profile acquisition component, and a generalization and recommendation component. A profile of a user’s demographic characteristics was acquired by a questionnaire-style interview. The generalization component consisted of a priori stereotype rules, mapping demographic categories into book preferences associated with those categories. The generalization algorithm consists of firing rules activated by the user’s profile in order to produce a generalized user model. For example, suppose that the category of women 26-35 years old stereotypically prefers “romance novels” with strength S. If a user’s profile implicates her into that category with uncertainty U, then a preference for “romance novels” is added to the user’s generalized model, with strength S and uncertainty U. By iterating through other attributes in the user’s profile, these preferences began to add up and overlap, eventually, the generalized user model should triangulate or converge onto several key recommendations.
GRUNDY’s approach resembled and paralleled developments in AI expert systems, as GRUNDY was essentially a system of hand-crafted rules backed by an uncertainty model, and was purposed for recommendation. Some subsequent user modeling shell systems formalized the stereotype and rule idea to allow user models to be assembled for any domain, by populating a rule base. GUMS (Finin & Drager 1986) and UMT (Brajnik & Tasso 1994) were two shells that allowed for the definition of hierarchical stereotypes and mapping rules, thus circumscribing the scope of GRUNDY. As both took a logic-based approach, they incorporated truth maintenance capabilities. The category-based approach of GRUNDY and later related modeling shells is a sound one, but the quality of models built from a priori stereotypes depend entirely on the insightfulness and profundity of those hand-crafted stereotypes.
Behavior-based modeling is interesting as an approach because it states that user models should be driven by actual ‘behavioral traces’ produced by the user, and not the result of a priori assumptions. Hence, user preferences are not deduced from stereotypes but are rather induced from a history of user actions. Social information filtering (Shardanand & Maes 1995), or collaborative filtering, as is the most preferred term today, is a popular algorithm for predicting user preferences. In that algorithm, a user’s history of actions (e.g. purchase history, browsing history) is represented as a high-dimensional vector of ratings over a field of items (e.g. products, webpages). By comparing one user’s vector with the vectors of all other users, mathematical metrics such as cosine similarity are used to find similar users and potentially desirable yet undiscovered items. User-user collaborative filtering has also been varied as item-item collaborative filtering (Sarwar et al. 2001), which is more commonly used in e-commerce applications such as Amazon’s^⁹ product recommendation mechanism. Today, there are several recommender toolkits built on variations of collaborative filtering, such as GroupLens (Breese et al. 1998; Herlocker et al. 1999), which supports predictive modeling from various behavioral traces such as the user’s profile elicited via online forms, navigation history, and transaction history.
Collaborative filtering is one popular technique for behavior-based modeling. Another technique is Bayesian modeling. One classic example, the Lumiere project (Horvitz et al. 1998) used Bayesian modeling over histories of users’ actions (and contexts) in the Microsoft Excel spreadsheet application to infer likely user goals. A key difference between collaborative filtering and Bayesian learning is that the Bayesian approach excels in complex decision spaces where the presence of contextual features changes drastically the interpretation of user actions. Collaborative filtering presumes that features are more or less of a homogenous class, and are not conditioned on one another. Other than Bayesian learning, a variety of other machine learning and statistical techniques may also be applicable—many are surveyed in (Zukerman & Albrecht 2001). For instance, the DOPPELGANGER user modeling shell system (Orwant 1995) allowed for interchangeable machine learning algorithms such as Markov models, clustering, and linear prediction, to predict current user state (e.g. ‘hacking’, ‘writing’, ‘idle’) from hardware and software sensors. However, DOPPELGANGER, along with knowledge-based intelligent tutoring systems such as COACH (Selker 1994) also exhibit qualities of the earlier described approach of category-based generalization. Each user’s profile of interests was generalized by the models of user communities that a user belonged to—such models were community-maintained. Significantly, DOPPELGANGER’S category-based models were not a priori, but were rather driven by community data.
A third approach for user modeling is assigned a working term here—discourse-based modeling. Discourse-based user modeling is likely the result of cross-over research from corpus-based computational linguistics and artificial intelligence, especially dialogue systems. One survey of research that lies between user modeling and natural language processing systems is presented in (Zukerman & Litman 2001). A discourse model is a model of the field of topics that are either covered by a conversation, or is a semantic field defining the vernacular of a domain. The former sense of the term is relevant to natural language dialog systems, in which users are modeled as their accrued topics and goal states from conversation with the system. The latter sense of the term has been used system-centrically in AI and HCI systems to describe a system’s field of allowable syntax and semantics, e.g. (Wahlster 1991); in other words, the field of possibility. However, discourse models can also exist for entities larger than a user’s conversation, and smaller than the total space of possibility; for example, communities and cultures have preferred discourses. Knowing the discourse models of several communities, by comparing the user’s discourse against those community discourse models, it is possible to characterize the user as inheriting from certain communities. This then becomes a basis for generalization. Cassell & Bickmore (2003) described embodied conversational agents that modeled a user’s discourse as breadth and depth of familiarity over a range of traversed topics (e.g. weather, baseball). This user discourse model then activated nodes in a spreading activation discourse planner, and a number of next conversation steps were calculated. One way to think about the spreading activation planner is that it has captured the common sense discourse of how conversation topics are related to one another. By plotting a user’s discourse onto the planner’s topology and spreading activation, a generalization is made.

1 2 3 4 5 6 7 8 9 ... 28