January

Businesss Meeting

Back to Top

TBA

Host: Bill Rapaport

Back to Top

Ann Bisantz , Department of Industrial and Systems Engineering, University at Buffalo

Assessment of Display Attributes

for Displaying Meta-Information

In many domains, operators need to understand and act on large volumes of information from a variety of sources. Operators are particularly challenged by the need to reason about the qualifiers of that information.For example, in military command and control tasks, commanders must reason about the location of threats. Information about those threats may come from sensors with associated uncertainties, be several hours (or days) old, and/or be derived from intelligence sources with varying degrees of trustworthiness. These critical data qualifiers are generally not presented, or are not incorporated into the primary information displays used by commanders. To date, we have conducted three studies which have investigated participants’ ability to rank order, rate, and anchor endpoints of graphical representations with respect to levels of meta-information. Across the experiments, it was clear that participants were able to consistently order and rate the representations of meta-information; but that the order direction could be influenced based on the task framing as well as (in one experiment) the type of meta-information. In particular, it appeared that there was a tendency (in the last experiment) for participants to select an order direction so that the level of meat-information that was most task relevant would be assigned a representation that had the greatest contrast with the background. The results from these studies suggest both that graphical characteristics such as color characteristics may be useful tools in representing meta-information on maps; however, designers cannot take the direction of the mapping for granted: instead, they should carefully consider their choice of representation levels according to the context of the task, and value or meaning of the meta-information being presented.

Host: Bill Rapaport

Back to Top

Brian Cantwell Smith, Faculty of Information Studies, University of Toronto

Rehabilitating Representation

Representation has come in for a bad rap, in recent decades. In this talk I will argue for a radical reconstruction of (something like) representation, in order to: (i) capture what was right and powerful in cognitive science's original representationalist approach -- something that I believe we decry at our peril; (ii) avoid what was wrong -- perhaps even fatal -- about the classical conception; and (iii) do justice to the flood of intuitions that motivate situated, embodied, embedded, social, and even post-structuralist alternatives.

Host: Bill Rapaport

Back to Top

Roelant Ossewaarde, Department of Linguistics, University at Buffalo

Incorporating an Instance-data

in an Ontology Alignment System

How to Make Use of What's Already There

Automatic alignment of ontologies with different structures is a well-known difficult problem in practical applications. If one ontology contains an assertion 'rowing is-a-kind-of sport', and another contains the assertion 'crew is-a-kind-of sport', how can we reason that 'rowing' and 'crew' are related, even identical?

With the availability of large-scale annotated ontologies on the rise, it becomes feasible to combine multiple sources of information as input to the reasoning process. One such information source may be the objects annotated by an ontology. For example, given a set of books categorized under the heading 'rowing', if a Dutch library categorizes the same books under heading 'roeien', one may conclude with some confidence that 'roeien' and 'rowing' are related, even without any knowledge about the structure of the ontology or about the words themselves.

My paper presents an attempt to incorporate information about such annotated instances in a evaluation initiative that compares different kinds of approaches to ontology alignment. I will show data that indicates that even very naive instance-based ontology alignment systems can be competitive with actual traditional systems that just compare the structure of ontologies or word-similarities.

Host: Jean-Pierre Koenig

Back to Top

Florian Jaeger, Brain & Cognitive Sciences, University of Rochester

Speakers make optimal choices
during language production

Theories of rational cognition (e.g. Anderson 1990; Simon 1990) consider the brain to be close to optimal in terms of efficient information processing. More recently, the idea of optimal information processing has also been applied to language processing (but see recent work by Aylett & Turk 2004; Genzel & Charniak 2002; Hale 2001; Jaeger 2006; Levy 2006). I present a series of studies that support the hypothesis that speakers make optimal choices (within the bounds defined by grammar) when formulating their utterances.

Most of the talk will focus on the link between redundancy and reduction in language production. As speakers encode their intended message into an utterance, they frequently choose between several ways to convey that message, including decision between full and reduced forms at many levels of linguistic processing:

(1) Phonetic & phonological reduction: e.g. full vs. weak vowels, t/d-deletion, syllable omission
(2) Morphosyntactic reduction
I'm ... vs. I am ...
He wouldn't ... vs. He would not ...
(3) Syntactic reduction
He thinks I am a looser. vs. He believes that I am a looser.
The guy known to be ... vs. The guy who is known to be ...

It follows from information theoretic considerations that optimal speakers should prefer choices that keep the amount of information¹ conveyed per unit uniform (the hypothesis of Uniform Information Density, Jaeger 2006; Levy & Jaeger 2006). For reducible structures (e.g. (1-3)), Uniform Information Density predicts that speakers are more likely to choose the full form (e.g. am) rather than the reduced form (e.g. Ôm) the more information the reducible unit contains. I present evidence for Uniform Information Density from corpus studies on syntactic and morpho-syntactic reduction in spontaneous speech. I conclude that the computational system underlying language production is - at least to some extent - optimized.

I will also present preliminary results from an ongoing study on the role of information content in speakers' planning beyond the level of the clause. The results provide evidence that (a) speakers prefer to split messages that contain a lot of information into several clauses, keeping the amount of information per clause relatively uniform, and (b) speakers
make this decision early during utterance planning (prior to linearization).

If time permits, I will close by presenting data from an ongoing production experiment investigating the consequences of apparently 'suboptimal' choices in sentence production.

¹Information theory defines the Shannon information content of a unit (e.g. a word) as the logarithm of the inverse of the probability of that unit in its context (Shannon 1948). This makes it possible to estimate the information of a word (e.g. from corpora or closes studies).

Host: Jean-Pierre Koenig

Back to Top

Barry Smith, Department of Philosophy, University at Buffalo

Towards Ontology of Science

The Star Trek Prime Directive asserts that it is the right of each sentient species to live in accordance with its normal cultural evolution; no Star Fleet personnel may interfere with the healthy development of alien life and culture. Such interference includes the introduction of superior knowledge, strength, or technology to a world whose society is incapable of handling such advantages wisely. In philosophy, this Prime Directive translates into the rule that philosophers of science should not interfere with the work of the scientists themselves. Rather, they should observe, comment and theorize from the outside. I will argue against this philosophical Prime Directive, and show how it is already being violated.

Hosts: Bill Rapaport

Back to Top

Peter Pfordresher, Department of Psychology, University at Buffalo

Modeling Retrieval in Sequence Production

When people produce long, complex action sequences they simultaneously engage in memory retrieval and production. Due to these task demands, and basic limitations of working memory, all sequence events are not retrieved simultaneously. At the same time, due to the temporal demands of production events are not retrieved one at a time. Instead, producers have access to a subset of sequence events at any point in time. I will describe a model that characterizes the sequence events that are accessible in production as the joint product of working memory limits and event-based similarity relationships. The model predicts patterns of serial ordering errors made in production that arise from competing activations that occur in the course of retrieval. To date the model's predictions have been tested on the production of musical sequences, though some data suggest that its predictions may generalize to speech. The model predicts working memory constraints based on production rate that lead to a speed accuracy trade-off that has been tested in a recent experiment. I will discuss these previous fits of the model and will also propose future directions that include comparisons between different similarity metrics, differential weighting of past and future events, and the possible inclusion of a feedback component to the model.

Host:Bill Rapaport

Back to Top

Peter Ludlow, Department of Philosophy, University of Toronto

Linguistic Intuitions Judgments

Linguists appeal to many sources of data in their theorizing, but none are| quite as contentious as so-called linguistic intuitions. My central claim will be that the controversy about linguistic intuitions is simply due to a number of misunderstandings, chief among them being the idea that linguistic intuitions are objects in the Cartesian theater of the mind - quales of acceptability as it were. In contrast, we can take a leaf from Timothy Williamson and maintain that an intuition that P is simply a judgment that P. In particular, if I have the intuition that a particular form is acceptable I am judging that it is acceptable.

Drawing on work by Bogen and Woodward, I'll argue that linguistic data (judgments) provide *evidence for* phenomena (like binding facts or "island effects") which are *explained by* the theory of grammar.

Host:Bill Rapaport

Back to Top

April

Thomas Bittner (Department of Philosophy), University at Buffalo

Ontology and Qualitative Medical Images Analysis

I describe a methodology for the analysis of radiographic images which is based on two major techniques: (1) qualitative geometric abstraction and ontological analysis of anatomical structures. The first technique is a bottom-up approach to extract qualitative spatial relations from medical radiographic images and the second technique is a top-down approach to determine which qualitative relations can possibly hold between the parts of (normal and pathological) anatomical structures. The process of image analysis is both a process of feature extraction, and the extraction of qualitative relations among features. These qualitative relations are then used to classify the images within the space of ontological possibilities.

Host:Bill Rapaport

Back to Top

Bimodal perception of visual and auditory expressions of emotion from early childhood to early adulthood

Emotion is multi-faceted and conveyed through verbal and nonverbal language, particularly the face and voice. Research has focused on single modality processing so little is known about how bimodal cues of emotion are integrated. Additionally, children and adults have been shown to present with different processing preferences during speech perception but these modality preferences have not been examined in emotion processing. The specific aims of this study are to 1) examine modality preferences in processing of congruent and incongruent visual and vocal cues of emotion; 2) compare responses of children and adults to examine perceptual preferences.

Host: Bill Rapaport

Back to Top

EMOTIONAL CONSCIOUSNESS

This talk proposes a theory of how conscious emotional experience is produced by the brain as the result of many interacting brain areas coordinated in working memory. These brain areas integrate perceptions of bodily states of an organism with cognitive appraisals of its current situation. Emotions are neural processes that represent the overall cognitive and somatic state of the organism. Conscious experience arises when neural representations achieve high activation as part of working memory. This theory explains numerous phenomena concerning emotional consciousness, including differentiation, integration, intensity, valence, and change.

Host: Bill Rapaport

Back to Top

Scaling Up Interventions Based on Research
The Building Blocks and TRIAD Projects

The Building Blocks project was a curriculum research-and-development project based on research in cognitive science and mathematics education woven into our Curriculum Research Framework. For example, research-based developmental progressions underlie the learning trajectories that form the backbone of this Framework. Initial randomized controlled trials showed this approach to have strong positive effects. However, we also recognize the "deep, systemic incapacity of U.S. schools, and the practitioners who work in them, to develop, incorporate, and extend new ideas about teaching and learning in anything but a small fraction of schools and classrooms" (Elmore, 1996 , p. 1). Therefore, we abstracted guidelines from research on scaling up successful interventions (see http://UBTRIAD.org). We used these guidelines to create and test our TRIAD intervention. TRIAD stands for Technology-enhanced, Research-based, Instruction,Assessment, and professional Development. Technology benefits students, teachers, and researchers. Research is the basis for all aspects of TRIAD: The instruction, the assessments, and the professional development. The goal of the TRIAD intervention is to avoid the dilution and pollution that usually plagues efforts to achieve broad success. We describe the TRIAD model andtwo randomized controlled trials evaluations.

Host: Bill Rapaport

Back to Top