Reid Swanson

Currently, my primary focus is to leverage current psychological theories of interpersonal conflicts to create natural and dynamic character models by blending large-scale data driven approaches with more targeted hand authored domain modeling methods. I am also involved in several other media related projects, which are looking at learning visual preferences from virtual photographs and another that is trying to model stylistic movement of professional performers in virtual characters.

At Disey, I was part of a team that developed technologies designed to bring rich story experiences to guests at their theme parks and resorts.

Say Anything is a collaborative writing system in which a human and computer agent interact to write a narrative story together. A new story is authored in a turned based process that begins with the person writing a single sentence. The computer responds to the input by providing the next sentence of the story. Sentences contributed by the computer are selected from a collection of millions of stories extracted automatically from Internet weblogs. The user may choose to end the story at any point in the process when they are satisfied (or sufficiently discouraged) with their story. After the completion of each story the user may rate their story on several criteria, such as its coherence. By leveraging large amounts of personal narrative content available on the web it is possible to produce compelling stories with users using a relatively simple architecture.

As part of a class with Elaine Chew I developed an interest in computational music processing. Although serving different purposes, music and language share many aspects in common. One of the more particularly interesting facets is that both music and language seem to share a similar mechanism for syntactic processing. High performance automatic syntactic parsing has been a great benefit to computational linguistic research. Despite these advancements there has not been as much research in musical parsing. In this work I used an unsupervised parsing model developed by Dan Klein and Christopher Manning for language processing to musical melodies. I ran a simple classification experiment to show the plausibility of the syntactic structures learned by the model.

For my Masters project I integrated new technology and improved several components of the Story Upgrade system. Story Upgrade is a system designed to help users and organizations find and retrieve stories relevant to their needs. The goal was to provide a tool for content developers of interactive training simulations that would help them write more realistic training simulations in a shorter period of time. By being able to retrieve many stories on a particular topic the author should gain insight into the range of experiences that happen to real people and provide new ideas for situations in the virtual environment.

While at IBM Watson Research Center I designed a visualization system for aiding in the development of large scale semantic networks. This was accomplished using clustering techniques to isolate semantically similar sub-groups of the network to layout regions of the knowledge base separately to ensure related nodes were spatially close together. After a suitable 2D global layout was configured the result was projected onto the surface of the Globe using Google Earth. This allowed the knowledge engineer an easier grasp on the inferences and connections in the KB through the ease of navigation provided by the application, the ability to easily isolate important sections of the graph by occluding less relevant sections beyond the horizon and to obtain a clearer visual depiction of the KB as a whole.

In order to catalog and develop linguistic resources relating to concepts in common sense psychology I helped organize and lead a brainstorming game designed to illicit natural language phrases expressing each concept in the ontology. This data was then used to author finite state automata for recognizing these concepts in English text.