Current Research Projects
In 1999, I began working on a project at the MIT Media Lab to collect common sense from volunteers on the internet. Since then, the Open Mind Common Sense project
(OMCS) has expanded. The English site over one million sentences from over 35,000 contributers. There are OMCSes in Korean, Japanese, Portuguese, and Dutch.
With others in the Common Sense Computing Initative, I mantain the semantic network ConceptNet
and work extensively on AnalogySpace
. AnalogySpace makes rough conclusions about new common sense knowledge based on similarities and tendencies by forming the
analogical closure of a semantic network through dimensionality reduction.
I work in an imprecise world: I start with structured knowledge representations and classic algorithms, and then I generalize them to be more compatible with the incomplete and noisy data of the real world. This style of machine learning is used in my work on AnalogySpace, Blending, Streaming AnalogySpace, and Blending for Spectral Association.
Currently, I am working with the OMCS team to develop a dimensionality reduction algorithm to be able to process and predict time series information and respond to feedback. Returning to the goals of Marvin Minsky's Emotion Machine and Society of Mind, we hope to create a metacognitive system made of individual planner agents whose interactions are governed by other planner agents. Applications of this work include metacognition music processing, story understanding, and behavioral and procedural modeling.
I developed an AI backend for the Glass Infrastructure
, or Charms, project at the Media Lab. A collaboration between four research groups, Charms allows visitors to the Media Lab to browse a directory using RFID-enabled touch screens throughout the building. I built a model of the topics that people work on in the lab by blending common sense with information "read" from the Media Lab's repository of project abstracts. Additionally, I created the ConnectMe project for the same event using the same model. ConnectMe creates an ad-hoc, effortless social network and helps participants at a company, conference, or research institution find others who share similar interests, and even suggests conversation topics.
Additionally, I am interested in using common sense to better understand free text information. Luminoso, is a tool developed by the OMCS team that uses common sense and Blending to "read between the lines", in order to better understand opinions and feedback expressed in free text. Luminoso creates a semantic space from the ideas in a set of documents, including common sense background information, and allows interactive exploration of this space.
Games and common sense seem to go hand in hand. Our current knowledge acquisition efforts are through crowdsourced "Games with a Purpose". On the flip side of this, games need common sense, planning, inference, and event modeling to provide unscripted, spontaneous, and reactive behavior. In addition to Games with a Purpose, I am interested in creating AI in games, especially those with interactive narratives that react coherently and spontaneously (that is, without explicit scripting) to events and behavior in the game. By understanding what is the normal behavior for a given situation, a game character or game world can become adaptive more able to react naturally and take into account the players goals and patterns of behavior.
Understanding language in any form requires understanding connections among words, concepts, phrases and thoughts. Many of the problems we face today in artificial intelligence depend
in some way on understanding this network of relationships, which represent the facts that each of
us knows about the world and how words relate to one another. When people communicate with each other, their conversation relies on many basic, unspoken assumptions, and they
often learn the basis behind these assumptions long before they can write at all. Only traces of
these assumptions are found in corpora.
Generative Lexicon and Common Sense Reasoning are related areas which are focused on learning such word meanings. This paradox of this vital unspoken information has lead to researchers in these fields to manually construct resources either using experts or volunteers from the internet. Such projects are major undertakings, using vast amounts of time and resources, but the resulting resources are invaluable to the field.
In my work, I hope to create a method, using principal component analysis, which can aid in the creation and bootstrapping of a large lexical resource by providing the human ontologist or volunteer contributer with a list of possible possible relations for each word in the system.