Lenat D.B.,Cycorp Inc. |
Durlach P.J.,Advanced Distributed Learning Initiative
International Journal of Artificial Intelligence in Education | Year: 2014
We often understand something only after we've had to teach or explain it to someone else. Learning-by-teaching (LBT) systems exploit this phenomenon by playing the role of tutee. BELLA, our sixth-grade mathematics LBT systems, departs from other LTB systems in several ways: (1) It was built not from scratch but by very slightly extending the ontology and knowledge base of an existing large AI system, Cyc. (2) The "teachable agent" - Elle - begins not with a tabula rasa but rather with an understanding of the domain content which is close to the human student's. (3) Most importantly, Elle never actually learns anything directly from the human tutor! Instead, there is a super-agent (Cyc) which already knows the domain content extremely well. BELLA builds up a mental model of the human student by observing them interact with Elle. It uses that Socratically to decide what Elle's current mental model should be (what concepts and skills Elle should already know, and what sorts of mistakes it should make) so as to best help the user to overcome their current confusions. All changes to the Elle model are made by BELLA, not by the user - the only learning going on is BELLA learning more about the user - but from the user's point of view it often appears as though Elle were attending to them and learning from them. Our main hypothesis is that this may prove to be a particularly powerful and effective illusion to maintain. © 2014 International Artificial Intelligence in Education Society. Source
Agency: Department of Defense | Branch: Army | Program: SBIR | Phase: Phase I | Award Amount: 99.87K | Year: 2004
Current document search techniques are inadequate to handle queries whose answer requires establishing links among a number of (potentially obscure) facts. We propose to remedy this situation with a prototype system that applies statistical and knowledge-based methods to the problem of identifying chains of connections between terms and entities. We will extend existing predictive annotation methods for indexing entities by tagging them with Cyc concepts. In addition, our system will be capable of indexing entire facts, by applying Cyc's parsers to text sentences. Parsing a sentence into the CycL representation language results in a full semantic translation, so links among entities mentioned in a sentence will be transparent in the CycL representation. We propose also to demonstrate the discovery of "anonymous" links, by using technology we developed during the ARDA AQUAINT project. This module, the Holistic Query Expansion Algorithm (HQEA), is designed to process a textual query and produce a set of pairwise highly-correlated terms that connect all the terms within the original query into a coherent topic. Finally, a user-feedback mechanism will be developed to refine the performance of the Cyc-based semantic annotators.
Agency: Department of Defense | Branch: Air Force | Program: SBIR | Phase: Phase I | Award Amount: 99.18K | Year: 2006
Analytical tasks at the all-source level, and above, generally require access to intelligence distributed among a variety of forms: structured databases with differing schemas, electronic maps with various metadata schemes, and textual reports in multiple languages. Knowledge bases that employ highly expressive formal languages, such as extensions of first order logic, offer a solution to the challenge of combining information from the current daunting variety of data forms. Such knowledge bases can, in principle, represent the content of all structured sources within a single structure. Such a structure can in turn be accessed by interfaces that allow information to be formed in a way that is natural to analysts - rather than in the various idiosyncratic forms of multiple structured sources. Moreover, the expressive power of such knowledge bases makes it possible for them to integrate existing structured sources as a virtual part of their content, by translating data in those sources. A complete Virtual Knowledge Base (VKB) of data for intelligence analysis would address the need for data to exist in a form that is intelligible to analysts, while circumventing the impracticality of constructing a single knowledge base in which all intelligence data actually resides.
Agency: Department of Defense | Branch: Air Force | Program: SBIR | Phase: Phase I | Award Amount: 96.67K | Year: 2006
In any complex environment – such as managing a battlespace, launching a satellite, or operating a global enterprise – critical decisions depend on a broad range of information sources, decision-making guidelines, and an array of operational and environmental factors. These challenges highlight the need for decision support systems whose decisions are based on both structured and unstructured information sources, and that can explain their decisions in a manner that garners trust from those relying on its conclusions. The Cyc knowledge-based environment supports many of the capabilities needed for such a system. Its Semantic Knowledge Source Integration functionality permits smooth integration with structured information sources, while its inference engine and NL generation capabilities provide textual justifications for its actions. Unstructured data (such as text documents, imagery, videos, etc.) can be mapped to the Cyc ontology to model their content as well as to identify key metadata (such as the source, creation date, scope, etc.), enabling material from unstructured sources to be seamlessly included in the decision process. We propose to design a decision support architecture around these existing capabilities that would gracefully incorporate a wide variety of information sources and offer greater transparency into its decision making process.
News Article | August 9, 2014
Editor’s note: Catherine Havasi is CEO and co-founder of Luminoso, an artificial intelligence-based text analytics company in Cambridge. Luminoso was founded on nearly a decade of research at the MIT Media Lab on how NLP and machine learning could be applied to text analytics. Catherine also directs the Open Mind Common Sense Project, one of the largest common sense knowledge bases in the world, which she co-founded alongside Marvin Minsky and Push Singh in 1999. Imagine for a moment that you run into a friend on the street after you return from a vacation in Mexico. “How was your vacation?” your friend asks. “It was wonderful. We’re so happy with the trip,” you reply. “It wasn’t too humid, though the water was a bit cold.” No surprises there, right? You and your friend both know that you’re referring to the weather in terms of “humidity” and the ocean in terms of “cold.” Now imagine you try to have that same conversation with a computer. Your response would be met with something akin to: “Does. Not. Compute.” Part of the problem is that when we humans communicate, we rely on a vast background of unspoken assumptions. Everyone knows that “water is wet,” and “people want to be happy,” and we assume everyone we meet shares this knowledge. It forms the basis of how we interact and allows us to communicate quickly, efficiently, and with deep meaning. As advanced as technology is today, its main shortcoming as it becomes a large part of daily life in society is that it does not share these assumptions. We find ourselves talking more and more to our devices — to our mobile phones and even our televisions. But when we talk to Siri, we often find that the rules that underlie her can’t comprehend exactly what we want if we stray far from simple commands. For this vision to be fulfilled, we’ll need computers to understand us as we talk to each other in a natural environment. For that, we’ll need to continue to develop the field of common-sense reasoning — without it, we’re never going to be able to have an intelligent conversation with Siri, Google Glass or our Xbox. Common-sense reasoning is a field of artificial intelligence that aims to help computers understand and interact with people in a more naturally by finding ways to collect these assumptions and teach them to computers. Common Sense Reasoning has been most successful in the field of natural language processing (NLP), though notable work has been done in other areas. This area of machine learning, with its strange name, is starting to quietly infiltrate different applications ranging from text understanding to processing and comprehending what’s in a photo. Without common sense, it will be difficult to build adaptable and unsupervised NLP systems in an increasingly digital and mobile world. When we talk to each other and talk online, we try to be as interesting as possible and take advantage of new ways to express things. It’s important to create computers that can keep pace with us. There’s more to it than one would think. If I asked you if a giraffe would fit in your office, you could answer the question quite easily despite the fact that in all probability you had never pictured a giraffe inhabiting your office, quietly munching on your ficus while your favorite Pandora station plays in the background. This is a perfect example of you not just knowing about the world, but knowing how to apply your world knowledge to things you haven’t thought about before. The power of common sense systems is that they are highly adaptive, adjusting to topics as varied as restaurant reviews, hiking boot surveys, and clinical trials, and doing so with speed and accuracy. This is because we understand new words from the context they are used in. We use common sense to make guesses at word meanings and then refine those guesses and we’ve built a system that works similarly. Additionally, when we understand complex or abstract concepts, it’s possible we do so by making an analogy to a simple concept, a theory described by George Lakoff in his book, “Metaphors We Live By.” The simple concepts are common sense. There are two major schools of thought in common-sense reasoning. One side works with more logic-like or rule-based representations, while the other uses more associative and analogy-based reasoning or “language-based” common sense — the latter of which draws conclusions that are fuzzier but closer to the way that natural language works. Whether you realize it or not, you interact with both of these kinds of systems on a daily basis. You’ve probably heard of IBM’s Watson, which famously won at Jeopardy, but it’s a lesser-known fact that Watson’s predecessor was a project called Cyc that was developed in 1984 by Doug Lenat. The makers of Cyc, called Cycorp, operate a large repository of logic-based common sense facts. It’s still active today and remains one of the largest logic-based common sense projects. In the school of language-based common sense, the Open Mind Common Sense project was started in 1999 by Marvin Minsky, Push Singh, and myself. OMCS and ConceptNet, its more well-known offshoot, include an information store in plain text, as well as a large knowledge graph. The project became an early success in crowdsourcing, and now ConceptNet contains 17 million facts in many languages. The last few years have seen great steps forward in particular types of machine learning: vector-based machine learning and deep learning. They have been instrumental in advancing language-based common sense, thus bringing computers one step closer to processing language the way humans do. NLP is where common-sense reasoning excels, and the technology is starting to find its way into commercial products. Though there is still a long way to go, common-sense reasoning will continue to evolve rapidly in the coming years and the technology is stable enough to be in business use today. It holds significant advantages over existing ontology and rule-based systems, or systems based simply on machine learning. It won’t be long before you have a more common-sense conversation with your computer about your trip to Mexico. And when you tell it that the water was a bit cold, your computer could reply: “I’m sorry to hear the ocean was chilly, it tends to be at this time of year. Though I saw the photos from your trip and it looks like you got to wear that lovely new bathing suit you bought last week.”