Artificial intelligence (AI) is the fodder of science fiction past, but reality may be creeping up. It is a big subject of research in today's universities and corporations. Some AI systems are designed to handle specific problems and tasks. Others take a more general approach. General intelligence, sometimes known as strong AI, usually involves research in reasoning, planning, learning, communication, perception, and movement. Integrating these areas creates what we know as AI systems.
However, an MIT Media Lab research team led by Catherine Havasi is looking to add another area of significance into the mix.
The researchers have developed a database called ConceptNet for common sense. Many AI systems currently use methods based on keywords and statistics; computer systems could pick up and understand basic facts using these methods but will have a difficult time understanding basic human communication. For example, if someone says, "I want some chips right now," humans will often interpret "chips" as meaning potato chips. But "chips" may easily confuse a computer system. Are we talking about potato chips? Computer chips? Poker chips?
The idea behind ConceptNet is to give systems and technical devices a better understanding of the human language. The researchers wrote in a paper that they formed a crowdsourced database meant to give ConceptNet an optimized approach for making "context-based inferences over real-world texts." It also helps computers grasp "unknown or novel concepts by employing structural analogies to situate them within what is already known" by the computer.
Robert Sloan and Stellan Ohlsson of the University of Illinois at Chicago recently tested the system. They used the Wechsler Preschool and Primary Scale of Intelligence, a test commonly used to measure a child's IQ. The test focused on the verbal categories, including information, vocabulary, and word reasoning. For a question such as, "What would you wear on your feet?," ConceptNet will search its database for the words must commonly associated with "wear" and "feet." Overall, ConceptNet's verbal IQ was equal to that of a 4-year-old child.
The MIT researchers said in a press release that they can improve ConceptNet's performance by using better algorithms. Furthermore, they think the latest version, which includes 17 million statements (the one tested by UIC had 1 million) can achieve an even higher score.
A system like ConceptNet could help engineers create extremely beneficial applications that would otherwise be impossible.