Commonsense knowledge (artificial intelligence)

• 1. Introduction

  ›

  Commonsense knowledge in artificial intelligence refers to the everyday facts that every person is assumed to know: things like "lemons are sour" or "cows say moo." These seem trivially simple. And yet, for decades, teaching a machine to know them has remained one of the hardest unsolved problems in AI research. How do you give a program the understanding that baking a cake implies someone will eat it? How do you help a system grasp that a comment about wigs and lipstick directed at a boy is probably an insult? The first AI researcher to seriously tackle this was John McCarthy, who built a program called Advice Taker in 1959. What he started set off a chain of efforts, crowdsourced databases, automated mining tools, and a nuclear treaty monitoring lab that are still running today.

• 2. Advice Taker to Today

  ›

  John McCarthy's Advice Taker, built in 1959, was the first AI program to directly address the problem of common sense. McCarthy recognized early that reasoning about the world required far more than logic and calculation. It required the background knowledge that human beings absorb without thinking. The challenge he identified has never fully been solved. All existing computer programs that attempt human-level AI perform extremely poorly on modern commonsense reasoning benchmarks, including the Winograd Schema Challenge. The problem is considered by many researchers to be "AI complete" - meaning that solving it would require building a fully human-level intelligence in the first place. Some researchers go further and argue that compassionate intelligence, not just cognitive ability, would also be necessary.

• 3. Default Reasoning

  ›

  One practical tool AI researchers developed to work around gaps in knowledge is the use of default assumptions. If a system knows that "Tweety is a bird," and also holds the general belief that "typically birds fly," it can reasonably assume Tweety can fly, even without additional evidence. These assumptions are expressed in AI systems as phrases like "normally P holds" or "typically P, so assume P." What makes this approach useful is that it is revisable. A process called truth maintenance allows a system to update its assumptions as new information arrives. If the system later learns that Tweety is a penguin, and separately knows that penguins do not fly, truth maintenance revises the earlier assumption automatically. Truth maintenance algorithms also create elaborate records of those presumptions, which gives them a secondary benefit: they can provide explanations for how a conclusion was reached, which matters deeply in the field of explainable AI.

• 4. Building the Knowledge Bases

  ›

  CYC and WordNet were among the earliest large-scale efforts to compile commonsense knowledge, driven by human experts painstakingly entering assertions by hand. A significant shift came with the OpenMind Commonsense project, which crowdsourced the collection of everyday knowledge from ordinary people across the internet. That project produced ConceptNet, a knowledge base that also functions as a natural language processing engine. As of 2012, ConceptNet contained 21 language-independent relations covering connections like IsA, UsedFor, HasA, CapableOf, Desires, CreatedBy, PartOf, Causes, and many more. A relation called DefinedAs, for example, can express that a cupcake is defined as a cake that is small, baked within a wrapper, and contains only one area of frosting or icing. Later automated approaches - including text mining tools called WebChild, Quasimodo, TransOMCS, and Ascent - produced databases significantly larger than ConceptNet, though researchers acknowledge these automated sources are of moderately lower quality. A project called AutoTOMIC took a different route entirely, harvesting commonsense assertions directly from pre-trained language models rather than from text. One notable outlier in this field is GenericsKB, which deliberately avoids further normalization and retains sentences in their full natural form.

• 5. From Bullying to Nuclear Treaties

  ›

  Around 2013, researchers at MIT built a system called BullySpace, an extension of ConceptNet designed to detect taunting on social media. BullySpace included over 200 semantic assertions built around stereotypes, allowing the system to infer that a comment like "Put on a wig and lipstick and be who you really are" is more likely to be an insult when directed at a boy than at a girl. ConceptNet found other uses as well, powering chatbots and computers that compose original fiction. At Lawrence Livermore National Laboratory, common sense knowledge was applied in a different direction entirely: an intelligent software agent used it to detect violations of a comprehensive nuclear test ban treaty. These applications span an enormous range - from playground cruelty to the monitoring of international arms agreements - and they all trace back to the same underlying challenge that McCarthy identified in 1959. The question now pressing AI researchers is how to represent commonsense knowledge in forms richer than the triple data model that most databases currently use, since triples are not always well-suited to the complexity of natural language assertions.