Commonsense reasoning

Commonsense reasoning

In artificial intelligence, commonsense reasoning is a special ability that helps machines think more like humans. It lets a computer make smart guesses about everyday situations.

This kind of reasoning is important because it helps AI devices act more naturally. It uses two key ideas from humans: folk psychology, which is how we think about what other people are thinking, and naive physics, which is how we understand how things work in the world.

A device with commonsense reasoning might guess that a ball will roll away if you throw it, or that someone is hungry if they look at food. This helps machines work better with people and solve problems in smart ways.

Definitions and characterizations

Commonsense knowledge is the basic understanding we all have about everyday things and events. It includes ideas about how objects behave, how people think and feel, and how the world around us works. For example, we know that water is wet, that people usually play cards for fun, and that animals have different habits.

Experts say that commonsense knowledge is what most people already understand without needing to learn it from books. It helps us make sense of new situations by using what we already know. A typical seven-year-old already has a lot of this commonsense knowledge about the world, like how things move, what plants and animals do, and how people interact with each other.

Commonsense reasoning problem

People are naturally good at guessing what will happen in everyday situations. For example, we know that if we roll a pen off a table, it will fall to the floor. We also understand why people act the way they do.

Current AI systems, like self-driving cars, do not think the same way humans do. They struggle to understand simple things about the world and people's intentions. This can lead to mistakes that seem strange to us. This difference means AI often sees things differently than people do.

Commonsense knowledge problem

Main article: Commonsense knowledge (artificial intelligence)

The commonsense knowledge problem is a big challenge for making smart machines. It means creating a database that holds everyday knowledge that most people know. This helps computers understand and do tasks like recognizing objects, translating languages, and finding information in text. Making a machine think and act like a human is very hard because it needs to know simple facts that people naturally understand.

Commonsense in intelligent tasks

In 1961, Bar Hillel talked about why practical knowledge is important for understanding language when using machines. Sometimes, a machine can get confused about words, but it can often figure things out by knowing a little about the world.

For example, when translating sentences like "The electrician is working" and "The telephone is working" into another language, a machine might know that "electrician" usually means someone is fixing things, while "telephone" means a device that needs to be working properly.

This kind of thinking is also important in computer vision. When a machine looks at a picture, it can guess what small things are by looking at what else is in the picture. For example, in a bathroom picture, things like towels or bottles might be hard to recognize unless there is a toilet or sink nearby to give clues.

In movies, understanding what is happening can be harder. To really understand a scene, a machine would need to guess what characters are thinking or planning to do next. Right now, machines can only recognize basic actions and follow characters, but they can't think through complex situations like humans can.

For robots that need to work in real places, like a robot waiter at a party, having this kind of thinking is very important. If the robot sees that a glass it picked up is broken, it should not try to pour a drink into it. Instead, it should pick up a new glass. These tasks seem easy for us, but they are tricky for robots to learn.

Successes in automated commonsense reasoning

Computers have made progress in understanding everyday ideas. They can now guess how things are grouped together, how actions change things, and how to think about time.

For example, computers can understand that a robin is a type of bird. Because robins can fly, a specific robin named Tweety can also fly. They can also understand that when events happen one after another, the world changes in certain ways. This helps computers predict what will happen next.

Computers can also make guesses about time. They know that if someone was born after another person, they might die sooner. This helps computers understand stories and events better.

Challenges in automating commonsense reasoning

Since 2014, some companies have tried to use commonsense reasoning in their systems. But these systems often use statistics instead of real understanding. They work with words but don’t truly understand what they mean.

According to Ernest Davis and Gary Marcus, there are five big challenges in creating a true commonsense reasoner:

• First, some areas like communication, relationships, and physical processes are not fully understood, even by humans.
• Second, some situations seem simple but can be very complex in ways we don’t fully know.
• Third, commonsense reasoning needs to make reasonable guesses based on what we know, which can be tricky because the information isn’t always certain.
• Fourth, there are many everyday situations that happen often, but also many rare ones that are hard to predict.
• Fifth, it’s hard to decide how detailed our guesses should be.

As of 2018, computer programs still struggle with commonsense tasks, doing much worse than humans. Some experts think solving this problem might require creating a true human-level intelligence. Some researchers believe that traditional learning methods aren’t enough and are exploring new techniques.

Approaches and techniques

We study how people reason with common sense in two main ways: using knowledge and learning from data.

Knowledge-based methods involve experts deciding what we need to know about certain topics. These methods can be very detailed and theoretical, or simple and based on everyday ideas.

Another way is to learn from large amounts of data, sometimes with help from many people. For example, COMET (2019)) uses smart computer programs and existing knowledge to make guesses that are close to how humans think. However, it still doesn’t fully understand everything the way people do. Some new methods also use pictures as well as words to learn more about how things work in the real world.