We work with common sense better than Microsoft

On July 23, 2019, Microsoft published an article The KnowRef Coreference Corpus: a resource for training and evaluating common sense in AI in which they tell how they are trying to work with common sense.

Today we want to study it in more detail and compare it with our approach.

AI has made major strides in the last decade, from beating the world champion of Go, to learning how to program, to telling fantastical short stories. However, a basic human trait continues to elude machines: common sense. Common sense is a big term with plenty of baggage, but it typically includes shared background knowledge (I know certain facts about the world, like “the sky is blue,” and I know that you know them too), elements of logic, and the ability to infer what is plausible. It looms large as one of the hardest and most central problems in AI. Machines can seem glaringly unintelligent when they lack common sense.

Yes, if we take the AI=ML for the truth, then this statement is true. Such systems are good where you need to calculate all the options.

This knowledge and logic are the most important elements by which we are able to understand the information and the environment with which we interact.

That is why we go the other way and grow our AI as a child, so that it also has the kind of knowledge about the world that is now available only to humans.

This is especially true when it comes to language because language is ambiguous. Common sense enables us to fill in the semantic blanks when a statement doesn’t fully specify what it describes. Imagine telling a machine:

The firemen arrived after the police because they were coming from so far away.

Does the machine recognize who was coming from so far away in this scenario? Only if it understands common concepts of distance and time; that is, that being more distant from a thing means taking more time to reach it. Humans acquire this knowledge from experience and learn to utilize and refer to it at will. The question is: How do we endow machines with similar abilities and, just as important, how do we measure progress towards this goal?

Those methods that are offered by our colleagues will not give the desired quality, that is, the human level. Because these solutions do not understand what they are working with. Now consider how we solve the same problem.

1. We find parts of sentence.

2. After that, the AI finds all the entities, defines the roles, as well as the connections between them in the main sentence.

Subject - firemen; action - arrives; object - police.

Also AI understands that firemen arrived after police.

3. We proceed to the second part of the sentence, which is associated with the first causal link.

Subject - they; action - were coming; place - far away.

The AI has an internal memory and before the AI proceeds to analyze the second part, it already has a picture of what is happening, formed on the basis of the first part.

The AI also knows that if the second part follows from the first part of the sentence, then the actions of the second part should be connected with the subject of the first part.

Lets stay in touch. Subscribe to our newsletter.