Do you know how the machine "chats" with you?

In science fiction, we can often see robots as intelligent as humans, but how do machines store knowledge behind them? How to infer knowledge and apply knowledge? Finally, how to achieve the interaction with people? This time in the Baidu Big Congress, Baidu Technology vice President Wang summed up the three key technologies to achieve artificial intelligence, basic answers to these questions, although the talk is relatively simple, but because the information is too large without a certain background of people is difficult to understand, so I simply to give you a popular science.

Key one, let machine understand language technology

Understanding language is a process from gradual progression of words to events.

1, Word segmentation technology

Let machine understand language, that is to let machine learn to think for themselves, then need to let machine understand language. and Chinese

It is more difficult to understand than English, and every word in English can be read by machine directly, but Chinese is more difficult in Word segmentation. Let's give a couple of columns.

"You | old Zhang |"

"You | | What's the mouth |?"

Above is a simple case, modify a word, meaning on the difference of 108,000, and the machine according to the different words can be forced to separate out. But the following is a toss.

"Table tennis Auction's over."

How to divide the machine? This is difficult, the machine can be divided into two kinds of meaning

"Table Tennis | The auction is over"

"Ping-Pong | racket | sold OUT."

So the difficulty comes, how can we tell the machine how to divide the words in this sentence? It's easy for people to recognize it right away, but it's very difficult for a machine to contextual.

2, the analytical technique of the sentence

After solving the problem of word segmentation is to analyze the sentence, please see the following two sentences

"Tse Feng | who | is | son"

"Tse Feng |"

For machines, these two words to obtain the key word segmentation information is the same, are "Tse Feng", "who", "son", these three key word segmentation information. How do we tell the computer that the semantics of the cause of the sequence are not the same?

This requires a very deep analysis of the language, the semantic understanding, so that they are looking for the answer is different. It's not easy to do this.

3, context-sensitive analysis techniques

After analyzing a sentence, the machine naturally has to be deduced to analyze the content of an article.

In "A Brief History of information" There is a very important principle of information, that is, when we need to transmit information, need a lot of redundant information to protect the accuracy of information, the more useless nonsense the more accurate transmission of information, the same for the machine to understand a sentence is also based on the same principle.

But the problem is that people can rely on intuition to grab critical information, and what does a machine do to grab critical information? How to identify real signals and noises? This is also a crucial technology.

Let's look at the following three columns

We see when searching for "Why the Sky is Blue", Baidu did not put the contents of the first half of the article, but automatically picked the middle of the key answer, and when Sogou search and 360 search, and did not digest the best answer, just the first half of the article put up. This shows that Sogou and 360 search and can continue to improve the space.

4, analyzing the technology of the event

Analysis of the problem of the article after the need to climb a higher altitude, so that the machine to systematically analyze an event, that is, plus the dimension of time, will be associated with an event-related key articles all set, you can restore the process of a historical event.

When we search for "Snowden", we will get the following title expanded by timeline.

On the left is the event collation of Baidu, the middle is 360 of the event collation, the far right is the result of Sogou search.

In the event classification technology, Baidu and 360 have been able to identify, and Sogou does not currently do this.

Event collation is the most difficult technology and how to make the machine understand the highest level of language.

Key two, knowledge mining technology

1, the establishment of knowledge map technology

Let's first assume that we store hundreds of millions of entity knowledge in the machine, this is not difficult for the machine, easy, difficult on the relationship of the storage entity, an entity corresponding to multiple attributes, such as a table corresponding to the brand, color, wood and so on attributes, these attributes are Bai, these relationships are intricately integrated , the data to be stored will refer to the increase in the number of levels, which is destined to be a super mass-scale map.

How to set up the atlas? Use the following sentence to illustrate

"Luxury brand Louis Vuitton 1854 established in Paris, France"

So how does a machine store knowledge?

1 Luxury and Louis Vuitton (recognizing that Louis Vuitton is a brand and a luxury, storing that knowledge)

2 Louis Vuitton was founded in 1854 (identifies the time when Louis Vuitton was set up to store the knowledge)

3 Louis Vuitton was founded in Paris, France (identified Louis Vuitton was founded in, and stored that knowledge)

4 Paris, France (to identify the French and Parisian relations, the storage of this knowledge)

5) ...

The above is only a rough idealized scenario, and Dr. Wang did not say anything more detailed. I'm here to add that in fact, this is only the map has been dynamic, there is a continuous increase in the process of deletion, each statement in the knowledge is based on the time line of large data keyword content, based on statistics before the establishment of the knowledge map, like the human brain, these relational knowledge Atlas appeared and disappeared, Finally, the definitive relationship is left behind, but these are still dynamic, and if the French capital is no longer Paris, the entire relational knowledge map database will update all the data.

2, Knowledge inference technology

When the Knowledge Atlas is established, it is the application of the actual level, which applies the knowledge map to the real implementation, when the user searches for a problem, retrieves the relational map in the database, and then presents the accurate answer of the highest relevance to the user.

1) Direct inference.

Let's take a look at how old Andy Lau is.

When we search this question, the search results will directly show the age of Andy Lau, this is the use of knowledge of the reasoning ability. The 53-Year-old is a dynamic result, the machine behind the scenes of a large number of operations know that the acquisition of age is a dynamic algorithm, you need to subtract the person's birthday and the current time before they can draw conclusions.

In the same way, when we search for "the father of the mother of the Tse-feng son," The result is "Xie", which is the reasoning behind the technology.

2) Classification Inference

This is a simple, intuitive inference model, but it works only when the user asks for the answer to a unique result, but it doesn't work when the user searches for a problem that doesn't have a single standard. Then the technique of classification inference is used here.

For example, when users search for "ornamental fish".

This search does not specify the only standard answer, so the machine from the background of the Knowledge system library extracts on the "ornamental fish" related to the classification of content, listed all the relevant results, gave the various types of ornamental fish results, so that users to find their desired results. Here and by the way three search results are compared, Baidu's most comprehensive search results, sogou second, and 360 did not classify.

Through large data, in the background to classify the entity knowledge, this is a linear direct inference above the high-level integration of reasoning.

In addition to being able to provide direct results, classification inference can bring additional relevance to users with helpful results, and when we search for "Guan Yu", the following results are displayed on the right side of the search results.

The upper left corner is the result of Baidu, the upper right corner is 360 of the search results, the lower left corner is Sogou search results.

And these results are not manual input, all through large data mining results, three can be dug out with the "Guan Yu" related information, but Baidu dug deeper, the relationship between Guan Yu and Liu Bei and Diao Chan dug out. This kind of hidden information mining is the core of large data value, the value of large data is not big, but it is to excavate valuable association, then pull other value. As a simple example, by discovering large data, it is found that a loaf of bread will sell better, there is such an implicit link, then the owner only need to prepare more of the cake will bring more benefits.

Key three, human modeling technology

The ultimate purpose of a machine to develop intelligence is to interact with others, so that the machine can understand human behavior, and only when understanding the behavior of the perfect can the machine be able to apply knowledge to the user's interaction, and then the value of the final commercialization of the landing.

1, individual modeling

The so-called individual modeling, that is, according to the operation of a single user to provide personalized services. "Today's headline" is the use of the individual modeling technology, when users view the news, it will be based on the user's behavior trajectory, the user to recommend their content of interest. Similarly, in the Baidu search engine as well, when users search for more keywords, Baidu has the ability to recommend its more interesting content.

For example, when a user searches for "SF" keywords.

If the user often visit Baidu animation related bar, search related cartoon characters, music and so on operation, then sort in the first will be a website about anime, but if the user often search and express related knowledge, then ranked first in the will be Shun Fung website.

This kind of accurate modeling for the individual in the future more sufficient data, everyone will be data retention, all of our behavior will be data visualization, and then come to their own all relevant conclusions. Current data acquisition and modeling this technology has matured, and the rest of the network of things to go into the tide.

2, Group modeling

Light is not enough for personal modeling, machine is the most important to group modeling, and group modeling is also the core of commercial value.

The so-called group modeling is to judge the behavior of the most people in a scene to collect, and then get the group intersection under each scene, and then come to have a certain part of the attributes of the people often make choices.

The attributes of these people include: geographical, comic-book lovers, American TV lovers, fathers, college entrance examination students ...

The behaviors of these people include: watching anime, watching American dramas, searching for parenting knowledge, searching for college entrance examination knowledge ...

The machine through the backstage judgment will they carry on the group crowd's attribute and the behavior classification, then lets the related high-level find can carry on the commercial decision-making support.

The above is still a little abstract, then we say more simple, for example, we use large data mining to predict a region of more than 30 years old in the recent period of time has a strong demand for autumn trousers, then businesses in the promotion of autumn trousers only need to increase in the region's advertising can get higher returns on profits. Of course, this is no longer the business initiative to find the answer, but the machine through the initiative to provide a series of options for it, businesses only need to passively accept can.

Conclusion：

In a sense, let machine understand language technology, knowledge mining technology, human modeling technology after a period of time of development, to crack Turing test is possible.

But this technology, in any case, needs to be driven by "batteries", the data produced by humans, and more specifically, human desires.

In my opinion, the only business and military driving technology is the two, which represent desire and fear, in this peacetime we are fortunate, we are freed from the fear of war, and then the rest is the endless desire, in technology and people more and more inseparable from the entanglement, We are going to usher in a new world of desire and technology super mixing.

(Responsible editor: Mengyishan)