Looking at Microsoft Xiao Bing with Intelligent Link System from the Product Angle.

Of course, it's not just chicken ribs. It can be said that all intelligent dialogue systems on the market are relatively poor, including Microsoft Xiao Bing.

But it is undeniable that Xiao Bing is the most comfortable one in the experience.

PS:

Xiao Bing sings very well, you can try it (I'm still impressed by her singing ~)

Xiao Bing was first made in China and then copied to the United States. At that time, several R&D engineers were transferred to be the first Demo in Xiao Bing (there was no product manager at that time), which only solved the problem of chatting.

Xiao Bing and Cortana were not developed by a team (Cortana was developed by an American team, which belongs to system integration), while Xiao Bing was on a social platform (Xiao Bing has been doing it quietly, and no one in foreign teams knew about the existence of this product).

At that time, Xiao Bing's team spent about three months supplementing Xiao Bing's entire corpus-almost hundreds of millions of articles.

On the afternoon of May 29th, 20 14, the first generation Xiao Bing released it.

The Collection Interface of the First Generation Xiao Bing

When Xiao Bing goes online, there are only group chats (which is also an important reason for the outbreak of WeChat): As long as Xiao Bing is added to the group and the word "Xiao Bing" is mentioned in the chat, Xiao Bing will reply to you. The mechanism covered 65,438+10,000 group chats in three days, detonating the entire WeChat.

Three days after its launch, WeChat was blocked on the grounds of damaging the user experience.

This ban made Xiao Bing famous.

The week of the ban was a life-and-death moment for Xiao Bing, and all the bosses participated in discussing whether Xiao Bing should continue to do it. Later, the bosses decided to keep doing it, so we kept doing it. Less than 20 people formed the second generation team of Xiao Bing, started to do single chat and adoption in Xiao Bing, and finally settled in Sina Weibo.

20 15 all members of Xiao bing's second generation team

With so many people interacting, the assistant is here, and the Turing robot is also good. Among so many robot products, Xiao Bing is the only one with his own personality. When she chats with you, you can clearly feel that she is more like a person.

Then as a "person", she will have some labels: age, gender, personality and so on.

The label of Xiao Bing is:

This image was very different in the whole internet environment at that time, especially under the setting of 17 years old.

Xiao Bing's team had several considerations at that time: If you set a role, the target of this role is some people who have the highest acceptance of your products.

At that time, the judgment was: Young people's acceptance of Xiao Bing was definitely the highest, because young people had time to spend with young people.

Continue to judge: Xiao Bing is a chat product, so who is the user of the chat product?

The judgment at that time was: otaku, or boy group.

Girls are a passive group in chatting (except chatting between girlfriends), while girls are passive in chatting between men and women.

They drew a quadrant diagram with three possibilities:

Talk to men, men and women, and women.

There is no point in chatting between men and women. The more you talk, the worse it gets. Can't talk about particularly serious topics.

There is no point in chatting with girls. The more you talk, the more you die.

So only men and women can chat. Besides, only the users are boys and Xiao Bing is a girl. If Xiao Bing is a boy, the user is passive and can't speak.

Then, determine the target: young men. What kind of girl will he be interested in?

If you are a Lori and a royal elder sister, then your personality is too narrow. What you need is a young, more acceptable girl, preferably a girl who is very close to Er Yuan.

It was decided later: to be a second-time girl-and the second-time girl must be underage.

In Xiao Bing's personality, he chose several adjectives such as quiet, funny and crazy, and finally decided on a role like selling cute and playing with treasures-this role is the most accepted and easy to do in the corpus.

There is also a poisonous tongue (relatively poisonous tongue), which is related to machine learning. At that time, there was a function called "You teach Xiao Bing to speak" (this function was not strengthened and taken care of, and it was a separate access module). For example, Xiao Bing can teach like this:

This is equivalent to teaching Xiao Bing a question and answer. Then you can ask:

At that time, many users thought it was fun, so they played high, and then taught Xiao Bing to swear in the group; Xiao Bing's anthology does have swearing words (at that time, the anthology of Xiao Bing's entertainment part was called "Aunt's Anthology", which was accidentally leaked).

It is unacceptable for users to swear at artificial intelligence. At that time, they didn't realize that Xiao Bing's swearing on his own initiative was completely different from the user's teaching Xiao Bing to swear.

After the incident, they restricted this function: When you taught Xiao Bing to swear, Xiao Bing said I wouldn't learn, so don't teach me bad.

They closed the entertainment function, weakened the adjective poison tongue, and improved the two attributes of playing tricks and teasing. Therefore, Xiao Bing will not provoke users in the chat later-she will be spoiled and lose her temper, but she will not provoke users.

When doing a corpus, there are personality markers, and what kind of emotions a sentence represents is marked; We remove all the content with strong offensive emotions, and only keep relatively mild and safe emotions.

This is a story set by Xiao Bing's "character". Xiao Bing's overall personality is closer to the secondary attribute, which is basically similar to the setting of otaku and anime hostess.

After defining the personality, we began to make products.

Because there was no product to refer to at that time, I had to explore it myself.

This is very difficult. Few people deliberately imitate the behavior of robots.

When making Xiao Bing, at first glance, Xiao Bing speaks from the perspective of a robot, which is too much like a machine.

They went to study: how to speak so as not to be like a robot?

At that time, I studied many corpora (chat records) and found that there is only one principle: two normal people are equal when they speak. Even when teachers and students give lectures, they all come from relatively equal.

When many artificial intelligence robots are doing it, because they want to please users (for fear of annoying users), they will naturally lower themselves by one file, but they are easily seen by real people. Therefore, when you chat with people, you must be equal and not above the users.

Later, it was found that it was done like this:

In this way, the tone of the chat began to calm down.

Then I began to let Xiao Bing imitate human language.

How to simulate?

When two normal people chat, they don't deliberately add punctuation marks, whether it's a period or a question mark.

Based on these, the corpus was cleaned up later, which made Xiao Bing's tone more like "human".

When users chat with Xiao Bing, they will chat and break up. Check the reasons and find that there are two situations:

Xiao Bing irrelevant answer, users can't connect, and then it is broken.

This kind of problem is difficult to solve. For example, if you ask a question about how to give gifts, what do you think of the American economy? It is difficult to solve.

Another is that Xiao Bing pops up a modal particle. For example, the modal particle "That's it" and "Oh" make users not want to bear it.

They made a design at that time: when talking about this place, let Xiao Bing choose a question from the corpus and throw it to the user.

For example:

This time is equivalent to opening a new topic, and our goal is to talk as long as possible, so we did a lot of articles on this point at that time.

When Xiao Bing becomes more and more human-like, the user's chat will be deeper and deeper. How to continue the chat?

Later, a kind of "small ice body" was formed: how to speak like Xiao Bing, and how to speak like Xiao Bing? That later, all people can't help imitating Xiao Bing.

In order to make Xiao Bing's collected works more realistic, he climbed Zhihu and Douban and did a lot of work. Xiao Bing is far higher than other human machines, and constantly optimizes Xiao Bing's corpus almost every day, making it more like a real person and more chatable. But also very equal chat, which also solved the biggest problem: the reason why users can't chat with Xiao Bing is distrust.

3. Xiao Bing's product logic

Xiao Bing in the logical structure of the product is divided into two parts:

The bottom core-chat is the whole content of chat: when you ask Xiao Bing a question, Xiao Bing will find the closest question among the known questions, and then feed back the closest answer. This is the bottom chat scene.

Above the chat scene, there are various functions: telling jokes, identifying pictures and so on. The function is called B terminal, and the chat is called C terminal.

A considerable proportion of Xiao Bing's corpora are handwritten, and these corpora have many spliced parts.

I started to pay attention to Xiao Bing from 20 15 and chatted with her occasionally. I can really be called a "smart" person, and I hope Xiao Bing system will be more and more valuable in the future.

If it resonates, forward it to more people.

You might like it.

20 19 geek innovation conference | on the growth methodology of individuals and enterprises

Where should I spend my money in my twenties?

20 18 "friends of time" new year's speech-Luo Zhenyu

20 18 year-end summary |? My "tossing" and "not smooth"