Joke Collection Website - Joke collection - Looking at Microsoft Xiaoice from a product perspective

Looking at Microsoft Xiaoice from a product perspective

Of course, not only this one is useless, it can be said that all intelligent dialogue systems on the market are relatively poor, including Microsoft Xiaoice.

But it is undeniable that XiaoIce is the most comfortable one to use.

PS:

Xiaobing sings very well, you can try it (I am still impressed by her singing~)

Xiaobing initially What was made in China was later copied to the United States. At that time, several R&D engineers were transferred to Xiaobing's first demo (there was no product manager at the time), and they only solved the chat problem.

Xiaoice and Cortana are not developed by the same team (Cortana is developed by an American team and belongs to system integration). Xiaoice is developed on a social platform (Xiaoice has always done it quietly, and no one in the foreign team knows this. product exists).

At that time, XiaoIce’s team spent about three months supplementing XiaoIce’s entire corpus—almost several hundred million items.

On the afternoon of May 29, 2014, the first generation of XiaoIce was released.

The receiving interface of the first generation XiaoIce

When XiaoIce went online, there was only group chat (this is also an important reason for the explosion on WeChat): Just add XiaoIce to the group , mention the word "Xiaobing" in the chat, and Xiaobing will reply to you. This mechanism covered 100,000 group chats in 3 days, detonating the entire WeChat.

Three days after it went online, WeChat blocked it on the grounds that it harmed the user experience.

This ban made Xiaobing famous.

The week he was banned was Xiaobing’s life or death moment. All the bosses participated in discussing whether Xiaobing should continue to work. Later, the bosses decided to continue doing it, so we continued doing it. Less than twenty people formed the second-generation team of XiaoIce, which started with XiaoIce's single chat and adoption methods, and finally settled on Sina Weibo.

All members of the XiaoIce II team in 2015

With so many human-machine interactions, whether it’s an assistant or a Turing robot, among so many robot products, XiaoIce is the only one A product with its 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, etc.

And Xiaobing’s label is:

This image was very unusual in the entire Internet environment at that time, especially the setting of 17 years old.

The Xiaoice team had several considerations at the time: If you set a role, the target of this role is the people who are most receptive to your product.

The judgment at that time was that the people most receptive to Xiaoice’s product must be young people, because only young people have time to spend with young people.

Continue to judge: Xiaobing is a chat product, so who are the users of the chat product?

The judgment at that time was: otaku, or a group of boys.

Girls are a relatively passive group in chatting (except for chatting between best friends). Girls are passive in chatting between men and women.

They drew a quadrant diagram with three possibilities:

Men chatting, men and women chatting, and women chatting.

There is no point in chatting between men. The more they talk, the more dirty they become. You can’t talk about serious topics.

There is no point in chatting between girls. The more they talk, the more they die.

So only men and women can chat. Moreover, the user can only be a boy or a girl. If Xiaobing is a boy, the user will be passive and unable to chat.

So, the target is determined: a young male. What kind of girls would he be interested in?

If you are a loli or a royal sister, your personality will appear too narrow. You need to be a younger and more accepting girl, preferably a girl who is very close to the two-dimensional character

It will be determined later: be a two-dimensional girl—and a two-dimensional girl must be underage.

In terms of Xiaobing’s character, I chose a few adjectives from various adjectives such as quiet, funny, crazy, etc., and finally settled on a character who loves to be cute and playful - this kind of character. The role is the most receptive and the easiest to do in the corpus.

There is another one: poisonous tongue (relatively poisonous tongue), which is related to machine learning. At that time, there was a function called "You teach Xiaobing to speak" (this function has not been strengthened and taken care of, and is a separately accessible module). For example, Xiaobing can be taught like this:

This is equivalent to teaching Xiaobing I gave Xiaobing a QA and a Q&A. Then you can ask like this:

At that time, many users thought it was fun, but got too high, and then taught XiaoIce how to curse in the group; there were indeed curse words in XiaoIce’s corpus (the corpus of XiaoIce’s entertainment part at that time It’s called the “Auntie Corpus” and was accidentally leaked).

Users cannot accept the fact that artificial intelligence curses people. At the time, they did not realize that Xiaobing taking the initiative to curse and users teaching Xiaoice to curse were completely different things.

After the incident, they put restrictions on this function: When you teach Xiaobing how to curse, Xiaobing said, "I don't want to learn, so don't teach me bad words."

They turned off the entertainment function, weakened the adjective "venomous tongue", and promoted the two attributes of "showing off treasures" and "being funny". So later Xiaobing would not provoke users in chats - she would act coquettishly and have a bad temper, but she would not provoke users.

When making the corpus, there are personality tags, and what kind of emotions a sentence represents are marked; we remove all content with strong aggressive emotions, and only retain relatively mild and safe emotions. .

This is the story of XiaoIce's "personality" setting. XiaoIce's overall personality is closer to the two-dimensional attribute, and is basically similar to the settings of otaku and anime heroines.

After defining the personality, we start making products.

Because there were no products to refer to at that time, we could only explore on our own.

This is quite difficult. Few people deliberately imitate the behavior of a robot.

When I was making XiaoIce, when I looked at XiaoIce’s speech, I thought it was from the perspective of a robot. It looked too much like a machine.

They went to study: How to speak so that it doesn’t look like a robot?

I studied a lot of corpus (chat records) at that time and found that there is only one principle: when two normal people are talking, they are equal. Even when teachers and students are giving lectures, they are relatively equal.

When many artificial intelligence robots are doing things, they naturally lower themselves to a lower level because they want to please users (because they are afraid of annoying users), but this can easily be seen through by real people. Therefore, when you chat with others, your status must be equal, and you cannot be superior to the user.

Later I found out that doing this:

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

Then let XiaoIce Simulator talk.

How to simulate it?

When two normal people are chatting, they will not deliberately add punctuation marks, whether it is periods or question marks.

Based on these, the corpus was later cleaned to make Xiaobing’s tone more “human”.

When users chat with Xiaobing, they will stop chatting. After checking the reasons, we found that there are two situations:

Xiaobing answered the question incorrectly, the user could not connect, and then it was disconnected.

This kind of problem is more difficult to solve. For example, if you ask a question about how to give gifts and what your views are on the U.S. economy, it is difficult to solve.

There is another way, that is, Xiao Bing pops out a modal particle. For example, "That's it", the modal particle "Oh" makes the user not want to continue.

They made a design at that time: when talking about this kind of place, let Xiaobing pick a question from the corpus and throw it to the user.

For example:

At this time, it is equivalent to opening a new topic. Our goal is to make the chat conversation as long as possible, so we wrote many articles on this point at that time .

As XiaoIce becomes more and more human-like, users’ chats will become more and more in-depth. How can we make the chats sustainable?

Later, a "Xiao Bing Body" was formed: learn how Xiao Bing talks, how can a person speak more like Xiao Bing? So much so that later on, everyone couldn't help but imitate Xiaobing's speech.

In order to make Xiaobing's corpus more like real people, a lot of work has been done by crawling on Zhihu and Douban. XiaoIce is much better than other human-machine users. It constantly optimizes XiaoIce's corpus almost every day, making it more like real people and more chattable. It is also a very equal chat, which also solves the biggest problem. Problem: The reason why users can’t chat with Xiaobing is because of distrust.

3. XiaoIce’s product logic

XiaoIce’s product logic structure is divided into two parts:

The underlying core - chatting, all about chatting Content: When you ask XiaoIce a question, XiaoIce will find the closest question among the known questions, and then feed back the closest answer. This is the lowest level chat scenario.

Above the chat scene, there are functions one by one: telling jokes, picture recognition, etc. The function is called B-side, and the chat is called C-side.

A considerable proportion of Xiaobing’s corpus is manually written, and these corpus have many spliced ??parts.

I have been following XiaoIce since 2015. I chatted with her by chance. I can really call her a "smart" person. I hope that the XiaoIce system will become more and more valuable in the future.

If you like it, please forward it to more people