Home > News content

Facebook tried to develop the most perfect chat robot, but AI still has limitations

via:博客园     time:2017/4/21 9:00:47     readed:1235

FacebookFacebook 试图开发最完美的聊天机器人,但 AI 仍有局限

[AI generation editor press] foreign media this week, published the Facebook virtual assistant M over the past two years of development. Facebook is trying to develop the most perfect chat robot, but M's progress also reflects the limitations of current artificial intelligence.

The following is the AI ​​generation (micro signal: tencentai) compiled the original article:

Amazon Alexa can be called Uber's vehicle, Siri can control your home smart thermostat. At present, these services attract millions of users every day. However, there are about 10,000 people know, Facebook M is the smartest smart assistant. Most of these people come from California.

Recommend and book a romantic hotel in Morocco, but also suitable for children? no problem. Looking for local service providers to trim the lawn price? Can do it. Facebook This experimental intelligence assistant is integrated in the Messenger application. M does not simply extract information from the database, but rather understand the complex instructions and take action, such as booking movie tickets, or contact a company to obtain information.

M looks so smart, but in fact it has resorted to cheating. M is similar to Siri. When you send a message to M, the algorithm will try to determine what your needs are. If the algorithm fails to succeed, then M will not simply return to the web search results, or say "I'm sorry, I do not understand this problem & rdquo ;. On the contrary, the artificial customer service will be in your case without hesitation, the demand for a reply, like the computer is still continuing to serve you the same. (Facebook declined to disclose the total number of these manual customer service.)

This design led to Facebook is difficult to promote M to all 1.2 billion Messenger users. Thus, in 2015, Facebook provided only M to thousands of users and positioned M as a semi-open R & D project. Combining manual services with algorithms will help Facebook to see how the user will respond to omnipotent virtual assistants, and the algorithm can continue to learn from the human trainer & rdquo;

Alex LeBrun is the initiator of this project. He said: "Everyone in this field dreams of developing a very, very intelligent virtual assistant. "M opened a way to achieve this goal.

At present, after two years, Facebook's research project can be considered very successful. Users like M, and software is also proven to be able to learn from human trainers. However, Facebook still can not face the remaining 99.9% of Messenger users to provide M, this progress is much slower than expected.

"We knew it was a daunting challenge, but it was even harder than I thought," said LeBron. Whether it is learning speed, or the development of automation, these are slower than we expect. "The story of M once again shows how much progress has been made in recent years, and how long has the future been going on?

M: "lend" on the moon & rdquo; abbreviation

Humans seem to like to talk with the machine very much. The world's first chat robot Eliza was born in 1964, the developer is the Massachusetts Institute of Technology professor Joseph & middot; Weizhenbaomu (Joseph Weizenbaum). The chat robot can scan a line of text, and a special keyword to respond, very suitable for the work of the clinician. But it is troubled by Wei Cenbaum, including his secretary, many of the trial users will be impressed by them, even if they know Eliza actually ignorant. He then said: "I did not realize that for ordinary people, a short time exposure to a simple computer program can also cause significant delusions. & Rdquo;

It is more difficult to let the chat robots talk to you and make them work for you. When the virtual assistant is asked to complete a job, the ambiguity or chicken with the duck's reply does not make sense. The current software is not good at understanding human language and the world, so similar to Siri and Alex's virtual assistant can only be programmed to master the specific skills.

FacebookFacebook 试图开发最完美的聊天机器人,但 AI 仍有局限

As a result, the virtual assistant on the market can do a limited job. At the end of last year, it was argued that chat robots would change the way humans interact with computers as mobile applications, and that would attract interest from Microsoft, Facebook, and some technology industry investors. But in fact, for the above reasons, to achieve this goal is also very difficult. Mobile service Burner CEO Greg & middot; Cohen (Greg Cohen) said: "The current chat robot is desperate. For industry watchers, chat robots are overly hype, but the development is poor. "Burner helps Airbnb platform homeowners create simple chat robots and answer customer FAQs.

Prior to the development of M, LeBron had spent more than 10 years developing traditional narrow chat robots. However, he wants chatting robots to do more. He joined Facebook at the beginning of 2015, when Facebook acquired Wit.ai, co-founded with others. Wit.ai helps companies develop chat robots, providing customer support services. Prior to this, LeBron had his own chat robot company sold to the voice recognition giant Nuance.

Lebron said: "Every chat robot on the market, including the development of this, are based on the rules." You will know that someday you will touch the ceiling and can no longer move on. Our children do not use rules or scripts, and they will be smarter than you in the future. & Rdquo;

M was originally provided only to Facebook employees, then opened to California part of the messenger of Messenger users. Soon, the algorithm proved to be able to learn some of the manual operation.

Facebook's artificial intelligence research department uses M to test new learning software called the memory network. This new software is good at answering questions about simple stories. The software can use the storage module to extract critical information for subsequent use. And Google is also testing such a design, optimize the ability to deduce the software.

In 1964, Wei Cenbaum had suggested that such a design would make Eliza smarter. In a few weeks time, M has been a significant change. LeBron remembered that he was surprised when he thanked M for his own ticket. At that time M automatically reply: "rude, enjoy this movie it "M has learned the memory of the task has been assisted and the environment at that time. LeBron said: "We were shocked, no one to write procedures to achieve this function. & Rdquo;

The memory network then helped M get more progress. For example, if someone asks for M to set up and distribute, M will automatically use the key information in the request, including the amount and address, according to the online shop information to generate recommendations. The artificial trainer then decides which option to recommend to the user.

In addition, M efficient performance also attracts users greater interest. For fully automated, limited virtual assistants such as Siri and Alexa, users will only use the most reliable features. And for M, this is not the case.

LeBron said: "At the beginning, people try to ask tomorrow's weather, then they will ask:" lsquo; whether there is a decent Italian restaurant? And then they will consult about immigration and wedding preparations. We know that doing so can be dangerous, which is broader than we expect. & Rdquo;

For difficult questions, the manual trainer will do his best to serve the user. But sometimes, they will directly reject the user's problem. M will automatically reject most of the complex user problems, and the diversification of user problems also lead to the algorithm is difficult to fully assume the work of the artificial trainer. In the near future, deep learning techniques are driving machine learning to become more powerful (memory networks are such applications). However, learning to deal with a variety of complex scenes in the case of very small amounts of data is not in the field of deep learning. LeBron said: "This is very smart, can learn to perform complex tasks, but also need a lot of data. & Rdquo;

Long journey

As the progress is slower than expected, Facebook is reconsidering the project. Last week, a feature called M Suggestion appeared in Messenger. This is similar to some simple chat robots, and that's exactly what M is trying to replace. M Suggesttion will focus on your chat with your friends, find clues, and decide if you want to do something, such as calling from Uber or transferring to others. M Suggestion provides a button that can be done with just one click.

Another co-founder of Wit.ai, Laurent Landowski, who currently works on Facebook and is responsible for the M project, said: "I decided to find an application scenario that accelerates the export of value to the user The "LeBron returned home in January of this year, joining Facebook at the Artificial Intelligence Research Lab in Paris.

Initially rely on artificial M still exists, and continue to those lucky users to provide a wealth of services. Facebook said it will be committed to this project, and the current development of artificial intelligence is very good, worthy of long-term investment. Justine Cassell said that over the past few years, the depth of learning to subvert the previous language processing technology, as well as the expectations of the relevant software. She said: "We are in the golden age of these new machine learning algorithms. In fact, the accuracy of Google translation has recently been close to human level.

But this does not mean that, by observing human activities only, the software can learn to perform complex tasks. Kassel said: "I do not think we have a full understanding. "However, Facebook researchers said that there are many concepts to be explored.

One of them is to let M's automation part learn to judge whether the message returned by the user is positive or negative. This mechanism is similar to a human training pet. If not just imitate the practice of human trainers, then the development of M may be faster. To promote a wider range of research, the Facebook team has released tools to help others test and compare helpless robots without scheduled scripts. With M Suggestion, new technology can now be tested on a larger scale.

LeBron and Landoski argue that they are still on track when they are promoting M to the public. Landoski said: "Sometimes we say that it takes 3 years or 5 years, but maybe it will be 10 years or more." & Rdquo;

"We're very hard, we're getting progress, but I think we have everything we need," says LeBron. "He may be right." But it is conceivable that in 1964, when Eliza was seen, many people had similar expectations. (Compiled / Yu Ye)

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments