domeet webmaster

Artificial intelligence has been around for more than 60 years, and some of the largest technology companies in the United States (Amazon, Microsoft, Google, Facebook, etc.) are just beginning to tap into the potential of AI and try to figure out how artificial intelligence will change. our future. This article is the fifth in a series of articles by Fast Company, “The New Rules for AI,” which introduces Facebook’s efforts to develop chat bots. The original author is Mark Sullivan, titled: Facebook is betting the next big interface is conversation

Back in 2015, chat bots were very hot.

The most hyped thing is Facebook’s M. The company’s original intention was to make it a flexible, general-purpose robot that could perform many different things, such as buying goods and arranging gifts. Booking and making travel plans, etc. But the hype is far greater than the essence. When Facebook conducted an M test of 2,500 people in the Bay Area, most of the tasks it required to perform were not up to the task.

After the initial enthusiasm for M and other chat bots (Microsoft CEO Satya Nadella once claimed that “chat bots are new apps”), there was a wave of disappointment. Chat bots are not so talkable. That’s because they are trained to talk only a small part of the story and perform limited specific tasks. They can’t communicate naturally with people, and they can’t make their own responses based on a general understanding of words and their meaning. They can only provide a general reply.

M’s beta test hasn’t finished yet, and Facebook has reduced the original blueprint for a chat bot, even though some of its natural language technologies have been integrated into the less daring Messenger chat bot. This chat bot can perform a single simple task, such as ordering or sending a Q&A message. Companies such as American Express and 1-800-FLOWERS are still using this low-profile chat bot to answer customer service questions, accept basic orders and provide account balance information. If you ask them anything beyond their limited understanding, many robots will transfer you to a human customer service representative.

But, FThe ace research team at acebook has got rid of that single-function chat bot. Antoine Bordes, a natural language researcher at Facebook, told me: “We have been saying for the past three to four years that goal-oriented dialogue should not be the path we need to explore, because this road is too difficult and the risk is too great.” The chat bot booked “the wrong plane, the wrong flight, so it is a huge mistake for money or travel.”

Bordes explained that Facebook is no longer focusing on specific task-based mechanisms, but rather taking a step back to solving a deeper problem – teaching virtual agents to talk like people. If chat bots can better understand and communicate with humans, they may eventually become better assistants who can help you with the actual tasks, such as booking a ticket.

Facebook has been seriously involved in this work, and they have hired some experts in natural language AI. What the company likes to point out is that, unlike some high-tech giants, it publishes AI research results online for use by the entire research community, which will help others who are developing the next generation of AI. However, this research will definitely launch Facebook’s own products.

Including Messenger and WhatsApp, the chat app is a natural destination, the latter has been acquired by Facebook and is still studying how to make money. As CEO Zuckerberg puts a new vision on the company and places more emphasis on private conversations, Messenger and WhatsApp need to add functionality to stay ahead of other messaging platforms such as WeChat, Telegram and Apple’s iMessage.

The development of algorithms that can chat with people has become the main goal of technology giants. Amazon, Gooogle and Microsoft have joined Facebook games. They all put their treasures on the power of dialogue with people – not just based on Chat in the text chat app, but also use voice assistants and other experiences. Due to recent research progress, the road to a true conversational computer suddenly became clear, but the reward for the first line of the line is still worth fighting for.

In other words, Facebook’s research on natural language is much more than just resurrection M or improving Messenger-based chat bots. This is about the future of the entire company.

Enter the neural network

Developing a digital agent that can engage in realistic dialogue with people is arguably the most difficult of all natural language problems. It takes a machine to learn a dictionary of all words, to understand all their usage and nuances, and then use it in real-time conversations with unpredictable people.Something.

Natural Language The AI ​​circle has only begun to take greater steps toward common-sense robots in recent years. Part of the reason is the tremendous advancement in neural networks, which are a type of machine learning algorithm that can identify patterns by analyzing large amounts of data.

In most of AI’s history, humans have been supervising the machine learning process of software. In a so-called supervised learning technique, human teachers slowly train the neural network by providing the correct answer to the problem, and then adjust the algorithm to let the machine achieve the same solution.

Supervised learning works well when you’re painstakingly tagging large amounts of data, such as photos that contain a cat or other item. But this method usually doesn’t work in the world of chat bots. Thousands of hours of conversations between people are difficult to find on a large scale, and the cost of a company to collect is very high.

Because it’s difficult to teach chat bots how to conduct conversations with these older methods, researchers are always looking for alternatives to supervised learning, hoping that neural networks can learn the data themselves without human intervention.

One way to reduce the need for training data is to teach the machine a basic common sense. If the computer has a certain understanding of the world, such as knowing the relative size of the objects, knowing how people use them, and how some physical laws will affect them, then it may be able to narrow the choice to the only possibility. Within the scope.

People will do this naturally. For example, suppose you drive on the road next to a cliff and suddenly see a large stone on the road ahead. You have to avoid hitting the stone. However, when considering the choice, you will never make the decision to hit the steering wheel to the side of the cliff. Because you know that due to gravity, the car will suddenly fall off the cliff, and the car will be destroyed.

Yan LeCun, vice president and chief AI scientist at Facebook, said: “Most of our learning is in… watching the world.” Yann LeCun is a legend in the field of artificial intelligence and has been dealing with AI since the 1980s. The biggest challenge. “We learn a lot from our parents and others, but through interaction with the world, we have learned a lot by trying, failing, and correcting mistakes.”

The AI ​​(called unsupervised learning) trained with this technique is also the way to learn. For example, just as a child learns about the world through five senses, an autonomous car collects data about the world through the deployment of many sensors and cameras. In this way, scientists provide a lot of training data for the machine to digest. They don’t ask the machine to generate the correct answer or trick it into reaching a certain goal. Instead, they only ask it to process and learn data, look for patterns and map relationships between different data points.

In many cases, it is difficult to get the necessary data. But one area of ​​AI is that neural networks can learn a lot about the world without sensors: that is natural language processing. Researchers can use a large amount of existing text to help the algorithm understand the human world, which is a necessary part of understanding the language.

Let’s say we assume the following two sentences for the neural network:

“The trophy won’t fit in the box because it’s too big.”

“The trophy won’t fit in the box because it’s too small.”

To know what “it” refers to in each sentence, the model needs to know a little about the knowledge of objects and the relationships between them. LeCun said: “There is already enough structure in the trained text. Through this structure, you can know that when a thing is put into another thing, if it is put too much, it will not be put in.” p>

This technology proves to be the secret of a new generation of more conversational and practical Facebook chat bots.

Get to know BERT and RoBERTa

The current progress in unsupervised training in the natural language system was initiated by Google in 2018. Their researchers created a deep learning model called BERT (Bidirectional Encoder Representations from Transformer).s, a two-way encoder representation from Transformer), and then provided it with uncommented text from 11038 books and 2.5 billion word entries from English Wikipedia. The researchers randomly covered certain words in the text and then challenged the model to see how they were filled back.

After analyzing the entire training text, the neural network finds patterns of words and sentences that often appear in the same context, helping it understand the basic relationships between words. And because words are representations of objects or concepts in the real world, the model not only learns the language relationships between words, but also learns more: it begins to understand the relationships between objects.

BERT is not the first model to train a computer to understand human language in an unsupervised way. But this is the first model to learn the meaning of words in a context.

Microsoft partner, researcher Gao Jianfeng, research manager of the deep learning department of the institute said: “I want to say that this is one of the two major breakthroughs in natural language processing. You have used this model as a model for developing all other natural language processing models. New benchmarks.” As other researchers build their own models based on Google’s models, the BERT research paper has so far more than 1,000 academic references.

LeCun and his team are also one of them. They built their own models and then made some optimization adjustments, greatly expanding the amount of training data and increasing the allowed training time. After the neural network has run billions of calculations, Facebook’s language model RoBERTa has outperformed Google’s model. The accuracy rate with BERT is 80.5%, which is 88.5%.

BERT and RoBERTa represent a new way to teach computer conversations. LeCun said: “In doing so, the system must express the meaning of the words it sees, the structure of the sentence, and the context. The result is that it has a little understanding of what the language is, which is strange because it is for the world. Physical reality is actually ignorant. It has no vision, no hearing, it has nothing.” It only knows the language – letters, words and sentences.

Getting closer to the real conversation

LeCun says that natural language models trained with BERT or RoBERTa still don’t have much common sense — that common sense is enough to generate chat responses based on a wide range of common sense. This can only be regarded as the beginning of the training algorithm to talk like a human.

Facebook’s natural language researchers are still trying to develop more conversational features based on RoBERTa. They start with a conversation between the researcher and the chat bot to get a good idea of ​​when and how the conversation was interrupted orboring. Their findings spurred the development of a study that proposed ways to train robots to avoid the most common conversational failures.

For example, chat bots often contradict themselves because they don’t remember what they have said in the conversation. The chat bot might say that he loved the re-release of “The Ranger” for a minute, and then said that he didn’t like watching TV in the next minute. Chatbots that build their own original responses (rather than getting examples from training data) tend to answer questions in a vague way to avoid errors. They often seem to be unable to distinguish emotions, which reduces their interactivity.

Chat bots must also be able to use knowledge to become interesting interlocutors. Chat bots that can take advantage of a variety of information are much more likely to talk to people. However, current chat bots are trained with a single domain knowledge corresponding to the tasks that the robot is to accomplish, and it becomes a problem if people begin to comment on topics other than knowledge of the robot domain. For example, if you ask the robot that sent the pizza about any topic other than pizza, the two sides will soon be unable to talk.

What should I do? Facebook researchers have been training natural language models to extract data from many areas of knowledge and then inject this information into the dialogue process in a natural way. Future research will focus on teaching robots when and how to move conversations from general topics to specific tasks.

One of the biggest challenges in developing chat bots is to allow chat bots to continue learning after they are deployed. The meaning of words changes over time, and new nouns and proverbs become culturally important. At the same time, chat bots can’t be too easily affected – Microsoft’s chat bot Tay is too fast to learn too much in conversation with people on the Internet, and it becomes a racist horrible within 24 hours. . Facebook is teaching its experimental chat bots, learning when the conversation is going well, and analyzing the language of human chat partners, and discovering whether the robot is stupid or boring.

Predicting when Facebook’s progress in the lab may bring even a chat bot that looks like a human being is dangerous. However, the time we can judge the results ourselves may not be that long. Facebook researcher Jason Weston told me: “We believe we are very close to having such a robot, we will be able to chat with it and will see its value.”