Dialogue robots: What else to consider besides algorithms?

Abstract

The ai industry is not purely algorithmic. Algorithms are important, but what goes beyond algorithms is even more important for a startup. Today’s sharing will let you know that although robotics is the core application of algorithms, there are more things to be considered in order to really do it well under the current technological conditions besides algorithms.

On August 12, 2017, Hong Qiangning, founder and CTO of Iin Interactive, delivered a speech titled “How to land dialogue Robot” at the “netease Learned Practice Day: Big Data and Artificial Intelligence Technology Conference”. IT big said as the exclusive video partner, by the organizers and speakers review authorized release.

Bot: A robot that performs a task automatically and communicates in natural language.

Chatbot: A robot that uses natural language for conversation purposes.

The concept of chatbots predates even artificial intelligence. The first dialogue robot, ELIZA, was born in 1966 as a simulated psychological counselor, marking the beginning of dialogue integration.

In 1970 there was a conversation robot called SHRDLU that confined conversations to a closed domain, understood and responded to what people were saying, and was better at communicating.

In 1988, Jabberwacky was introduced, also based on pattern matching. Jabberwacky was added to be entertaining, with humorous answers and enjoyable conversations.

In 1995, A.L.I.C.E., technically speaking, A.L.I.C.E. Based on pattern matching, a set of artificial intelligence markup language engine is created, which makes artificial intelligence become more standardized.

In 2001, SmarterChild brought new changes. SmarterChild took chatbots to the Web, reaching 30 million users at its peak, and its impact expanded considerably.

In 2006, Watson beat a real person on a question-and-answer variety show because its knowledge base was so well organized that it was already more accurate than humans.

In 2010, Apple released Siri. Siri is a spoken personal assistant that brings new applications to chatbots.

By 2016, Facebook had released the Messager Platform. 2016 was the year of artificial intelligence, and Facebook decided it had a chance, so it launched a platform where people could build robots. There were tens of thousands of bots running on the platform in the first month.

Since 1966, so many scientists have been making robots talk to humans in natural language, because they have a rich ideal that robots can possess human intelligence. One definition of artificial intelligence is that an intelligence that passes the Turing Test is artificial intelligence. The Turing test itself is a test of a talking robot.

Around April 2016, there was a lot of enthusiasm for Facebook’s Messager Platform, where a lot of people were developing robots.

But by February 2017, Facebook found that the bots were so ineffective that they had an error rate of 70 percent. So Facebook scaled back.

So are we too early to be talking bots? I think it is, but it’s not. The reason is that artificial intelligence is still in the early stages of natural language understanding, and it is still difficult for an algorithm to understand exactly what a human is saying. I say “no” because nowadays, due to the development of dialogue, people have been used to interact on the dialogue platform. As long as we choose the right scene, algorithm and dialogue mode, we can still achieve good accuracy.

How does the dialogue robot land?

Because the current algorithms are not mature enough, companies that make conversational robots need to keep up with any possible achievements in industry and academia and constantly try new optimizations.

This leaves us in a very uncertain state at the level of the algorithm framework. Engineering and algorithms are closely tied together, so engineering needs to provide an algorithm with conditions that are so good that the algorithm can easily try different approaches.

Development domain vs. closed domain

In the chatbot industry, there are basically two schools of problem domains that we solve, one is open domain, the other is closed domain.

Open domain is not limited to what users ask, we have to give a correct answer. For now, the scene does little more than make small talk.

We hope to enable robots to truly understand what human beings say and execute instructions, and to recommend products to users in a personalized way. Under the existing technical framework, it can only be closed. Closed fields can do either retrieval or generative responses.

Closed domain conversations need to clearly define several possibilities for what the human is saying, and all the machine has to do is define what domain the sentence is, what intention it is, and what slot information it needs to know in order to accomplish that intention, which is predefined and a lot of work.

To be able to do this at a relatively low cost, you need to standardize the data structure, extract it automatically through raw corpus analysis, and then incrementally increase the intent.

Topic divergences can be avoided and the content of the user’s conversation is predictable. Steer the conversation to the goal and avoid endless flirting.

Tailor feedback to users’ answers and gracefully break the ice. When there is something abnormal, we can deal with it by acting cute. We need to express positive emotions to users and get their understanding and solutions.

A very significant difference between CUI, which is conversational, and GUI, which is graphic-based, is that CUI is a one-to-one conversation, which is more private. GUI is particularly good for breadth, while CUI is better for depth in conversational scenarios. GUI is more spatial, CUI is more temporal. Because of its privacy, it can also do a lot of personalized operations to strengthen the user’s sense of participation.

During the conversation, the robot needs to remember what the user has said, which is very good for user experience, so it needs a long-term memory in which it can make personalized recommendations.

When the machine can’t accept 100 percent of what the human says, it can switch to the human service. At the same time, when the user input the prompt can input what content, so that the user input content must be in the understandable range, greatly improve the accuracy. Suggestions can be made for replies, even while being served by a human. This in itself is a labeling process that tells the robot whether its response is positive or negative.

There is as much intelligence as there is labor. Not everything is suitable for machines, leaving only repetitive, certain tasks to machines. There are bound to be inaccuracies, and inaccuracies are acceptable as long as we have a good solution when inaccuracies occur so that users can still get things done. Data is AI’s food. Without data, there is no AI. Strong AI is still a long way off.

In the process of practice, we found that the factors that really affect the success of the project, although the algorithm is very important, but the ones I introduced before are not very closely related to the algorithm. A lot of it is product selection, scene selection, dialogue design, and so on, but they are no less important than algorithms.

That’s all for today’s sharing, thank you!