Overview of chat bots

Chat robot is a program that simulates human beings through natural language and then talks with people.

1950, Allen m turing published an article about computer machinery and intelligence in mind magazine. At the beginning of this article, he proposed "Can machines think? (Can the machine think? ) ",and put forward the classic Turing test. Passing the Turing test is considered to be the ultimate goal of artificial intelligence research, and Turing himself is therefore called "the father of artificial intelligence".

1966, the earliest chat robot program ELIZA was born. It was developed by Joseph Weizenbaum of Massachusetts Institute of Technology (MIT) and developed a basic script program for clinical simulation of Rogers psychotherapy. The realization technology is only to match the keywords of the text input by the user into the computer and manually write the reply rules.

From 65438 to 0972, American psychiatrist Kenneth Colby wrote a computer program PARRY with LISP to simulate paranoid schizophrenia.

1988, British programmer Rollo Carpenter founded Jabber Wakey, a chat robot. The goal of this project is to "simulate natural man-machine chat in an interesting, entertaining and humorous way". This project is also an early attempt to create an artificial intelligence chat robot by interacting with humans, but Jabber Wakey has not been used to perform any other functions. This technology uses context pattern matching technology to find the most suitable reply content.

From 65438 to 0988, Robert Wilensky of the University of California, Berkeley and others developed a chat robot system called UC(UNIX Consultant). The purpose of UC chat robot is to help users learn UNIX operating system.

From 65438 to 0990, Hugh G. Loebner, an American scientist and philanthropist, established the Loebner Prize, an annual artificial intelligence competition. The Lobner prize aims to test the thinking ability of machines through dialogue. It is regarded as an opportunity to test Turing. Competition prizes are divided into three categories: gold, silver and bronze. So far, no entry procedure has reached the standard of gold or silver award.

Driven by the Lobner Prize, the research on chat bots reached a climax, among which Alice (Artificial Language Internet Computer Entity) was born on February 23rd. Aiml (artificial intelligence markup language) released with ALICE has been widely used in the development of mobile phone virtual assistant.

200 1, the popularity of SmarterChild in short messages and instant messaging tools makes chat bots used in the field of instant messaging for the first time. In 2006, IBM began to develop super brain Watson, which can answer questions in natural language. As a supercomputer based on IBM's "deep question and answer" technology, Watson can use hundreds of algorithms to find the answer to a specific question in 3 seconds.

In 20 10, Apple launched Siri, an artificial intelligence assistant. Siri's technology comes from the CALO plan announced by the Advanced Research Projects Agency of the US Department of Defense: a virtual assistant that simplifies the complex affairs of the military and has the ability of learning, organization and cognition. The civilian version of the software derived from the CALO project is Siri Virtual Personal Assistant.

Since then, chat bots such as Microsoft Xiao Bing, Microsoft Cortana (Xiao Na), Ali Xiaomi, JD.COM Jimi and Netease Seven Fish have emerged one after another, and these chat bots have gradually penetrated into all fields of people's lives.

In 20 16, major companies in China began to launch open platforms or open source architectures that can be used to build chat robot systems.

Since 20 10, the iconic chat robot products are shown in the following figure.

Summary: With the gradual rise of artificial intelligence related technology "Dongfeng", the research on natural language processing has achieved fruitful results, and the related technologies of chat robots have developed rapidly. At the same time, as a novel way of human-computer interaction, chat robot is becoming one of the portals of mobile search and service. After all, the final form of search engines is likely to be chat bots. Many explorers and developers in the field of artificial intelligence want to seize and seize the new interactive portal of chat bots.

The following introduces the classification from the arrangement of several dimensions.

The main function of online customer service chat robot system is to automatically answer questions related to products or services raised by users, thus reducing the operating cost of enterprise customer service and improving the user experience. Representative commercial online customer service chat robot systems include Xiaoyi robot, JD.COM Jimi customer service robot, Ali Xiaomi and so on. Taking JD.COM JIMI customer service robot as an example, users can know the specific information of products, the activity information of the platform and the problems existing in shopping by chatting with JIMI. In addition, JIMI has a certain refusal ability, and can know which questions users can't answer, and can transfer users to manual customer service in time. Alibaba Group released an artificial intelligence shopping assistant virtual robot named "Ali Xiaomi" on July 24th, 20 15. Ali Xiaomi provides a good customer experience based on vertical areas (service, shopping guide, assistant, etc.). ) where customers need it.

The main function of the chat robot system in the entertainment scene is to chat with users without topic restrictions, thus playing the role of companionship and comfort. Its application scenarios focus on social media, children's entertainment, game sparring and other fields. There are representative systems such as Xiao Bing of Microsoft, Xiao Wei of WeChat, Xianer Robot and Monk of Beijing Longquan Temple.

Chat robot system in educational scene can be further divided according to different educational contents. The application scene of this kind of chat robot is a learning and training product with human-computer interaction function, and a children's intelligent toy.

Personal assistant application can interact with users through voice or text, and realize the inquiry and agency of users' personal affairs, such as weather inquiry, short message technology, positioning and route recommendation, alarm clock and schedule reminder, ordering food, etc., so that users can handle daily affairs more conveniently.

The intelligent question-answering chat robot system can answer the factual questions raised by users and other complex questions that need calculation and logical reasoning in the form of natural language, so as to meet users' information needs and help them make decisions. Not only factual questions and answers, such as what, who, which, where, when and so on. It also includes non-factual questions and answers, such as how and why. It should be considered, so the intelligent answering chat robot is usually used as a service module of the chat robot.

From the perspective of implementation, chat bots can be divided into retrieval type and generation type. The answer to the search chat robot is defined in advance. In the process of chatting, the robot uses rule engine, pattern matching or classifier trained by machine learning to select the best answer from the knowledge base and show it to the user. Generative chat robots don't rely on predefined answers, but in the process of training robots, a large number of corpora are needed, including context-related chat information and replies.

Although the search-based chat robot system is generally used to provide chat services in a specific production environment, the emergence of the Seq2Seq model based on deep learning may make the generation-based chat robot system become the mainstream.

Function-based chat robots can be divided into four types: question answering system, task-oriented dialogue system, chat system and active recommendation system.

At present, the evaluation indexes of question answering system and active recommendation system are objective and the evaluation methods are mature. However, given the same input, the task-oriented dialogue system and the filling system can respond in various forms. For the same input of users, there are usually a variety of reasonable and uncertain responses, which makes it difficult to evaluate it through an objective mechanism, so it is necessary to add people's subjective judgment as one of the basis for evaluation.

Usually, a complete chat robot system framework is shown in the figure, which mainly includes five main functional modules: automatic speech recognition, natural language understanding, dialogue management, natural language generation and speech synthesis. It should be pointed out that not all chat robot systems need voice technology.

For example, the chat robot system that realizes human-computer interaction in text mode does not need automatic speech recognition module and speech synthesis module.

Amazon Lex is a service that can use voice and text to build a dialogue interface in any program. Amazon Lex provides an extensible, secure and easy-to-use end-to-end solution for building, publishing and monitoring robots released by developers. The picture below shows how the chat robot can help users complete the demand for flower ordering through dialogue.

Another typical chat bot framework is Wit.ai of Facebook. Wit.ai has accumulated a large number of high-quality conversation data, which effectively promoted the development of chat robot system. Through the combination of artificial intelligence and human intelligence, the intelligent level of chat robot has been further improved.

There are four types of chat bots, including question answering system, task-oriented dialogue system, chat system and active recommendation system.

Siri is positioned as a task-oriented dialogue system, providing users with services such as making phone calls, ordering food, booking tickets and playing music. Siri docked many services and set up "bottom" operations. When Siri can't understand the user's input, it will command the search engine to return to related services. The emergence of Siri has led the commercialization development trend of personal affairs assistants in mobile terminals.

The following figure shows Siri's technical framework:

20 1 1 In February, IBM Watson, which was developed by IBM at a cost of 30 million dollars, appeared on the famous American quiz show Jeopardy. Facing the English problems with double meanings in the program, IBM Watson can analyze and find clues in the huge natural language knowledge base and combine these clues into answers. In the end, IBM Watson defeated the smartest brain in the program with overwhelming advantage, setting the highest score in the 27-year history of this knowledge contest series. As a question answering system developed by IBM, IBM Watson integrates the application of natural language processing, information retrieval, knowledge representation, automatic reasoning, machine learning and other technologies. And form an in-depth question-and-answer technology of hypothetical cognition and large-scale evidence collection, analysis and evaluation. IBM Watson can analyze data in natural language and provide personalized service for users through large-scale learning and reasoning.

20 12 On July 9th, Google released its intelligent personal assistant, Google Now. Google Now provides users with functions such as page search and automatic instruction through natural language interaction. Allo is a voice assistant released by Google on the basis of the above work. Allo can learn user behavior over time.

April 2, 2065 438+04

Active recommendation system adopts a technical way to realize personalized information push. The active recommendation system does not need users to provide clear needs, but builds user portraits by analyzing users' historical behavior data, so as to actively recommend information that the system thinks meets users' interests and needs based on user portraits. It has been widely and successfully applied in e-commerce shopping (such as Alibaba and Amazon), social networks (such as Facebook and Weibo), news information (such as today's headlines), music movies (such as Netease Cloud Music and Douban) and other fields. Active recommendation system is essentially a tool to help people solve the information overload problem. The so-called information overload means that what users really need and are really interested in is submerged in the ocean of similar items. Active interaction can significantly improve the user experience, and the active interaction of robots is closer to the real dialogue between people, making the dialogue more natural.

One way of active recommendation is an active recommendation system based on knowledge graph. For example, when establishing an active recommendation system in the music field, we can first establish the knowledge map of the music field and the user's knowledge map, and then establish the portrait of the user's music preference in the process of user information search, so as to push the user's music more accurately.

As can be seen from the figure, the active recommendation system can give the best music recommendation comprehensively by combining the music knowledge map, the user's personal knowledge map and the user's historical dialogue data in the process of ordering songs.

Active recommendation system and question answering system, task-oriented dialogue system and chat system are considered as four categories of chat robot products.