Realizing the natural interaction between man and machine has always been the common aspiration of human beings all over the world, and countless scientists devote their lives to realizing this goal. Language, as the most effective tool for people to communicate with the outside world, has become the focus of machine intelligence research object, and the conversational robot has become the earliest breakthrough to achieve human-computer interaction.

Since the Turing Test, the quest for conversational robots has never stopped

Ever since Alan Matheson Turing, a pioneer of computer science and cryptography, proposed the Turing test in 1950, the quest to make machines human-like has never stopped. For example, Eliza Chatbot, a psychotherapist developed by MIT in 1966, Alice, a NLP Chatbot based on pattern matching in 1995, Siri, Watson, Google Now, which emerged from 2011 to 2012, Amazon Alexa, Microsoft Cortana and Microsoft Xiaoice, which rose to fame in 2015, as well as Baidu Dumi, Ali Xiaomi, Tencent Cloud Xiaowei, Xiaoai Classmate and Tmall Genie, which started the craze of chat robots in China in 2017, are all achievements of the continuous maturity and commercialization exploration of conversation robots.

In both commercialization and implementation, online intelligent customer service is one of the earliest application fields of conversational robots. The application of dialogue robots in foreign countries is earlier, but due to the more developed domestic e-commerce, the application of intelligent customer service robots in China is more extensive and more mature. There are various commercial products of intelligent customer service, which are widely used in many industries such as medical care, education, intelligent cars and so on.

Easy chat technology intelligent AI customer service, is also one of the outstanding.

One of the first people in the country to get in

Easy to talk about science and technology is the earliest a batch of incoming intelligent customer service in the field of science and technology enterprises, in 2014 started with SaaS and custom from the online customer service, now in more than 20 industries, such as medical treatment, education cultivates intelligent customer service for many years, the domestic market share of more than 80% before robots, high conversion rate is in an industry, such as in the medical field, The conversion rate of subdivision scene could reach: plastic surgery 35%, male 30%, oral (dental) 30%, vitiligo 35%, psychiatric 38%; In the field of education, the conversion rate of segmentation scenarios can be achieved: 55% for qualification, 60% for vocational skills, and 58% for education.

In 2014, foreign bot topic is quite hot, amazon Alexa, Microsoft Cortana, small ice bot often appear on the media headlines, at home, ali, tencent, baidu before domestic technology giants began to large area layout bot, easy to talk about science and technology has pioneered a into the bureau of pre-sale intelligent customer service area, To take the initiative, it has launched two intelligent customer service products: Easy Liao Intelligent online customer service system (IM) and Each Bot AI intelligent reception robot.

Pre-sales robot control logic is more complex than after-sales robot control logic, and it is difficult to catch up with the advantage of backwardness based on large-scale data

In the broad field of customer service robots, Easy chat chose pre-sales robot arrays with more extensive market applications and relatively higher difficulty. Easy Liao intelligent online customer service system provides traditional customer service system solutions for pre-sales data access, lead follow-up control and after-sales customer management. Each Bot AI is based on customer service chat big data as deep learning language. Application of natural language understanding, neural network and other artificial intelligence technology as the core of the commercial application of intelligent marketing service robot.

Schematic diagram of EachBot operation

Here, the difference between pre-sale and post-sale robots must be mentioned. From the technical point of view, the two kinds of robots are very different in focus and difficulty of implementation. In short, the logic control of pre-sale robots is more complex than the latter. This also means getting pre-sales robots right is harder.

The main differences between pre-sale robots and after-sale robots of Yitchat are as follows:

  1. It’s usually multiple rounds of dialogue with more depth. Pre-sale scenario is not just solutions to the problems of sentence, by contrast, after facing the visitors, often has to have a certain understanding, can put forward more specific questions, and organize, pre-sale visitors many cases don’t know what to ask questions, or only puts forward the problem of fuzzy, or initial questions cannot meet their own needs, The question is just the beginning of a series of questions. The pre-sales robot needs to guide the conversation and keep the visitor clarifying the question, so it is not to explain one thing, but to help the visitor find the question and then answer it.
  2. The internal logic of multiple conversations must be tight, regardless of whether the topic is the same. Pre-sales robots must control the essence of the dialogue, with their own goals as the guide, the dialogue always around the theme, or temporarily seem to deviate from the theme, but the internal logic remains unchanged, because the change of the topic is also to achieve the final goal.
  3. Robots need to react more quickly. In the after-sale scenario, visitors have a clear and strong desire to solve problems, so they are relatively tolerant of the response speed of after-sale feedback. In the pre-sale scenario, visitors’ desire is vague, and if they do not respond in time, the loss will be serious.

Because the earliest effective entry into the pre-sale robot field, Easy chat technology in the pre-sale robot field formed a certain technical and product barriers. Stone in the king’s letter, it seems, the three elements of the AI for the algorithm, data, calculate the force, the algorithm and data very close together, the combination of data to a large extent determine the effect of the algorithm, which is characteristic of the era of big data, leading to obtain customer data due to easy to chat, for the first time of optimized algorithm, model, knowledge base, thus improved the performance of the algorithm, Improved customer experience brings more customers and data, which in turn provides more room for algorithm improvement. In this positive cycle, the application of AI products by customers is the optimization of the performance of AI products. This Matthew effect on AI technology is bound to distance Itech from potential competitors.

On the other hand, as more and more players enter the field of intelligent customer service, it is difficult to break down the barrier of first-mover advantage simply through the late-mover advantage of massive data.

Although the algorithm effect is based on data, it cannot fully automate the transformation of data into knowledge and wisdom, and it is difficult to solve pre-sales problems by relying solely on data and computing power. The logic control of pre-sales robots is more complicated than that of post-sales robots, and multiple rounds of dialogues result in insufficient data required after entering a deeper level of activity. In other words, data sparseness exists, which requires human expert knowledge and involves the understanding of customer value. In this regard, our operation department has rich experience.

The problem of insufficient data after such segmentation conditions still exists widely in the era of big data. In order to improve the effect, data segmentation is inevitable, and the relative data shortage has also become inevitable. From a purely theoretical point of view, data and computing power are possible only when the effect of continuous improvement is not considered, that is, when viewed statically. A major goal of algorithm research and development is to solve problems infinitely close to automatic, the path is to use the algorithm to constantly improve human efficiency, reduce human labor.

— Wang Hanshi, chief scientist of Echat

Technical “vanguard corps” to solve core algorithm model problems

How to build this more technical challenge of the system, is the task of the technical team. Led by CTO Bing, a team of natural language processing scientists and machine learning algorithm researchers with 13 years of im technology worked with teams incorporating the latest AI technologies to tackle core algorithm models.

There are many intelligent customer service products on the market, but the real “intelligent”, can achieve the goal of the product is very few, the problem lies in the core system architecture. The transaction rate of the intelligent customer service of Yitchat is more than 60%. The reason why it can guide users to reach a transaction naturally is related to its unique system architecture and core AI algorithm innovation.

System architecture

EachBot AI is developed based on Python, using TensorFlow engine and Apache Spark big data platform. Natural semantic understanding and industry knowledge map can be realized without acosmia sense of art synthesis, meaning toki, several rounds of dialogue, contextual understanding, emotion recognition, as well as the length of time memory network, with automatic induction, characteristic value, omit words and language behavior, etc, which makes the robots could be more flexible, comprehensive, precise and intelligently deal with visitors information. By using Python’s simple and efficient processing logic, flexible and diverse portability and rich scalability, easy chat robot platform is more concise and clear, and it can easily integrate other modules and constantly improve itself.

Easy chat instant messaging underlying system architecture, the use of J2EE system structure, than the industry general. NET/PHP is more flexible and stable. The operating environment is Unix/Linux, which is more stable and secure than Windows. In the communication security using HTTPS and SSL encryption, security is stronger.

The underlying system architecture of instant messaging

Core algorithm innovation: intention recognition, conversation control and knowledge base construction automation

In terms of the core algorithm, many innovations of itchat intelligent customer service system have greatly improved the effect, such as the adoption of automatic methods in intention recognition, conversation control and knowledge base construction.

The correct rate of intention recognition is more than 98%

Generally speaking, intention recognition refers to analyzing visitor discourse, extracting features (semantic labels), and identifying visitor intention and information. Dialogue logic control refers to the transition between scenes, feature migration, and response generation based on more detailed features and rule generation according to intention and visitor information. This process forms a graph structure and controls the flow of dialogue to lead to a given goal, such as linking.

At present, the model method is mainly applied in the part of intention recognition, and improving the recognition accuracy is the goal of model optimization. After 6 iterations of big versions, the recognition accuracy of Yichao pre-sales robot has been improved from 86% to more than 98%, and the recognition accuracy is close to 100% in some scenes.

Combined with BERT and other pre-training models, a number of core algorithm innovations

Our machine learning process is based on a variety of training models, from simple linear discriminant models (such as LinearSVC), integrated discriminant models (such as XgBoost), to more complex deep neural networks (DNN, such as Transformer), dynamic Bayesian networks (DBN, such as coupled Markov chains), Combined with Bert and other pre-trained language models, dependency grammar analysis, semantic role analysis and other computational linguistics methods, and some original model algorithms.

Model algorithm on innovation, of course, is easy to talk the improved performance of key customer service robot, vulnerable to chat since research such as applicable to figure language text clustering algorithms, a variety of linguistic features of hierarchical discriminant algorithm and fusion, in effect at the same time, the time and space complexity control at lower levels, and according to the models to calculate the force difference, with layered structure identification, It reduces the computation cost and the requirement of data scale. Compared with the earlier simple recognition method, the new method has almost no decrease in recognition rate while greatly improving the recognition accuracy.

Self-developed automatic method of constructing knowledge base

, on the other hand, from a more macro perspective, the AI pre-sale robot is a typical expert system, including the knowledge base and reasoning logic, and the expert system is a kind of knowledge engineering and application form, so the pre-sale technical characteristics and difficulties of the robot and knowledge engineering are similar, in addition to the robot operation intention recognition and dialogue in the process of logical control, which in addition to use of knowledge, The real difficulty is how to acquire knowledge and build knowledge base.

The dialogues in the customer service scenario are constantly updated in real time, so what the medical treatment needs to build is a knowledge base of production multi-round dialogues control. Since 2014, Yitchat has accumulated massive dialogues data, and the system needs to mark more than 3 billion dialogues records and more than 500 million dialogues. Such a large volume of conversation data needs to be built into a knowledge base more easily.

Are so easy to talk gave up time-consuming way of pure artificial building, since the research the knowledge base constructing semi-automatic method, such as automatic line lian discriminant, words extraction, extraction such as QA, can without any increase in the premise of human increase efficiency of the building, and provides independent scaffolding and prepared template, will build the function of the knowledge base is open to customers, In order to improve the efficiency and effect of building knowledge base, to meet the personalized needs of customers. At present, Echat says it is developing a less manual method to automatically build a knowledge base, including automatically extracting contextual pairs of questions from the data provided by customers and using them directly for reasoning. The goal is to gradually approach the automatic method and complete the knowledge base construction with very little manpower.

Echat pre-sales robot production knowledge base (currently) includes:

1. 1000000+ semantic tag (features, divided into 100+ version, 1000+ dimension);

2. Nearly 60,000 scenes;

3, nearly 1 million migration generation formula.

Having been deeply engaged in the medical and education industries for more than 7 years, Echat has accumulated industrial knowledge base in these two fields, which enables robots to better understand users’ intentions and improve the conversion rate.

It is worth noting that the customer service of Echitchat smart energy also uses a variety of machine learning technologies, such as active learning, small sample learning, transfer learning, reinforcement learning, etc.

What are these concepts? Let’s give a simple example. First of all, when the data is entered into the learning system, the system will automatically mark the data. If the accuracy of the labeling is not high enough, the training model will be connected, and the algorithm will extract the data with low confidence and close to the discriminant boundary, and then hand it to manual labeling and labeling, and then learn. This is actually a process of active learning. The algorithm automatically screened out the labeling data with poor reliability and handed over to manual correction. By triggering this mechanism repeatedly, more high-quality labeling data could be obtained and the system discrimination accuracy could be improved.

In the aspect of small sample learning, the system firstly uses heuristic rules to mark, and combines the active learning process with manual labeling to correct the labeled samples with poor quality. Finally, the goal of sufficient sample learning can be basically achieved. The transfer learning ability of the model means that the trained model can maintain its processing effect even when there is a difference between the actual data and the developed data.

Principles of reinforcement learning

In addition, easy to chat said is still in research and development applicable to dialogue logic control of the reinforcement learning method, by processing results of a backstepping procedure and their combination, to find the optimal process, the optimization mechanism is very suitable for pre-sale training, dialogical robot for pre-sale scenarios, although easy to judge whether a final purpose, For example, it is difficult to judge whether each line of dialogue and process is reasonable in terms of achieving the final goal.

Innovation of pre-sales robot computing logic

For pre-sale robots, Wang Hanshi also gave a new idea from the level of computational logic. He believes that, from a more modern AI perspective, the pre-sales scene can be viewed as a game between two agents (robots) and visitors (not a zero-sum game), so that the traditional computational logic can be reconstructed using intelligent decision theory. Intelligent decisions based on utility theory of economics, based on AI repetition in the process of this theory, a bayesian network is adopted and the utility function, and the expected utility maximization as the optimization goal of calculation model, the bayesian network through dialogue and data calculation under different response in the context of the specific conditions of probability, the modeling of objective world, The utility function describes the pros and cons of the dialogue results, that is, to model the customer subjective value, and the combination of the two can fit the marketing scene to the maximum extent.

In the process of optimization model, easy chat customer service robot system will use active learning method to easily deviate from the standard (prone to error) of the sample to focus on learning, using reinforcement learning, dialogue results as the starting point, reverse trace the dialogue process of different links and reward and punishment, results-oriented optimization process.

Compared with the original calculation logic and method of building knowledge base, the new calculation logic as based on the nature of the decision making process, is not limited the general method of the industry, not confined to dialogue scenes before or after sales, to adapt to the range up to the whole field, is adapted to the new customers, the problem of the industry, other links can also be used for marketing.

With the support of these AI technologies, Yichao customer service robot system has gradually overcome many technical “problems” in the FIELD of AI, such as ambiguity, emotion recognition, internal representation, potential identification, etc.

The future of conversational robots: Intelligent decision making

At present, the knowledge logic of Echat pre-sales robot has supported a single conversation ability of 30-50 rounds. In the future, Echat said that it will carry out innovation in the dialogue logic, and the biggest change will be the introduction of intelligent decision-making ability. Because dialog control mainly reasoning in knowledge base, is actually the condition judgment under the scenario of migration, the future introduction of intelligent decision-making, it can not only according to the rules of the judgment is or is not, you can also determine the probability have how old problems, working backward from the efficiency in reinforcement learning way, find out the problem behind the probability, to achieve optimum results of dialogue, It is the target of the next stage of easy chat intelligent system.

Review from the first generation since the inquiry system to the system update iteration for many times, now easy to chat system that intelligent robot is the most important is the logic, the logic is the core part of the knowledge, and now the process of building a knowledge, most of the vendors in fact is a large amount of manual intervention, and because of easy incoming chat earlier, the prophase work is adequate, so have first-mover advantages in the industry. Speed up the build, and you have the high ground.

Maintaining this barrier and creating more barriers will require more innovation in an increasingly competitive smart customer service market.

In the future, the real intelligent dialogue robot should be more flexible. At present, the robot has a single purpose, but in the future, it needs to be more intelligent, complete more diversified tasks, and achieve the level of strong artificial intelligence.

— Bing Lixin, CTO of Ecotone Technology

With the advent of the era of big data and cloud computing, conversational robots are experiencing profound changes in technology, application and commercialization, and the goal of realizing a more intelligent human-computer interaction system has become ever clearer. Let us look forward to the next dialogue robot qualitative transformation!

guests

Bing Lixin, Yitao CTO, graduated from Shenyang University of Technology in 2005, and joined Beijing Tianrongxin Network Security Co., LTD in 2005 as a senior development engineer, researching and developing Internet security monitoring products. At the beginning of 2006, DUOyou was founded as a founding partner and served as CTO to develop online customer service products, which is the predecessor of THE IM product of Yicchat. In 2012, this product took the first place in the market, with the turnover of pure IM reaching about 20 million yuan and the share of education industry reaching 70%. In 2014, I started my second business and established Yicchat. In 2018, I adjusted the strategy to focus on SaaS customers. At the beginning of 2018, we launched the first version of AI robot products, and quickly iterated three versions. The AI conversion rate exceeded the human level, and it was able to be on duty independently. So far, AI products have completed six iterations of major versions, and the robot conversion rate has further improved.

Wang Hanshi, chief scientist and AI algorithm leader of Itchat. He graduated from Beijing Institute of Technology in 2011. His research interests include artificial intelligence, natural language processing and intelligent question answering. After six years of teaching at Capital Normal University, Wang Hanshi changed his academic career to industry in 2017. He worked as a data scientist at Xiaomi Intelligent Technology, engaging in structured and unstructured data analysis and modeling, as well as model development in artificial intelligence, natural language processing, machine learning, data mining and other fields. After joining Echat, he served as AI algorithm leader, responsible for the research and development, selection, tuning and testing of statistical models and machine learning models, and applied the research results of AI computing models to echat’s pre-sales robot, SCRM, CRM, outbound call and other businesses to optimize operational efficiency and effect improvement tools.