Build a conversational robot in one minute

What is a robot factory?

Since 2016, the world has entered the "Chatbot Era", and technology industry giants have also released their related products in the field of Chatbot, such as Apple Siri, Microsoft Xiaoice, Baidu Xiaodu, Ali's Tmall Genie and so on. The emergence of conversational robots has brought about a revolution in the way users interact with computers, from a sequence of computer-centered orderly operating instructions to user-centered natural language conversational AI.

The robot factory is the Chatbot intelligent robot incubation platform that emerged in the wave of "conversational AI First". With the robot factory, anyone can build an exclusive conversational robot in just one minute, and all kinds of problems such as not knowing NLP technology and not knowing how to program do not exist!

The overall architecture diagram of the robot factory is shown below, and the following will give a detailed introduction around the two major parts of product scenarios and core capabilities.


What product scenarios are used for?

The application scenarios of the robot factory are mainly divided into three categories: answering questions, operations, and operation and maintenance.

AI+ Intelligent Q&A

Answering questions is the most common of the three scenarios. There will always be a large number of common and repetitive questions that need to be answered when users inquire about the products of merchants and employees about the rules and regulations of the company. These common questions are extracted and sorted into question and answer pairs, which are deposited on the robot factory platform. The answering robot can help you Answering most questions saves time and effort, and reduces labor costs. For example: the robot factory provided the conference guidance robot for the Yunqi Conference.

AI+ content scene

The operation scenario mainly uses robots to accurately and quickly reach all users in a unified manner. For example, Xiao Ming operates 100 merchant groups all over the world. He only needs to configure the message content and sending time once by the robot factory, and he can reach 100 merchant groups on time. Easily solve the inefficiency and time delay of 100 manual operations.

AI + daily operation and maintenance

The operation and maintenance scenario is even more powerful. The robot can not only answer questions but also execute commands. For example, if you buy a server on Alibaba Cloud and run some tasks, you can ask the robot to help you query the status of the task execution, find abnormalities, terminate the tasks, and so on. "User questions—understanding instructions—calling services (executing instructions)—returning answers" is the link of the operation and maintenance scenario.

There are also more novel ways to play, such as combining voice-to-text conversion technology, and working with Alibaba Cloud Communications and Cainiao Post to create an intelligent outbound robot. Scenes such as questionnaires, telephone return visits, door-to-door service to confirm whether you are at home, etc. can all be completed by robots, and the number of calls that robots can make in a day is several times more.

What are the core competencies?

Having covered this much, you should have some basic idea of what a robot factory can do. The following will introduce to you the core capabilities of the robot factory and how to play them.

Smart Q&A
The housekeeping skill of intelligent conversational robots is QA questions and answers. Intent and entity are the two most basic concepts. Intent consists of three parts: user input, action, and response. Among them, the user input defines the user's question; the response defines the corresponding answer; the action is optional, and defines a series of instructions that need to be executed after understanding the user's intention. Entity acts on user input, extracts structured information from user input, and efficiently solves the problem of a large number of similar intent matching scenarios. For example, "How's the weather in Hangzhou on March 8, 2020?" Hangzhou can be extracted as an enumerated entity, and 2020-03-08 can be extracted as a regular entity, and the action can be defined as calling a weather query based on the city and time service, and finally responds with weather information.

The above example shows that QA questions and answers must first achieve a precise understanding of user intentions. The bottom layer of the robot factory has a complete set of algorithm frameworks. Traditional machine learning algorithms are combined with deep neural network-based natural language processing algorithms. Offline feature extraction model training is combined with online real-time prediction. Plain text FAQ intent matching and entity-based The intent matching of slots is combined to improve the accuracy of intent matching. I won’t go into details here, and there will be a special article introduction in the future.

Corpus Management

The intelligence of a conversational robot depends to a certain extent on the richness of the corpus it understands. But most of the knowledge is deposited in the form of unstructured text, rather than the intentional form of question and answer that conversational robots can understand. Therefore, how can a newly created conversational robot quickly construct a corpus with the ability of intelligent question answering? The robot factory provides three ways to solve the cold start problem. Within a single robot application, enrich the corpus through corpus crawling (automation) and batch import (manual) and preset intentions (system public corpus); realize corpus between multiple applications Sharing (app assembly).


Corpus crawling

Corpus crawling refers to automatically extracting and sorting the user's existing unstructured knowledge base or documents into a form of question-and-answer pairs that can be understood by conversational robots through machine reading and comprehension capabilities. Corpus crawling can not only replace manual input to quickly enrich the corpus of conversational robots, but also greatly reduce the migration cost of knowledge base docking robot factories. At present, 80% of the corpus on the robot factory platform is generated by corpus crawling. In addition, it also supports manual batch import of corpus in excel or json format to automatically generate intents.

preset intent

The preset intention is that the robot factory sinks high-frequency and common scenarios of users to the platform level, so that it can empower robot applications on all platforms and enhance QA question-answering capabilities. For example, gossiping, checking the weather, checking duty, etc. Users only need to tick the box on the platform to enable their robots to have the ability to answer these questions.

Corpus sharing

Corpus sharing refers to the ability to reuse corpus between different conversational robots. For example, the member manuals of all Intime department stores are the same, but different stores have different discount promotions. The robot factory supports the creation of a robot A with common and universal corpus, and the creation of robots with different corpora, but everyone shares and reuses the corpus of robot A. Corpus sharing can improve the reuse rate of corpus and allow users to focus more on differentiated parts.

online programming

As mentioned above, the intent is composed of three parts: "user input-action-response", where the action defines a series of instructions that need to be executed after understanding the user's intent. Usually an action will call a user-defined service through an HTTP request. However, it is found that the following problems are often encountered:

If the user already has a service interface, the service format does not match; the special processing logic of the robot factory is strongly coupled with the business logic and other problems.
If the user does not have a service interface, a series of processes need to be developed, deployed, jointly debugged, and released, and problems such as machines, networks, and the environment will also be encountered. If the online verification fails, the above steps need to be repeated.
Adding some time-sensitive temporary functions needs to affect the whole body.
In order to solve the above problems, Robot Factory cooperated with AppStudio, the online development platform of Alibaba Cloud Computing Platform, to develop an online service development IDE based on AppStudio, which provides users with an online programming platform on the cloud, and helps you connect with downstream services to realize data query, command execution, Knowledge base retrieval, content recommendation and other functions. can provide you with:

Flexibility: Support online programming, custom business logic, security verification, etc., decoupled from the business system itself;
Openness: SDKs that need to be relied upon can be introduced, and services such as odps and hsf are supported;
Simplicity: encapsulates basic classes and openApi for easy development;
Immediateness: Does not rely on any publishing system, and takes effect as soon as the change is made;
Sharing: support collaborative editing and development, code sharing;
Debugging: support online debugging, service testing and other functions;

The future has come

After 2 years of development and polishing in the Alibaba Group, the robot factory has incubated 20,000+ robots and served 440,000+ users. Participating in the Yunqi Conference for two consecutive years has made us feel the strong appeal of users for intelligent robots. In 2020, the robot factory will officially release the public cloud version. Finally, I will present the demo of the one-click answering robot created by Robot Factory for Feitian big data development platform DataWorks.