WO2017041372A1

WO2017041372A1 - Man-machine interaction method and system based on artificial intelligence

Info

Publication number: WO2017041372A1
Application number: PCT/CN2015/096599
Authority: WO
Inventors: 王海峰; 吴华; �田�浩; 赵世奇; 孙雯玉; 吴甜; 忻舟; 马艳军; 吕雅娟
Original assignee: 百度在线网络技术（北京）有限公司
Priority date: 2015-09-07
Filing date: 2015-12-07
Publication date: 2017-03-16
Also published as: CN105068661B; CN105068661A

Abstract

A man-machine interaction method and system based on artificial intelligence. The man-machine interaction method based on artificial intelligence comprises the following steps: receiving input information input by a user through an application terminal (S1); acquiring intent information about the user according to the input information about the user, and distributing the input information to at least one interaction service subsystem according to the intent information (S2); receiving a return result returned by the at least one interaction service subsystem (S3); and generating a user return result according to the return result in a pre-set decision strategy, and providing the user return result to the user (S4). In the method and system, the man-machine interaction system is anthropomorphic instead of being instrumentalized, and the user can obtain relaxed and pleased interaction experience in the intelligent interaction process through chat, research and other service. The search in the form of keywords is improved into natural language based search, the user can express demands through flexible and free natural languages, and a multi-round interaction process is closer to the interaction experience among humans.

Description

Human-computer interaction method and system based on artificial intelligence

Cross-reference to related applications

This application claims the priority of the Chinese patent application No. 201510563338.2, which is filed on September 7, 2015 by Baidu Online Network Technology (Beijing) Co., Ltd. and whose invention is entitled "Human-Computer Interaction Method and System Based on Artificial Intelligence".

Technical field

The invention relates to the field of artificial intelligence technology, in particular to a human-computer interaction method and system based on artificial intelligence.

Background technique

Artificial Intelligence is a branch of computer science, abbreviated as AI. It is a new technical science that studies and develops theories, methods, techniques and application systems for simulating, extending and extending human intelligence.

With the continuous advancement of technology, search engines have become an indispensable part of people's lives, and they are becoming more and more intelligent. At present, the traditional search engine interaction mode is that the user inputs a search keyword, and the search engine returns search results related to the user's needs, and sorts according to the order of relevance from high to low. Users can browse and click on search results and select information and content that are of interest or need. Among them, some search engines use box computing technology and knowledge mapping technology. The box computing technology is mainly that the search engine directly provides results or services for the query keywords input by the user. For example, users searching for "Beijing Weather", "Renminbi Dollar Exchange Rate", "May Day Holiday" and other keywords in the search engine can display the results at the top of the search results page. Knowledge Mapping technology aims to organize and present knowledge related to user needs into a “knowledge map” to meet the needs of users for background knowledge and extended requirements. For example, searching for "Andy Lau", through the knowledge map technology, the search engine can display the background knowledge of Andy Lau's height, birthday, film and television works, as well as "Zhang Xueyou", "Zhu Liqian" and other related figures.

In addition, some search systems can also provide users with the required resources by interacting with users based on natural language. For example, on the smartphone side, users can obtain the required resources through mobile applications such as Apple Siri, Google google now, Baidu Voice Assistant. The above application mainly uses voice as a carrier to issue local services, online search and other instructions to the system in the form of natural language, and feeds the results to the user in the form of voice broadcast.

In addition, users can also ask questions in the deep question and answer system to get the corresponding answer. For example, "Which provinces are the Yellow Rivers?", "Which city is the capital of the United Kingdom?"

However, in the process of implementing the present invention, the inventors have found that at least the following problems exist in the prior art: the current system can only be used to answer simple questions existing in the existing knowledge base, and has high complexity and timeliness. Subjective with the user It is difficult to make an effective answer to the deep questions related to the viewpoint. For a natural language-based search system, after the current topic ends, the system needs to wait for the next topic raised by the user before answering. Due to the lack of information related to the topic, the system cannot actively continue or lead new topics, and cannot continuously interact like people, lacking initiative and association.

Summary of the invention

The present invention aims to solve at least one of the technical problems in the related art to some extent. To this end, an object of the present invention is to provide a human-computer interaction method based on artificial intelligence, which can perform multi-round interaction and search based on natural language, and transform the human-computer interaction system from instrumentalization to anthropomorphic intelligent system.

A second object of the present invention is to provide a human-computer interaction system based on artificial intelligence.

A third object of the invention is to propose an apparatus.

A fourth object of the present invention is to provide a non-volatile computer storage medium.

In order to achieve the above object, an embodiment of the first aspect of the present invention provides a human-computer interaction method based on artificial intelligence, including: receiving input information input by a user through an application terminal; and acquiring an intention of the user according to input information of the user And distributing the input information to the at least one interactive service subsystem according to the intent information; receiving a return result returned by the at least one interactive service subsystem; and generating a user according to the returned result according to a preset decision policy The result is returned and the user returned results are provided to the user.

The human intelligence interaction method based on artificial intelligence in the embodiment of the invention comprises the following advantages: (1) realizing the transformation of the human-computer interaction system from instrumentalization to personification, and letting the user interact in the process of intelligent interaction through chat, search and other services. Get a relaxing and enjoyable interactive experience, not just search and question and answer. (2) From keyword-based search to natural-based search, users can express their needs in a flexible and natural language. The multi-round interaction process is closer to the interactive experience between people. (3) To realize the evolution from user active search to all-weather companion service, based on the user's personalized model, you can provide users with recommendations and other services anytime, anywhere.

The embodiment of the second aspect of the present invention provides a human-computer interaction system based on artificial intelligence, comprising: a first receiving subsystem, configured to receive input information input by a user through an application terminal; and a distribution subsystem, configured to be used according to the user The input information acquires the intent information of the user, and distributes the input information to at least one interactive service subsystem according to the intent information; and the second receiving subsystem is configured to receive the return of the at least one interactive service subsystem Returning a result; generating a subsystem for generating a user return result according to the returned result according to a preset decision policy; and providing a subsystem for providing the user return result to the user.

The human intelligence interaction system based on artificial intelligence in the embodiment of the invention comprises the following advantages: (1) realizing the transformation of the human-computer interaction system from instrumentalization to personification, and allowing users to interact in the process of intelligent interaction through chat, search and other services. Get a relaxing and enjoyable interactive experience, not just search and question and answer. (2) Improved search from keyword form to based on nature Language search, users can use flexible natural language to express demand, multiple rounds of interaction process is closer to the interactive experience between people. (3) To realize the evolution from user active search to all-weather companion service, based on the user's personalized model, you can provide users with recommendations and other services anytime, anywhere.

A third aspect of the present invention provides an apparatus comprising: one or more processors; a memory; one or more programs, the one or more programs being stored in the memory when When the plurality of processors are executed, the artificial intelligence based human-computer interaction method of the first aspect of the present invention is executed.

A fourth aspect of the present invention provides a non-volatile computer storage medium storing one or more programs, when the one or more programs are executed by one device, causing the device An artificial intelligence based human-computer interaction method for implementing the first aspect of the present invention.

DRAWINGS

1 is a flowchart of a human-computer interaction method based on artificial intelligence according to an embodiment of the present invention;

2 is a flowchart of generating a user return result according to a return result according to a preset decision policy according to an embodiment of the present invention;

3 is a flow diagram of a demand fulfillment service subsystem performing steps in accordance with one embodiment of the present invention;

4 is a flow chart showing steps performed by a vertical class service module according to an embodiment of the present invention;

5 is a flowchart of a specific process of obtaining at least one round of interaction with a user to obtain a query result required by a user according to an embodiment of the present invention;

6 is a flowchart of a process of acquiring related information of a corresponding query word according to an embodiment of the present invention;

7 is a first schematic diagram of a user interface including related information of a query term;

8 is a schematic diagram 2 of a user interface including related information of a query word;

9 is a schematic diagram 3 of a user interface including related information of a query word;

10 is a schematic diagram 4 of a user interface including related information of a query term;

11 is a flow chart showing steps performed by a chat service subsystem in accordance with one embodiment of the present invention;

12 is a flow diagram of steps performed by a bootstrap and recommendation service subsystem in accordance with one embodiment of the present invention;

FIG. 13 is a schematic diagram showing the effect of a topic map according to an embodiment of the present invention; FIG.

14 is a schematic diagram showing the effect of network text data as semi-structured data according to an embodiment of the present invention;

15 is a schematic diagram showing the effect of network text data as structured data according to an embodiment of the present invention;

16 is a schematic diagram of an effect of acquiring user browsing behavior data according to an embodiment of the present invention;

17 is a schematic diagram showing the effect of establishing a topic map according to an embodiment of the present invention;

18 is a schematic structural diagram 1 of a human-computer interaction system based on artificial intelligence according to an embodiment of the present invention;

19 is a schematic structural diagram 1 of a generation subsystem according to an embodiment of the present invention;

20 is a second schematic structural diagram of a generation subsystem according to an embodiment of the present invention;

21 is a second schematic structural diagram of a human-computer interaction system based on artificial intelligence according to an embodiment of the present invention;

22 is a schematic structural diagram 3 of a human-computer interaction system based on artificial intelligence according to an embodiment of the present invention;

23 is a fourth structural diagram of a human-computer interaction system based on artificial intelligence according to an embodiment of the present invention;

24 is a schematic structural diagram of a demand satisfaction service subsystem according to an embodiment of the present invention;

25 is a schematic structural diagram of a vertical service module according to an embodiment of the present invention;

26 is a schematic structural diagram 1 of an interaction sub-module according to an embodiment of the present invention;

FIG. 27 is a second schematic structural diagram of an interaction submodule according to an embodiment of the present invention; FIG.

28 is a schematic structural diagram of a fourth acquiring unit according to an embodiment of the present invention;

29 is a schematic structural diagram of a deep question and answer service module according to an embodiment of the present invention;

FIG. 30 is a first schematic structural diagram of a generating submodule according to an embodiment of the present invention; FIG.

FIG. 31 is a second schematic structural diagram of a generating submodule according to an embodiment of the present invention; FIG.

32 is a third schematic structural diagram of a generating submodule according to an embodiment of the present invention;

33 is a schematic structural diagram 1 of an information search service module according to an embodiment of the present invention;

FIG. 34 is a second schematic structural diagram of an information search service module according to an embodiment of the present invention; FIG.

35 is a schematic structural diagram 1 of a chat service subsystem according to an embodiment of the present invention;

36 is a second schematic structural diagram of a chat service subsystem according to an embodiment of the present invention;

37 is a schematic structural diagram of a search-based chat module according to an embodiment of the present invention;

FIG. 38 is a schematic structural diagram of a rich knowledge chat module according to an embodiment of the present invention; FIG.

39 is a schematic structural diagram 1 of a portrait-based chat module according to an embodiment of the present invention;

40 is a second schematic structural diagram of a portrait-based chat module according to an embodiment of the present invention;

FIG. 41 is a third schematic structural diagram of a chat service subsystem according to an embodiment of the present invention; FIG.

FIG. 42 is a schematic structural diagram 4 of a chat service subsystem according to an embodiment of the present invention; FIG.

43 is a schematic structural diagram 5 of a chat service subsystem according to an embodiment of the present invention;

FIG. 44 is a schematic structural diagram 6 of a chat service subsystem according to an embodiment of the present invention; FIG.

45 is a schematic structural diagram 1 of a guidance and recommendation service subsystem according to an embodiment of the present invention;

FIG. 46 is a second schematic structural diagram of a boot and recommendation service subsystem according to an embodiment of the present invention; FIG.

Figure 47 is a third schematic diagram of the structure of the bootstrap and recommendation service subsystem in accordance with one embodiment of the present invention.

detailed description

The embodiments of the present invention are described in detail below, and the examples of the embodiments are illustrated in the accompanying drawings, in which the same or similar reference numerals indicate the same or similar modules or modules having the same or similar functions. The following is described by reference to the drawings The embodiments described are illustrative and are not intended to limit the invention.

An artificial intelligence based human-computer interaction method and system according to an embodiment of the present invention will be described below with reference to the accompanying drawings.

1 is a flow chart of a human-computer interaction method based on artificial intelligence according to an embodiment of the present invention.

As shown in FIG. 1, the human intelligence interaction method based on artificial intelligence may include:

S1. Receive input information input by the user through the application terminal.

The application terminal may include a PC end, a mobile terminal, or an intelligent robot. The input information may be text information, image information, or voice information.

S2: Acquire intent information of the user according to the input information of the user, and distribute the input information to the at least one interactive service subsystem according to the intent information.

The interactive service subsystem may include a requirement satisfaction service subsystem, a guidance and recommendation service subsystem, and a chat service subsystem. In an embodiment of the present invention, the user's intent information may be acquired according to the user's input information, and then the input information is distributed to the above-described interactive service subsystem according to the intent information.

In addition, the user's customized task information may also be received, and the input information may be distributed to at least one interactive service subsystem according to the customized task information. For example, if a user's task only performs a search, then only the customized requirements satisfy the service subsystem; some users' tasks need to search and chat, and the customized requirements satisfy the service subsystem and the chat service subsystem. All of the above can be customized according to the actual needs of users.

S3. Receive a return result returned by at least one interactive service subsystem.

S4. Generate a user return result according to the returned result according to a preset decision policy, and provide the user return result to the user.

Specifically, as shown in FIG. 2, generating a user return result according to the returned result according to a preset decision policy may include the following steps:

S41. Acquire a demand analysis feature of the input information.

S42. Acquire a confidence feature of the returned result returned by the interactive service subsystem, a context feature of the user's dialog interaction information, and a personalized model feature of the user.

S43. Perform a decision on the returned result according to the requirement analysis feature, the confidence feature of the returned result, the context feature of the user's dialogue interaction information, and the user's personalized model feature to determine the user return result.

Specifically, the decision is made on the returned result to determine that the user returns the result mainly based on the following characteristics: 1. The requirement analysis feature, by performing a demand analysis on the user's problem information, selecting a question and answer result provided by the question and answer service module that more closely matches the user's needs. . 2, question and answer results confidence characteristics, each Q & A service module provides a question and answer results with confidence, you can choose a high confidence question and answer results. 3. The contextual characteristics of the user's conversational interaction information can select a question and answer result that is more in line with the context information. 4, the user's personalized model features, you can choose a question and answer results that better meet the user's individual needs. Among them, the characteristics of the demand analysis, the confidence characteristics of the question and answer results, the contextual characteristics of the user's dialogue interaction information, and the use The individualized model features of the households respectively have their own decision weights. Based on the above characteristics, the Q&A results are determined to determine the final question and answer results. After determining the final Q&A result, it can be fed back to the user to meet the user's needs. Among them, the question and answer results can be reported to the user by means of voice broadcast or by means of on-screen display. The use of voice broadcast makes the process of human-computer interaction more convenient and natural.

In addition, according to the user's log, based on the enhanced learning model, the demand analysis feature, the confidence feature of the question and answer result, the context feature of the user's dialogue interaction information, and the decision weight of the user's personalized model feature can be trained to provide the user with more Q&A results that meet user needs.

The demand analysis feature, the confidence feature of the returned result, the contextual feature of the user's dialogue interaction information, and the user's personalized model feature respectively have respective decision weights.

After generating the results returned by the user, the user returns the results into a natural language and broadcasts to the user. Of course, the text corresponding to the result returned by the user can also be directly presented to the user.

In one embodiment of the invention, if an execution instruction is included in the user return result, the execution instruction may be sent to the corresponding execution subsystem and executed by the execution subsystem. The execution instructions may include, but are not limited to, hardware action instructions, playing music instructions, and reading story commands. For example, hardware action instructions are mainly for intelligent robots, which can have hardware components such as heads, torso, limbs, etc., so that control intelligence such as "nodding", "laughing", "lifting hands" can be performed. The instructions of the robot hardware components. Playing music instructions can generally include starting playback, stopping playback, previous, next, big, small, and so on. It should be noted that the user's search for music of a specific type or style (such as "music suitable for listening to bedtime", "song of Jay Chou", etc.) does not belong to the music command. The reading of the story is mainly aimed at children-oriented applications, such as intelligent robots need to replace the parents to tell children stories. Similar to playing music instructions, searching for stories of specific topics, characters, plots, etc. is not a reading of story instructions.

In addition, after receiving the input information input by the user through the application terminal, the above information of the interaction with the user may also be acquired, and then the input information may be complemented according to the information above. Specifically, in a multi-round interaction process, the user usually omits a part of the input information based on the conversation, and therefore needs to complete the input information to clarify the user's needs. For example, the dialogue above is “What snacks are there in Beijing?”, and the input information is “What about special products?”, you need to complete the input information and generate new question information “What special products does Beijing have?”.

In the embodiment of the present invention, if there is no user return result satisfying the user's requirement in the network resource, the input information of the user may be recorded, and then the user returns a result in the network resource that meets the user's requirement in a preset period. And when the user returns a result, the user returns the result to the user. For example, if a user searches for a movie that has just been released, but there is no corresponding resource in the current network resource, the user's demand can be recorded, and the corresponding resource is searched for in the network resource according to a certain period. After the corresponding resource is searched, the resource can be pushed to the user, that is, the asynchronous demand is satisfied.

As shown in Figure 3, the requirements fulfillment service subsystem can perform the following steps:

S31. Obtain problem information input by the user.

The problem information may be text information or voice information. For example, the question information entered by the user "What snacks are there in Beijing?"

S32. Acquire user demand information of the user according to the problem information.

Specifically, the problem information may be analyzed for requirements to obtain user demand information of the user. For example, user demand information can be vertical category requirements, Aladdin requirements, deep question and answer requirements, information search needs, and the like.

S33. Distribute the problem information to the corresponding at least one question and answer service module according to the user demand information.

The question and answer service module may include an Aladdin service module, a vertical service module, a deep question and answer service module, and an information search service module.

In an embodiment of the present invention, when the user requirement information is an Aladdin requirement, the problem information may be distributed to the Aladdin service module; when the user demand information is a vertical type requirement, the problem information may be distributed to the vertical class service module. When the user demand information is a deep question and answer requirement, the problem information may be distributed to the deep question and answer service module; when the user demand information is the information search requirement, the problem may be distributed to the information search service module.

Among them, Aladdin service is a general term for a type of service that can provide users with precise satisfaction, such as the exchange of US dollars to RMB, and the Spring Festival holiday in 2015. For example, if the user's question information is “Who is Andy Lau's wife?”, the problem information can be analyzed, and the demand type can be analyzed as “person”, the query subject is “Andy Lau”, and the query attribute is “wife”. , and can be attributed to the query attribute, the query attribute is grouped as "wife." Then search and obtain the result field as "Zhu Liqian", and then generate a question and answer based on Natural Language Generation (Andy Lau's wife is Zhu Liqian). For another example: the user's question information is “Would it hot tomorrow in Beijing?”. By searching and obtaining the result field as “35 degrees Celsius”, the question and answer result can be generated based on the common sense knowledge base and preset rules. “The weather is very hot tomorrow, the highest temperature. For 35 degrees Celsius, it is recommended to pay attention to heatstroke prevention." Among them, common sense knowledge base can include common sense knowledge, such as temperature above 30 degrees Celsius is hot weather.

The vertical service is a service that performs multiple rounds of interaction for the vertical type of demand, such as "booking a ticket". The service of the vertical class mainly clarifies the user's needs through Dialogue Management and Dialogue Policy, so as to provide users with questions and answers that meet the needs of users. For example, if the user's problem information is “Beijing to Shanghai ticket”, the problem information can be analyzed, and then the user is asked “Which day is your departure date?”, the user answers “Tomorrow” and then continues to ask questions. “Do you have any requirements for airlines?”, etc., gradually clarify the needs of users, and finally return the Q&A results that meet the needs of users.

S34. Receive a question and answer result returned by at least one question and answer service module, and make a decision on the question and answer result to determine a final question and answer result.

Wherein, a decision is made on the result of the question and answer to determine the final question and answer result and the return knot in step S43 The technical means used to make the decision to determine the implementation method of the user return results are consistent, and will not be described here.

Specifically, as shown in FIG. 4, the vertical service module can perform the following steps:

S331. Acquire a query word input by a user.

In one embodiment of the invention, the user can enter query terms in a variety of ways, for example, the user can enter query terms in text, voice, or images.

When the user inputs by voice or image, the input voice or image can be converted into a natural language query word that the user can easily understand, and the corresponding text is displayed on the interactive interface.

For example, after the user inputs the query word by voice, the voice input by the user may be converted into a corresponding text based on the language model, and the query word input by the user is displayed on the interactive interface in the form of a natural language.

S332, determining a vertical class to which the query term belongs.

Specifically, after obtaining the query word input by the user, it is necessary to determine the vertical class to which the query word belongs, so as to facilitate subsequent interaction with the user under the vertical class to which the query word belongs, or obtain relevant information of the query word. At present, the categorization class of the query term can be determined in a variety of ways, and the user can select according to actual needs, as illustrated by the following:

(1) Determine the vertical class to which the query term belongs based on the machine learning method.

Specifically, firstly, the query related to the vertical class is mined and labeled from the search engine log (including the voice search), the training data set related to the vertical class is constructed, and then the training data is extracted and the machine learning classifier is trained (for example, The maximum entropy model and the support vector machine classify the vertical class demand query according to the extracted features to determine the correspondence between the query words and the vertical class, and save the correspondence between the query words and the vertical class.

Among them, it should be noted that in the process of classification, for a plurality of vertical classes, all categories of unified model multi-classification can be used, or each vertical class can be used to separate two models, and finally unified decision-making.

Specifically, after obtaining the query word, the corresponding class of the query word may be determined by the correspondence between the query word and the vertical class. For example, after receiving the query word input by the user as "the novel of the silkworm potato", since the query word includes the author name, the novel and the like, the machine learning method can determine that the vertical class corresponding to the query word is a novel.

(2) Determine the vertical class to which the query term belongs based on the mode analysis method.

In order to determine the vertical class to which the query word belongs based on the mode analysis method, before determining the vertical class to which the query word belongs, for each type of vertical class (for example, novel vertical class, gourmet vertical class, location vertical class, restaurant vertical class, etc.) Build a keyword list and save the correspondence between the vertical class and the keyword.

After receiving the query word input by the user, the entity and the keyword in the query may be parsed based on techniques such as word segmentation and named entity recognition, and the parsing result is used to match the pattern set of the vertical category. If the matching is successful, the corresponding image is sent to the corresponding Vertical category.

Take the restaurant as an example: suppose the user's current query is “quiet restaurant near Sanlitun”. First, the basic lexical analysis of the query, named entity identification, etc., can be determined through analysis. For: [Location]_[Style]_[Restaurant]. Each category separates a collection of patterns. That is to say, for the query to be distributed, firstly, the query is analyzed by basic lexical analysis methods such as word segmentation and named entity recognition, and then the analysis result is matched with the pattern set of the vertical category, and if the matching is successful, the corresponding is distributed to the corresponding Vertical category.

S333, in the vertical class to which the query term belongs, perform at least one round of interaction with the user to obtain a query result required by the user, wherein, in each round of interaction, the information displayed to the user includes: a query result corresponding to the query word, and, Boot information.

In an embodiment of the present invention, in the vertical class to which the query term belongs, a specific process of performing at least one round of interaction with the user to obtain a query result required by the user, as shown in FIG. 5, may include:

S3331, the query word is parsed into structured information that can be represented by the vertical class knowledge system of the vertical class to which the query word belongs.

Among them, the vertical knowledge system of each vertical class is pre-established, and the vertical knowledge system is based on the information provided by the vertical category structured webpage and the user demand representation system.

Among them, the user requirement representation system is a semantic representation system of user requirements. Specifically, semantic and structural knowledge can be mined from the user requirement representation system.

It should be noted that the user requirements are determined based on the query terms. That is to say, the user requirement representation system contains a large number of query words corresponding to the user's needs. By analyzing the query words, the semantic and structural knowledge of the query words can be obtained.

The structural form of the vertical class knowledge system of each vertical class is different. The following is an example to illustrate the structural form of the vertical class knowledge system.

For example, the structure of the vertical knowledge system of the restaurant is shown in Table 1.

Table 1 Structure of the vertical knowledge system of the restaurant

It can be seen from Table 1 that the vertical knowledge system of the restaurant category includes multiple dimensions of information related to the location, cuisine, taste, environment, etc. of each restaurant, as well as possible values of each dimension.

S3332, according to the structured information, the vertical knowledge system, and the vertical class resource library of the vertical class to which the query term belongs, obtain relevant information.

The related information may include, but is not limited to, a query result and a guide information of the corresponding query word.

In an embodiment of the present invention, in order to obtain the vertical class resource of the vertical class to which the query word belongs, before step S3332, the structured resource and the unstructured resource of the vertical class to which the query word belongs may also be acquired and structured. Resources and non Structured resources constitute a library of vertical classes.

The structured resource is a full-scale data resource obtained by fetching integrated data from a plurality of corresponding vertical-type websites, and the unstructured resources are supplemented or expanded according to structured query resources obtained by user query words or Internet text mining.

The following takes the novel as an example to illustrate the process of composing the primitive resources of the novel according to the structured resources and unstructured gestures of the novel.

Generally, the structured resources of the vertical category present a complex architecture. In the process of composing the primitive resources of the novel, the structured resources of the novel can be obtained first. Specifically, the Chinese network can be crawled from the starting point. The information on the novels on the mainstream Chinese novel websites such as Net, Jinjiang, Red Sleeve, 17K Novel Network, and Novel Reading Network establishes a full amount of data resources.

Then, for the unstructured resources of the novel, you can get the novel name, author, category, tag words, resource satisfaction links, novel introduction, novel surrounding and encyclopedic information, and integrate the above resources. Finally, the integrated resources and the above-mentioned full data resources can be saved to the vertical resource library to complete the storage of the vertical resources of the novel.

Among them, it should be understood that, for other vertical classes, the process of obtaining the corresponding vertical class resources is the same as the process of obtaining the vertical class resources of the novel class, and will not be described herein.

In an embodiment of the present invention, the process of acquiring related information of the corresponding query word, as shown in FIG. 6, may include:

S33321: Update the current state information of the user according to the structured information and the previous state information of the user.

According to the common dialogue process in the vertical scene, the state space construction and interactive strategy initialization of the dialogue system are implemented. Specifically, after the user inputs the query for the first time, the initialization state information of the user may be obtained according to the user's preference or the interaction history.

S33322: Generate a candidate action corresponding to the current state information according to the vertical class knowledge system and the vertical class resource library.

The candidate actions may include: an action that satisfies the user's needs, or an action to further clarify the user's needs, or provide horizontal or vertical guidance information for the user's needs. Among them, the user needs are determined according to the query word.

S33323: Select a preset number of candidate actions that match the current state information in the candidate action according to the preset model, and select the selected candidate action as related information.

Specifically, after the candidate action corresponding to the current state information, the degree of matching with the current state information may be selected from the plurality of candidate actions based on a preset model, such as a POMDP (partially observable Markov decision processes) model. a high preset number of candidate actions, and returning the selected candidate action as a query result and guiding information of the query word to the user, and displaying the query result of the query word in the current interface of the application having the dialog function used by the user and Boot information.

Among them, the action that satisfies the user's demand, or the action that further clarifies the user's demand is selected as the query result, and the horizontal or vertical guidance information for the user's demand is selected as the guide information after being selected.

The preset number is preset, for example, the preset number is 5. It is assumed that the candidate action for generating current state information is 10 according to the vertical knowledge system and the vertical resource library, and at this time, the POMDP model can be adopted. Five candidate actions that match the current state information are selected, and the selected candidate actions are returned to the user as related information.

S3333, showing the query result and the guiding information to the user.

S3334, after the user inputs the query word again according to the guiding information, repeat the above process of obtaining relevant information according to the query word, until the query result required by the user is obtained.

In an embodiment of the present invention, the parameters of the preset model may also be updated according to the feedback of the user, so as to select different candidate actions when the parameters are different. That is to say, after the user inputs the query word again, the parameters of the preset model may be adjusted according to the query word input by the user, so that the preset model selects different candidate actions for the user according to the adjusted parameters. That is, the guidance information and the satisfaction information are provided according to the current state information, and the guidance information and the satisfaction information corresponding to the different state information are different, and the system provides the optimal satisfaction information and the guidance information according to the current state information and the user requirements, so as to guide the user to query the vertical information. .

For example, the query word input by the current user is “Western Restaurant”, and it can be determined that the vertical category corresponding to the query word is a gourmet category, and at the same time, the query user can determine that the current user's user demand is to find a western restaurant to eat, due to the query according to the query. The word cannot determine what type of western restaurant the user needs. At this time, according to the vertical knowledge system and the vertical resource library, a plurality of candidate actions can be performed, and 13 candidate actions with a higher degree of matching with the current state information are selected through the POMDP model. The selected 13 candidate actions are returned to the user for related information of the query. The query result displayed in the current user's current interface is as shown in FIG. 7. According to the query word, it is not possible to determine what type of western restaurant the user needs. At this time, the user may be guided to provide more detailed first guiding information, and provide A possible answer corresponding to the guidance information, that is, the second guidance information, is convenient for the user to select or input. Among them, the user can also view other answers corresponding to the first guidance information by clicking the next instruction button. After the user clicks “please ask the customer to eat”, a restaurant that meets the user's needs can be determined according to the query word currently input by the user, and the query result and the guiding information with the current query word are obtained, wherein the interface containing the relevant information of the current query word is obtained. As shown in Figure 8, at this time, the user can make a question based on the guidance information, and ask questions about the restaurant, such as whether there is wifi, whether it is convenient to park or not.

For another example, if the query word input by the current user is "Fiction of Tianshen Potato", after receiving the query word of the user, the semantic analysis can determine the name of the novel including the author of the novel, and the query word can be determined according to the query word. The vertical class is a novel class. At the same time, it can be determined by the query word that the user wants to query the book according to the author name, and the corresponding candidate action can be obtained according to the author name, and the related information corresponding to the query word is displayed in the application used by the user. The interface form of the related information including the query word is as shown in FIG. 9. At this time, the user can click the corresponding book name according to the requirement. In addition, the user also logs in by using the first button, or clears the message record.

For another example, if the query word input by the current user is “a delicious Korean barbecue”, after receiving the query word input by the user, the vertical category corresponding to the query word may be a restaurant gourmet category, specifically, the query may be Word parsing The structured information that can be represented, and the query result and the guiding information corresponding to the query word are obtained according to the structured information, the vertical knowledge system and the vertical class resource library of the vertical class to which the query word belongs, and the obtained query word query is obtained. The result and the guiding information are returned to the user, wherein the user interface including the related information of the query word is as shown in FIG. 10, at this time, the user may select another one according to the guiding information, or directly determine the store according to the demand. In addition, users can also view other boot information by clicking the next prompt button.

In summary, the artificial intelligence-based information query method of the embodiment has the following beneficial effects: (1) Compared with searching through the search engine, the information query mode of the embodiment does not require the user to have a vertical category in the query process. Deeper understanding, through multiple rounds of interaction, guide users to accurately describe the requirements, and provide users with corresponding query results and guidance information according to needs. (2) Compared with the browsing method of the vertical website, the information query mode of the embodiment does not require the user to browse a large number of web pages, and does not need to manually filter useless information. The query method intelligently filters useless information and provides only the user with the query words. Related information.

(3) Comparing the related dialogue system, the information query mode of this embodiment specifically processes the complexity of the vertical category resource structure, and generates a state space based on the vertical class entity structure, which can satisfy the deep level problem in the vertical class. And prompting the user to input the query word again through the guiding information to perform the next round of querying, that is, the information query mode of the implementation can effectively guide the user to provide the correct question by displaying the guiding information.

The deep question and answer service is a service for users to input accurate information based on in-depth semantic analysis and knowledge mining technology. When the user requirement information is a deep question and answer requirement, the deep question and answer service module can receive the problem information, and obtain a corresponding question type according to the problem information, and then select a corresponding question and answer mode according to the question type, and generate a pattern and a problem information according to the selected answer generation mode. Corresponding question and answer results. Among them, the problem type may include an entity type, a viewpoint type, and a fragment type.

More specifically, when the problem type is an entity type, the entity class problem information may be generated according to the problem information, and the entity class problem information is expanded based on the summary and historical presentation log captured by the search engine to generate a cluster problem information cluster. Among them, the information cluster of the same family entity problem corresponds to the candidate answer. Then, the candidate entities are extracted from the candidate answers of the same family entity problem information cluster, and then the confidence degree of the candidate entities is calculated, and the candidate entities whose confidence is greater than the preset reliability threshold are used as the question and answer results. For example, the question information is "Who is the wife of Andy Lau?" The candidate's answer is "In fact, as early as 1992, there were reports that Andy Lau and Zhu Liqian have been secretly registered in Canada...", in which the candidate entity is "Andy Lau" "Zhu Liqian" and "Canada". Then, based on the entity knowledge base and the question-and-answer semantic matching model, the confidence of each candidate entity is calculated. It can be calculated that the confidence level of the candidate entity “Zhu Liqian” is greater than the pre-set reliability threshold, and then “Zhu Liqian” can be determined as the question and answer result. In addition, the phrase "Zhu Liqian" for the first time in the candidate answer can also be used as the answer summary.

When the question type is a viewpoint type, a candidate answer corresponding to the problem information may be obtained, and the candidate answer is segmented to generate a plurality of candidate answer phrases, and then the plurality of candidate answer phrases are aggregated to generate a view aggregation cluster. specifically, The keyword in the candidate answer phrase can be extracted according to the IDF (anti-document frequency) score of the vocabulary in the short sentence, and the keyword containing the negative word is generalized and a negative label is generated, and then the keyword is vectorized based on the negative label Representing that the vector angle and/or semantic similarity between each two keywords is calculated, and then the candidate answers whose vector angle is smaller than the preset angle or whose semantic similarity is greater than a preset threshold are aggregated to generate a view aggregation cluster.

After that, the opinion type of the opinion cluster can be judged. Among them, opinions may include non-classes, evaluation classes, suggestion classes, and the like. Specifically, the opinion type of the opinion aggregation cluster may be determined by a preset rule or based on a statistical model. Then, based on the type of opinion, the answer point is selected from the corresponding view aggregation cluster. Among them, the rules for selecting the answer point of view may include, but are not limited to, selecting the information to cover the most comprehensive answer point of view, selecting the answer point with the lowest IDF*log (IDF) value, and selecting the answer point with the most occurrence in the article corresponding to the candidate answer. Among them, IDF is the anti-document frequency. After that, a summary corresponding to the answer point can be generated, and then the answer point can be scored, and the answer point with the score greater than the preset score threshold is fed back as a question and answer result. For example, the problem information is “Precautions for Pregnancy”. One of the candidate answers is “When you are pregnant, you should follow the principle of medical treatment, and the principle of war, that is, see a doctor regularly, rest in bed, and overcome your bad habits.” The candidate answers are divided into four short answer phrases: "When you are pregnant, you should follow the doctor's principles, and you should follow the principle of regular treatment," "that is, see the doctor regularly," "multiple bed rest," and "to defeat your bad habits." The duplicated content or approximate content in the candidate answer phrase can then be aggregated to generate a view aggregation cluster and the answer opinion can be selected. After that, the answer viewpoint can be scored according to the information richness, the sufficientness of the argument, the information redundancy, and the answer viewpoint whose score is greater than the preset score threshold is used as the question and answer result. In addition, after selecting the answer point, the sentence in the source article can be obtained, and then the sentence is intercepted according to the predetermined length, thereby generating a summary corresponding to the answer point. The summary can then be sorted based on content richness and authoritative answers.

When the problem type is a segment type, a candidate answer corresponding to the problem information may be obtained, and the candidate answer is segmented to generate a plurality of candidate answer short sentences, and then the importance scores are scored for the plurality of candidate answer short sentences to generate the candidate answer short The short sentence importance feature of the sentence is generated, and the answer summary is generated according to the short sentence importance feature, and then the answer quality can be based on the short sentence importance feature of the answer summary, the authority of the answer, the relevance of the question information, and the richness of the answer. Score. The short sentence importance feature may include an aggregate feature, a relevance feature, a type feature, and a question answer matching feature. Among them, the aggregation feature is used to measure the repetition of short sentences, such as: word vector centroid feature, NGram (calculated probability of occurrence) feature, Lexrank (multi-text automatic summary) feature. Type characteristics are type characteristics of the problem, such as WHAT (what) type, WHY (why) type, HOW (how) type, and so on. The authoritativeness of the answer is the authority of the website from which the answer originated. After that, the user's behavior data can be obtained, and then the candidate answers are sorted according to the user's behavior data and the scoring result, and finally the sorting result is fed back as a question and answer result. The user's behavior data is historical behavior information of the user, including the user's click behavior on the question and answer result, the time spent on the question and answer result, and the current question and answer result jump to other question and answer results.

When the user demand information is information search demand, the information search service module can receive the problem information and according to the problem The information is searched to generate a plurality of candidate web pages, and then the candidate web pages are subjected to text analysis to generate corresponding candidate chapters. Specifically, the candidate web page may be subjected to chapter content extraction, chapter topic segmentation, and chapter relationship analysis to generate corresponding candidate chapters. Among them, the chapter content extraction mainly identifies the body part of the candidate webpage, and deletes the content irrelevant to the user demand information. The chapter topic is divided into the analysis of the topic structure of the chapter, which can divide the chapter into multiple sub-themes. Chapter relationship analysis is to analyze the relationship between multiple subtopics in a chapter, such as a side-by-side relationship. After the candidate chapter is generated, the sentences in the candidate chapter can be scored. Among them, the scoring order is mainly based on the importance of the sentence in the candidate chapter and the correlation between the sentence and the user's demand information. After that, the user's demand scenario information can be obtained, and a summary is generated according to the demand scenario information and the score sorting result, and finally the summary is fed back as a question and answer result. The scene information may include a mobile terminal scene and a computer scene. When the scene information is a mobile terminal scene, the sentence can be compressed and abbreviated, so that the generated summary is as concise as possible; when the scene information is a computer scene, the sentences can be spliced and merged, so that the generated abstract is clear and detailed. Of course, when the candidate chapter is generated, since the content in the candidate chapter is related to the user requirement information, there may be repeated or complementary content, and the information of the plurality of candidate chapters needs to be aggregated.

As shown in Figure 11, the chat service subsystem can perform the following steps:

S51. Receive input information input by a user.

The input information may be voice information or text information.

After receiving the input information input by the user, the input information may be corrected and/or rewritten, used to correct the typos in the input information, and to rewrite the irregular colloquial expression.

In addition, the above information that is chatted with the user may also be obtained, and then according to the above information, it is determined whether the dependency of the input information and the above information is greater than a preset relationship threshold. If it is greater than the preset relationship threshold, the input information can be complemented according to the above information, thereby ensuring the smoothness of the human-machine chat. Specifically, complementing the input information may include referring to the digestion. For example, if the input information is "Has he got married?", he can replace "He" in the input information with "Andy Lau" according to the above information "Andy Lau". Completing the input information may also include omitting the completion. For example, the above information "Andy Lau's wife is called Zhu Liqian.", the input information is "I don't know.", then the input information can be completed as "I don't know Zhu Liqian."

In addition, the current topic information of the user may be obtained according to the above information, so that the subsequent chat service module guides the chat topic.

S52. Distribute the input information to the chat service module.

Specifically, domain analysis may be performed on the input information to obtain an area corresponding to the input information. The input information can then be distributed to chat service modules having the same or similar fields based on the fields corresponding to the input information.

The chat service module may include one or more of a search-based chat module, a rich knowledge chat module, a portrait-based chat module, and a crowdsourcing-based chat module.

Specifically, the search-based chat module may perform a word-cutting on the input information to generate a plurality of chat phrases, and then query the chat corpus according to the plurality of chat phrases to generate a plurality of chat corpus sentences and multiple chat corpus upper sentences. More Chat corpus. The chat corpus is pre-established, and the source of the chat can include, but is not limited to, “posting-replying” in the forum data such as posting bar, “blog-reply” in the microblog, and “question-answer” in the question and answer community.

After that, multiple chat corpus sentences can be filtered. Specifically, the similarity between the input information and the sentences of the plurality of chat corpora can be calculated. If the similarity is less than the first preset similarity threshold, the corresponding chat corpus can be filtered; if the similarity is greater than or equal to the first preset similarity threshold, the corresponding chat corpus can be retained.

After filtering the upper corpus of the chat corpus, the chat corpus corresponding to the sentence in the chat corpus after filtering may be classified. Specifically, the similarity between the input information and the plurality of chat corpus sentences is calculated, and the machine based on the similarity is based on GBDT (Gradient Boost Decision Tree), SVM (Support Vector Machine), and the like. The learning model classifies multiple chat corpora sentences. The similarity between the input information and the sentences of the plurality of chat corpora may be a literal similarity between the input information and the sentence of the chat corpus, or may be similar to the input information and the chat corpus based on the deep neural network training. Degree can also be the similarity between the input information and the chat corpus based on the machine translation model. It should be understood that the similarity between the input information and the plurality of chat corpus in the present embodiment and the machine learning model such as GBDT and SVM are well-known technologies, and are not described herein.

Then, the search-based chat module can reorder the chat corpus under the classification, and generate a candidate reply according to the sort result. Specifically, the chat attribute of the user may be obtained according to the above information of the user chat, and then the chat quotation sentence after the classification is reordered based on the learning order model (Learning-To-Rank) according to the chat attribute. The chat attribute may include a chat occasion such as a time and place, a chat fun, a chat style, and the like. Of course, the chat attribute is not limited to the above information from the user chat, but also can be obtained according to the user's long-term historical chat record. It should be understood that the learning ordering model in this embodiment is a well-known technology, and details are not described herein.

The rich knowledge chat module may generate a search word according to the input information, and perform a search according to the search word to generate a plurality of search results, and then perform sentence extraction on the plurality of search results to obtain a similarity with the search word greater than the second preset similarity. A set of candidate sentences for a sentence of a threshold. After that, the sentences in the set of candidate sentences can be rewritten to generate candidate responses. In addition, sentences in the candidate sentence set may be reordered according to the user's chat attribute. For example, if the input information is "I hope to have the opportunity to travel to Mount Fuji", the input information can be parsed and the corresponding search term "Mount Fuji, Tourism" can be generated, and then multiple search results can be obtained according to the search term, and the search words can be extracted and searched. A sentence with high similarity. Among them, some sentences may include obvious texts such as “journalists”, so these sentences need to be rewritten to make them more fluid, more like the sentences of natural language chat, and the final candidate for the response is “Mt. Fuji due to Due to the weather, there is only a certain period of summer in the year to climb the mountain. Compared with the traditional reply, "I also want to go to Mount Fuji, let's go together.", with certain knowledge and timeliness, can enable users to Get useful knowledge during the chat.

In order to better realize anthropomorphization and provide personalized services for users, the human-machine chat system can set its own attributes, status, interests, etc., that is, the system portrait model. You can also set the user's attributes, status, interests, etc. model. Of course, in the case of facing different users, the system portrait model used may be the same, or a system image model corresponding to each user may be set. Both the system portrait model and the user portrait model are based on the image knowledge map. The portrait knowledge map is a hierarchical knowledge system. For example, the "family member" node may include two child nodes "brothers" and "parents", and the "parent" child nodes include two child nodes "father" and "mother". Each node corresponds to a plurality of input information template clusters, such as "Who is your father", "Who is your father", and "What is your father's name" belongs to the same input information template cluster. Each input information template cluster corresponds to one or more candidate responses. The input information template cluster and the candidate reply may include variables, such as interest, hobbies, and preferences corresponding to the same attribute "INTEREST", and the attribute values of "INTEREST" may include climbing, music, reading, sports, and the like.

Specifically, the portrait-based chat module may acquire the chat context of the user, and determine whether the collection condition is satisfied according to the chat context. If it is determined that the collection conditions are met, the question can be sent to the user. After that, the user's answer information according to the question can be received, and the user portrait model is updated according to the answer information. For example, when talking to a user about a movie-related topic, you can send the question "What movie do you like?" or the user asks the human-machine chat system "What do you like to eat?", the human-machine chat system can ask the user "You like to eat What? After the user answers, the user portrait model can be updated based on the user's answer information, which is more in line with the user's personalized needs.

In addition, the portrait-based chat module may also acquire the chat content of the user, extract user image data according to the chat content, and then update the user portrait model according to the extracted user portrait data. For example, the user said in the chat process, "I like to climb the mountain and catch fishing when I am fine." I can extract the user portrait data "hobby climbing, hobby fishing" to update the user portrait model. At the same time, an appropriate answer can be extracted based on the user's portrait data, and the appropriate answer information can be returned to the user.

Crowdsourcing is a method of outsourcing specific tasks to non-specific users on the Internet. For human-machine chat, the problem that the machine is difficult to answer can be distributed to the performer to manually respond in real time to meet the user's needs. Actual demand.

Specifically, the crowdsourcing-based chat module can determine whether the input information is suitable for crowdsourcing completion, for example, if the user's mood is low and needs comfort, etc., it is suitable for crowdsourcing completion. For example, if the user's input information contains personal information such as personal identification information, password, and telephone, it is not suitable for crowdsourcing.

If it is judged that the crowdsourcing is completed, the input information can be distributed to the corresponding performer. Of course, the above information can also be sent to the performer together, and the performer can reply according to the above information and input information. The crowd-based chat module can then receive the reply information of the performer and make a quality judgment on the reply information. If the quality requirement is met, the reply message is used as a candidate reply. For example, if the reply message contains vulgar, reactionary, or pornographic content, the quality is not enough. Or if the executor replies for more than the predetermined duration, the replies of the executor will not be used, and the reply information can be saved to the chat corpus.

In addition to this, it can also be judged whether the input information belongs to chat information without actual content, such as "hehe", "hoho" Wait. If it is determined that the chat information belongs to the content without the actual content, the current topic can be obtained, that is, the current topic is calculated according to the historical chat record based on the Topic Model. After acquiring the current topic, a guided topic may be generated based on the current topic based on the topic chat map. Among them, the topic chat map is a directed graph with a topic as a node. For example, as shown in FIG. 2, the node "leisure" can point to the node "watching movie" and the node "listening to the song", indicating that the topic "leisure" can be guided to the topic "watching movie" or the topic "listening to the song". The topic "watching movies" and the topic "listening songs" all have a certain guiding probability, and the guiding of the topic can be realized according to the guiding probability, thereby ensuring the diversity of guiding topics.

A candidate reply can then be generated based on the guided topic. Specifically, the template of the candidate reply may be generated based on the Natural Language Generation, and the guide topic may be filled into the template to generate a candidate reply; or the sentence including the guide topic may be selected as a candidate reply, thereby implementing the user. Actively chat topic guide.

S53. Receive a candidate reply returned by multiple chat service modules.

Among them, the candidate reply has a corresponding confidence.

S54. Sort the responses according to the confidence degree, and generate chat information according to the sorting result, and provide the chat information to the user.

Specifically, the characteristics of the user's input information may be acquired, and the selected responses are sorted based on the characteristics and confidence of the input information. The characteristics of the input information may include classification features, literal features, topic features, and the like. The higher the confidence, the better the quality of the candidate reply is. The responses can be sorted according to the order of confidence, and finally the chat information that meets the user's needs is provided to the user.

In addition, the Reinforcement Learning can also be updated according to the feedback information of the user, so that the user can be provided with more satisfactory chat information. For example, adding a comment button such as “Like” or “Tread” to the user's chat information to collect feedback information of the user; or analyzing the input information of the user in the chat based on the sentiment analysis technology, thereby obtaining the user's evaluation, For example: "You are really smart", etc.; or judge the user's satisfaction by recording the number of interactions with the user.

As shown in Figure 12, the boot and recommendation service subsystem can perform the following steps:

S61. Receive interaction information input by the user, and determine a current topic according to the interaction information.

Specifically, the interaction information input by the user may be received first, for example: “Is the dream space look good?”, and then the interaction information is subjected to requirement identification and correlation calculation, thereby determining that the current topic is “the dream space evaluation”.

S62. Obtain a plurality of candidate guidance topics related to the current topic based on the topic map.

The topic map may include a plurality of topics and associations between topics.

Specifically, a plurality of candidate guidance topics related to the current topic may be acquired based on the pre-established topic map. For example, if the current topic is "Privacy Dream Space Evaluation", you can obtain a number of guiding topics related to "Pirates of Dreams" based on the topic map, such as "Nolan's film", "Leonardo's movie", etc. And their relationship with the "dream space evaluation".

S63. Acquire user image data of the user.

The user portrait data is a collection of data such as attributes, states, interests, and the like of the user, and can be acquired by the user actively input or according to the historical interaction record of the user, and then integrated, thereby generating personalized user portrait data about the user.

S64. Select a guiding topic from a plurality of candidate guiding topics related to the current topic according to the user portrait data, and feed back the guiding topic to the user.

Specifically, the user's intention information may be determined according to the user image data and the context information of the interaction information, and then the guidance topic is selected from a plurality of candidate guidance topics related to the current topic according to the user's intention information, and the guidance topic is fed back to the user.

For example, a guided topic can be an extension of the current topic. For example, if the interactive information is “How do chickens do?”, the current topic can be “Chicken Practice”. After the current topic interaction ends, the current topic can be extended. In combination with the user portrait data such as “user is a pregnant woman”, the user can be fed back to the topic “How to eat chicken in a pregnant woman is better”.

Of course, the guiding topic can also be a recommendation based on the current topic. For example, if the interactive information is “Is the dream space to look good?”, the current topic can be “the dream space evaluation”. After the current topic interaction ends, based on the current topic, and combined with user portrait data such as "users like to watch movies", the user can be fed back to guide the topic "Nolan's movie."

When the user's intention information cannot be determined based on the user image data and the context information of the interaction information, the user's intention information needs to be clarified. For example, the interactive information is “How to go to the Forbidden City?”, while Beijing, Shenyang, and Taipei all have “Forbidden City”. Therefore, it is necessary to clarify the user’s intention information, and can return a question of clarification to the user according to the interactive information. “Excuse me. Which is the Forbidden City to go to?".

In addition, before step S62, step S65 may also be performed.

S65. Establish a topic map.

As shown in FIG. 13, one node in the topic map represents a topic or a requirement put forward by the user, and each node may include a reply of the corresponding topic and a resource satisfying the user's requirement, and the associated node may pass the edge. Correlation is made to form a meshed topic map.

Specifically, the method for establishing the topic map is as follows: the topic association data can be obtained, and then the topic map is established according to the topic association data.

More specifically, obtaining topic related data can be divided into two cases.

The first case: the network text data can be obtained first, and the topic related data is obtained from the network text data. Among them, network text data can be divided into unstructured data, semi-structured data and structured data.

When the network text data is unstructured data, the topic association data can be obtained based on entity extraction and syntax analysis. Among them, unstructured data may include news, forums, blogs, videos, and the like. For example: for the text text data "the most eye-catching Nobel Prize for Literature is the winner, the Frenchman Modiano becomes the lucky winner of the new class. Of course, multiple nominations are always Murakami Haruki, who missed the award, is still the "most recent person who has won the Nobel Prize." The Chinese poet North Island has only made the Chinese people crazy. "Based on entity extraction technology, the entity information "Nobel Prize for Literature", "French Modiano", "Murao Haruki", "Chinese poet North Island" can be extracted, and based on syntactic analysis, the existence of the above entity information is known. Further, it can be analyzed that the Frenchman Modiano is a Nobel laureate in literature, and Haruki Murakami and the Chinese poet North Island did not receive the Nobel Prize in Literature.

When the network text data is semi-structured data, the topic-related data is obtained based on page structure analysis, tag extraction, and entity recognition. Among them, the semi-structured data may include Wikipedia, Baidu Encyclopedia and other encyclopedic data, or thematic data. For example, as shown in FIG. 14, the "off-the-life" of "Djokovic", including "family life" and "charity activities", can be obtained based on page structure analysis, tag extraction, and entity recognition.

When the network text data is structured data, the topic related data is obtained from the knowledge map. Among them, the structured data may include knowledge map data. For example, as shown in Figure 15, the director of the movie "Pirates of the Dream" and the movie "Interstellar Crossing" is "Christopher Nolan".

The second case: the user's search behavior data or browsing behavior data may be obtained first, and then the topic association data is generated according to the search behavior data or the browsing behavior data.

Specifically, the search behavior data of the user may be acquired, and the corresponding search object is obtained according to the search behavior data, and then the topic association data is generated according to the search object. For example, if the user continuously searches for "Nolan", "Nolan's Movie" and "Christian Bell", the above topics can be correlated to generate topic related data.

Certainly, the browsing behavior data of the user may also be obtained, and the corresponding browsing object is obtained according to the browsing behavior data, and the topic related data is generated according to the browsing object. For example, as shown in FIG. 16, a plurality of news or videos that are clicked when a user browses a webpage may be associated to generate topic related data.

After the topic association data is obtained, the topic map may be established according to the topic association data by one or more of the RandomWalk algorithm, the association analysis algorithm, and the collaborative filtering algorithm. For example, as shown in Fig. 17, q1, q2, q3, and q1', q2', q3', and q4' are topics, and d1, d2, d3, and d4 are resource data. As can be seen from FIG. 13, the resource data d1 and d2 are associated with the topic q1; the resource data d1, d2, d3 are associated with the topic q2; the resource data d4 is associated with the topic q3, and the topic with the associated relationship and the resource data are used. Solid lines are connected. Based on the RandomWalk algorithm, it is possible to iteratively calculate the relationship between the topic q1 and the resource data d3, which are connected by a dotted line. The topic q1' is a topic that the user issues after the resource data d1 or d4 is browsed according to the resource data d1 or d4, and the relationship between them has a sequential relationship. Similarly, the topic q2' is a topic issued based on the resource data d2, the topic q3' is a topic issued based on the resource data d2 or d3, and the topic q4' is a topic issued based on the resource data d3 or d4. Further, it can be inferred that the topic q1 and the topic q1' have an association relationship, the topic q1 and the topic q2' have an association relationship, etc., and finally the topic map as shown in Fig. 13 is established.

In addition, the guidance and recommendation service subsystem can also parse the interaction information and obtain keywords in the interaction information. segment. The keyword segment may include one or more of time information, location information, and reminder events. Then, the reminder information can be established according to the keyword segment, and when the time information reaches the preset time, the reminder information can be sent to the user. For example, if the user's interaction information is "Remind me to write a work plan at 6 o'clock tomorrow night", the time information "18:00 on August 8, 2015" can be parsed, and the reminder event is "write work plan". When this time is reached, the user can be alerted.

To achieve the above object, the present invention also proposes a human-computer interaction system based on artificial intelligence.

FIG. 18 is a first schematic structural diagram of a human-computer interaction system based on artificial intelligence according to an embodiment of the present invention.

As shown in FIG. 18, the artificial intelligence based human-computer interaction system may include: a first receiving subsystem 10000, a distribution subsystem 20000, an interactive service subsystem 30000, a second receiving subsystem 40000, a generating subsystem 50000, and a provider. System 60000.

The first receiving subsystem 10000 can receive input information input by the user through the application terminal.

The distribution subsystem 20000 may acquire the user's intent information according to the user's input information, and distribute the input information to the at least one interactive service subsystem 30000 according to the intent information.

The interactive service subsystem may include a demand fulfillment service subsystem 31000, a chat service subsystem 32000, and a boot and recommendation service subsystem 33000. In one embodiment of the present invention, the distribution subsystem 20000 may acquire the user's intent information according to the user's input information, and then distribute the input information to the interactive service subsystem 30000 according to the intent information.

In addition, the first receiving subsystem 10000 can also receive customized task information of the user, and the distribution subsystem 2000 can distribute the input information to the at least one interactive service subsystem 30000 according to the customized task information. For example, if a user's task only performs a search, then only the customized requirements satisfy the service subsystem; some users' tasks need to search and chat, and the customized requirements satisfy the service subsystem and the chat service subsystem. All of the above can be customized according to the actual needs of users.

The second receiving subsystem 40000 can receive the returned result returned by the at least one interactive service subsystem 30000.

The generating subsystem 50000 may generate a user return result according to the returned result according to a preset decision policy.

A provisioning subsystem 60000 can provide user return results to the user.

As shown in FIG. 19, the generation subsystem 50000 can include a first acquisition module 51000, a second acquisition module 52000, and a first decision module 53000.

The first obtaining module 51000 can acquire a demand analysis feature of the input information.

The second obtaining module 52000 can obtain a confidence feature of the returned result returned by the interactive service subsystem, a context feature of the user's dialog interaction information, and a personalized model feature of the user.

The first decision module 53000 can make a decision on the returned result according to the requirement analysis feature, the confidence feature of the returned result, the context feature of the user's dialogue interaction information, and the personalized model feature of the user to determine the return result of the user.

Specifically, the decision is made on the returned result to determine that the user returns the result mainly based on the following characteristics: 1. The requirement analysis feature, by performing a demand analysis on the user's problem information, selecting a question and answer result provided by the question and answer service module that more closely matches the user's needs. . 2, question and answer results confidence characteristics, each Q & A service module provides a question and answer results with confidence, you can choose a high confidence question and answer results. 3. The contextual characteristics of the user's conversational interaction information can select a question and answer result that is more in line with the context information. 4, the user's personalized model features, you can choose a question and answer results that better meet the user's individual needs. Among them, the demand analysis feature, the confidence feature of the question and answer result, the contextual feature of the user's dialogue interaction information, and the user's personalized model feature respectively have their respective decision weights. Based on the above characteristics, the Q&A results are determined to determine the final question and answer results. After determining the final Q&A result, it can be fed back to the user to meet the user's needs. Among them, the question and answer results can be reported to the user by means of voice broadcast or by means of on-screen display. The use of voice broadcast makes the process of human-computer interaction more convenient and natural.

Additionally, as shown in FIG. 20, the generation subsystem 50000 can also train the module 54000. The training module 54000 can train the user based on the enhanced learning model based on the enhanced learning model, the confidence feature of the question and answer result, the contextual feature of the user's dialogue interaction information, and the decision weight of the user's personalized model feature. Q&A results that meet user needs.

After generating the user return result, the provisioning subsystem 60000 can convert the user return result into a natural language and broadcast it to the user. Of course, the text corresponding to the result returned by the user can also be directly presented to the user.

In an embodiment of the present invention, as shown in FIG. 21, the artificial intelligence based human-computer interaction system may further include a transmitting subsystem 70000 and an execution subsystem 80000.

If the user returns a result including an execution instruction, the transmitting subsystem 70000 can send the execution instruction to the corresponding execution subsystem 80000, which can execute the execution instruction. The execution instructions may include, but are not limited to, hardware action instructions, playing music instructions, and reading story commands. For example, hardware action instructions are mainly for intelligent robots, which can have hardware components such as heads, torso, limbs, etc., so that control intelligence such as "nodding", "laughing", "lifting hands" can be performed. The instructions of the robot hardware components. Playing music instructions can generally include starting playback, stopping playback, previous, next, big, small, and so on. It should be noted, The user's search for music of a particular type or style (such as "music for listening to bedtime", "good song for Jay Chou", etc.) is not a music instruction. The reading of the story is mainly aimed at children-oriented applications, such as intelligent robots need to replace the parents to tell children stories. Similar to playing music instructions, searching for stories of specific topics, characters, plots, etc. is not a reading of story instructions.

In an embodiment of the present invention, as shown in FIG. 22, the artificial intelligence based human-computer interaction system may further include a complement subsystem 90000.

After receiving the input information input by the user through the application terminal, the completion subsystem 90000 can acquire the above information of the interaction with the user, and then can complete the input information according to the above information. Specifically, in a multi-round interaction process, the user usually omits a part of the input information based on the conversation, and therefore needs to complete the input information to clarify the user's needs. For example, the dialogue above is “What snacks are there in Beijing?”, and the input information is “What about special products?”, you need to complete the input information and generate new question information “What special products does Beijing have?”.

In an embodiment of the present invention, as shown in FIG. 23, the artificial intelligence based human-computer interaction system may further include a recording subsystem 100000 and a monitoring subsystem 110000.

If there is no user return result in the network resource that satisfies the user's requirement, the recording subsystem 100000 can record the user's input information, and then the monitoring subsystem 110000 monitors, in a preset period, whether there is a user return result satisfying the user's requirement in the network resource, and The provisioning subsystem 60000 can provide the user return results to the user when the user returns a result. For example, if a user searches for a movie that has just been released, but there is no corresponding resource in the current network resource, the user's demand can be recorded, and the corresponding resource is searched for in the network resource according to a certain period. After the corresponding resource is searched, the resource can be pushed to the user, that is, the asynchronous demand is satisfied.

As shown in FIG. 24, the demand satisfaction service subsystem 31000 may include a third acquisition module 31100, a fourth acquisition module 31200, a first distribution module 31300, a question and answer service module 31400, and a second decision module 31500.

The third obtaining module 31100 can acquire problem information input by the user. The problem information may be text information or voice information. For example, the question information entered by the user "What snacks are there in Beijing?"

The fourth obtaining module 31200 can acquire user demand information of the user according to the problem information. Specifically, the problem information may be analyzed for requirements to obtain user demand information of the user. For example, user demand information can be vertical category requirements, Aladdin requirements, deep question and answer requirements, information search needs, and the like.

The first distribution module 31300 may distribute the problem information to the corresponding at least one question and answer service module 31400 according to the user requirement information. The Q&A service module may include an Aladdin service module 31410, a vertical class service module 31420, a deep question and answer service module 31430, and an information search service module 31440.

In an embodiment of the present invention, when the user demand information is an Aladdin requirement, the problem information may be distributed to the Aladdin service module 31410; when the user demand information is a vertical demand, the problem information may be distributed to the vertical service. Module 31420; when the user requirement information is a deep question and answer requirement, the problem information may be distributed to the deep question and answer service module 31430; When the user demand information is an information search request, the problem can be distributed to the information search service module 31440.

Among them, Aladdin service is a general term for a type of service that can provide users with precise satisfaction, such as the exchange of US dollars to RMB, and the Spring Festival holiday in 2015. For example, if the user's question information is "Who is Andy Lau's wife?", the Aladdin service module 31410 can analyze the problem information, and analyze the demand type as "person", and the query subject is "Andy Lau", query The attribute is "wife", and the query attributes can be normalized, and the query attributes are grouped into "wife". Then search and obtain the result field as "Zhu Liqian", and then generate a question and answer based on Natural Language Generation (Andy Lau's wife is Zhu Liqian). For another example: the user's question information is “Would it hot tomorrow in Beijing?”. By searching and obtaining the result field as “35 degrees Celsius”, the question and answer result can be generated based on the common sense knowledge base and preset rules. “The weather is very hot tomorrow, the highest temperature. For 35 degrees Celsius, it is recommended to pay attention to heatstroke prevention." Among them, common sense knowledge base can include common sense knowledge, such as temperature above 30 degrees Celsius is hot weather.

As shown in FIG. 25, the drop class service module 31420 can include a fifth get submodule 31421, a determine submodule 31422, and an interaction submodule 31423.

The fifth obtaining sub-module 31421 can obtain the query word input by the user.

The determination sub-module 31422 can determine the vertical class to which the query term belongs.

Specifically, firstly, mining and annotating the query related to the vertical class from the search engine log (including the voice search), constructing the training data set related to the vertical class, and then extracting the feature from the training data, and training the machine learning classification The device (for example, the maximum entropy model and the support vector machine) classifies the vertical class demand query according to the extracted features to determine the correspondence between the query words and the vertical class, and saves the correspondence between the query words and the vertical class.

Take the restaurant as an example: suppose the user's current query is “quiet restaurant near Sanlitun”. Firstly, the basic lexical analysis of the query, named entity identification, etc., can be determined by analysis to determine the corresponding pattern of the query: [Location]_[Style]_[Restaurant]. Each category separates a collection of patterns. That is to say, for the query to be distributed, firstly, the query is analyzed by basic lexical analysis methods such as word segmentation and named entity recognition, and then the analysis result is matched with the pattern set of the vertical category, and if the matching is successful, the corresponding is distributed to the corresponding Vertical category.

The interaction sub-module 31423 can perform at least one round of interaction with the user in the vertical class to which the query term belongs, and obtain the query result required by the user. The information displayed to the user in each round of interaction includes: the query result corresponding to the query word. , as well, guide information.

In an embodiment of the present invention, as shown in FIG. 26, the interaction sub-module 31423 may include: a parsing unit 314231, a fourth obtaining unit 314232, a displaying unit 314233, and a fifth obtaining unit 314234.

The parsing unit 314231 can parse the query word into structured information that can be represented by the vertical knowledge system of the vertical class to which the query word belongs.

It should be noted that the user requirements are determined based on the query terms. That is to say, the user requirement representation system contains a large number of query words corresponding to the user's needs, and the semantics and structure of the query words can be obtained by analyzing the query words. Knowledge.

Table 1 Structure of the vertical knowledge system of the restaurant

The fourth obtaining unit 314232 may acquire related information according to the structured information, the vertical knowledge system, and the vertical resource library of the vertical class to which the query word belongs.

In an embodiment of the present invention, as shown in FIG. 27, the interaction sub-module 31423 may further include a component unit 314235.

In order to obtain the vertical class resource of the vertical class to which the query word belongs, the component unit 314235 can acquire the structured resource and the unstructured resource of the vertical class to which the query word belongs, and form the structured resource and the unstructured resource into the vertical class resource library. .

In an embodiment of the present invention, as shown in FIG. 28, the fourth obtaining unit 314232 may include an update subunit 3142321, a generating subunit 3142322, and a matching subunit 3142323.

The update subunit 3142321 may update the current state information of the user according to the structured information and the previous state information of the user.

The generating subunit 3142322 can generate a candidate action corresponding to the current state information according to the vertical class knowledge system and the vertical class resource library.

The matching sub-unit 3142323 may select, in the candidate action, a preset number of candidate actions that match the current state information according to the preset model, and select the selected candidate action as the related information.

The presentation unit 314233 can present the query results and the guidance information to the user.

The fifth obtaining unit 314234 may repeat the above process of acquiring related information according to the query word after the user inputs the query word again according to the guiding information, until the query result required by the user is obtained.

For another example, if the query word input by the current user is “a delicious Korean barbecue”, after receiving the query word input by the user, the vertical category corresponding to the query word may be a restaurant gourmet category, specifically, the query may be The word analysis is structured information that can be represented by the vertical knowledge system, and the query result and guidance information corresponding to the query word are obtained according to the structured information, the vertical knowledge system and the vertical class resource library of the vertical class. The obtained query result and the guiding information of the query word are returned to the user, wherein the user interface including the related information of the query word is as shown in FIG. 10, and at this time, the user may select one according to the guiding information, or may directly determine according to the requirement. This store. In addition, users can also view other boot information by clicking the next prompt button.

(3) Comparing the related dialogue system, the information query mode of this embodiment is complex for the vertical category resource structure Sexually doing specific processing, generating a state space based on the primitive structure of the vertical class, can satisfy the deep-level problems in the vertical class, and prompt the user to input the query word again through the guiding information to perform the next round of query, that is, The information query mode of the implementation can effectively guide the user to provide the correct question by displaying the guide information.

In an embodiment of the present invention, as shown in FIG. 29, the deep question and answer service module 31430 may include a first receiving submodule 31431, a first obtaining submodule 31432, and a generating submodule 31433.

Specifically, the deep question-and-answer service is a service for user input, based on in-depth semantic analysis and knowledge mining technology, to provide users with accurate Q&A results. When the user requirement information is a deep question and answer requirement, the first receiving submodule 31431 can receive the problem information, the first obtaining submodule 31432 obtains the corresponding question type according to the problem information, and then the generating submodule 31433 selects the corresponding question and answer mode according to the question type. And generating a corresponding question and answer result according to the selected answer generation mode and the problem information. Among them, the problem type may include an entity type, a viewpoint type, and a fragment type.

More specifically, when the question type is an entity type, as shown in FIG. 30, the generation sub-module 31433 may include a first generation unit 314331, an extension unit 314332, an extraction unit 314333, a first calculation unit 314334, and a first feedback unit 314335.

The first generating unit 314331 may generate entity class problem information according to the problem information, and the expanding unit 314332 expands the entity class problem information based on the digest and history presentation log captured by the search engine to generate a congenital entity problem information cluster. Among them, the information cluster of the same family entity problem corresponds to the candidate answer. Then, the extracting unit 314333 extracts candidate entities from the corresponding candidate answers in the same family entity problem information cluster, the first calculating unit 314334 calculates the confidence of the candidate entity, and the first feedback unit 314335 uses the candidate entity with the confidence greater than the preset reliability threshold as the question and answer. The results are fed back. For example, the question information is "Who is the wife of Andy Lau?" The candidate's answer is "In fact, as early as 1992, there were reports that Andy Lau and Zhu Liqian have been secretly registered in Canada...", in which the candidate entity is "Andy Lau" "Zhu Liqian" and "Canada". Then, based on the entity knowledge base and the question-and-answer semantic matching model, the confidence of each candidate entity is calculated. It can be calculated that the confidence level of the candidate entity “Zhu Liqian” is greater than the pre-set reliability threshold, and then “Zhu Liqian” can be determined as the question and answer result. In addition, the phrase "Zhu Liqian" for the first time in the candidate answer can also be used as the answer summary.

When the problem type is a viewpoint type, as shown in FIG. 31, the generating submodule 31433 may include a first obtaining unit 314336, a first segmentation unit 314337, a first aggregation unit 314338, a determination unit 314339, a selection unit 3143310, and a score feedback unit. 3143311.

The first obtaining unit 314336 may acquire candidate answers corresponding to the problem information, the first segmentation unit 314337 segments the candidate answers to generate a plurality of candidate answer phrases, and then the first aggregating unit 314338 aggregates the plurality of candidate answer phrases. To generate a view aggregation cluster. Specifically, the keyword in the candidate answer phrase can be extracted according to the IDF (anti-document frequency) score of the vocabulary in the short sentence, and the keyword including the negative word is generalized and a negative label is generated, and then the keyword is based on the negative label Representing a vector, calculating the vector angle and/or semantic similarity between each two keywords, The candidate answers whose vector angle is smaller than the preset angle or whose semantic similarity is greater than the preset threshold are aggregated to generate a view aggregation cluster.

After that, the judging unit 314339 can judge the opinion type of the opinion aggregation cluster. Among them, opinions may include non-classes, evaluation classes, suggestion classes, and the like. Specifically, the opinion type of the opinion aggregation cluster may be determined by a preset rule or based on a statistical model. The selection unit 3143310 then selects an answer point from the corresponding view aggregation cluster based on the opinion type. Among them, the rules for selecting the answer point of view may include, but are not limited to, selecting the information to cover the most comprehensive answer point of view, selecting the answer point with the lowest IDF*log (IDF) value, and selecting the answer point with the most occurrence in the article corresponding to the candidate answer. Among them, IDF is the anti-document frequency. After that, a summary corresponding to the answer point can be generated, and then the answer point can be scored, and the answer point with the score greater than the preset score threshold is fed back as a question and answer result. For example, the problem information is “Precautions for Pregnancy”. One of the candidate answers is “When you are pregnant, you should follow the principle of medical treatment, and the principle of war, that is, see a doctor regularly, rest in bed, and overcome your bad habits.” The candidate answers are divided into four short answer phrases: "When you are pregnant, you should follow the doctor's principles, and you should follow the principle of regular treatment," "that is, see the doctor regularly," "multiple bed rest," and "to defeat your bad habits." The duplicated content or approximate content in the candidate answer phrase can then be aggregated to generate a view aggregation cluster and the answer opinion can be selected. Thereafter, the score feedback unit 3143311 can score the answer viewpoint according to the information richness, the argument sufficient degree, the information redundancy degree, and the like, and feedback the answer viewpoint whose score is greater than the preset score threshold as the question and answer result. In addition, after selecting the answer point, the sentence in the source article can be obtained, and then the sentence is intercepted according to the predetermined length, thereby generating a summary corresponding to the answer point. The summary can then be sorted based on content richness and authoritative answers.

When the problem type is a segment type, as shown in FIG. 32, the generating submodule 31433 may include a second obtaining unit 1413132, a second segmentation unit 3143313, a scoring unit 3143314, a second generating unit 3114315, a first sorting unit 3143316, and a Two feedback units 3143317.

The second obtaining unit 3143312 may acquire candidate answers corresponding to the problem information, the second segmentation unit 3143313 splits the candidate answers to generate a plurality of candidate answer phrases, and then the scoring unit 3143314 scores the importance of the plurality of candidate answer sentences. To generate a short sentence importance feature corresponding to the candidate answer phrase, the second generating unit 3143315 generates an answer summary according to the short sentence importance feature, and then the first sorting unit 3143316 can according to the short sentence importance feature of the answer summary, the answer authority, The relevance of the problem information and the richness of the answers score the quality of the answer. The short sentence importance feature may include an aggregate feature, a relevance feature, a type feature, and a question answer matching feature. Among them, the aggregation feature is used to measure the repetition of short sentences, such as: word vector centroid feature, NGram (calculated probability of occurrence) feature, Lexrank (multi-text automatic summary) feature. Type characteristics are type characteristics of the problem, such as WHAT (what) type, WHY (why) type, HOW (how) type, and so on. The authoritativeness of the answer is the authority of the website from which the answer originated. After that, the user's behavior data can be obtained, and then the candidate answers are sorted according to the user's behavior data and the scoring result, and finally the second feedback unit 3143317 feeds back the sorting result as a question and answer result. Among them, the user's behavior data is It may include the user's click behavior on the question and answer results, the time spent on the question and answer results, and the historical behavior information of the user who jumped to other question and answer results through the current question and answer results.

In an embodiment of the present invention, as shown in FIG. 33, the information search service module 31440 may include a second receiving submodule 31441, a first search submodule 31442, and an analysis feedback submodule 31443. The analysis feedback sub-module 31443 may include an analysis unit 314431, a second sorting unit 314432, and a third generation unit 314433.

When the user requirement information is the information search requirement, the second receiving sub-module 31441 can receive the problem information, the first search sub-module 31442 performs a search according to the problem information to generate a plurality of candidate web pages, and then the analyzing unit 314431 of the analyzing feedback sub-module 31443 can Perform chapter analysis on the candidate web pages to generate corresponding candidate chapters. Specifically, the candidate web page may be subjected to chapter content extraction, chapter topic segmentation, and chapter relationship analysis to generate corresponding candidate chapters. Among them, the chapter content extraction mainly identifies the body part of the candidate webpage, and deletes the content irrelevant to the user demand information. The chapter topic is divided into the analysis of the topic structure of the chapter, which can divide the chapter into multiple sub-themes. Chapter relationship analysis is to analyze the relationship between multiple subtopics in a chapter, such as a side-by-side relationship. After generating the candidate chapter, the second sorting unit 314432 may sort the sentences in the candidate chapters. Among them, the scoring order is mainly based on the importance of the sentence in the candidate chapter and the correlation between the sentence and the user's demand information. After that, the third generating unit 314433 can acquire the user's demand scene information, generate a summary according to the demand scene information and the score sorting result, and finally feed the summary as a question and answer result. The scene information may include a mobile terminal scene and a computer scene. When the scene information is a mobile terminal scene, the sentence can be compressed and abbreviated, so that the generated summary is as concise as possible; when the scene information is a computer scene, the sentences can be spliced and merged, so that the generated abstract is clear and detailed.

As shown in FIG. 34, the analysis feedback sub-module 31443 can also include a second aggregation unit 314434.

Specifically, when the candidate chapter is generated, since the content in the candidate chapter has relevance to the user requirement information, there may be repeated or complementary content, and the second aggregating unit 314434 is required to aggregate the information of the plurality of candidate chapters.

The second decision module 31500 can receive the Q&A results returned by the at least one Q&A service module and make a decision on the Q&A results to determine the final Q&A result.

As shown in FIG. 35, the chat service subsystem 32000 can include a first receiving module 32100, a second distribution module 32200, a chat service module 32300, a second receiving module 32400, and a sorting module 32500.

The first receiving module 32100 is configured to receive input information input by a user.

The input information may be voice information or text information.

The second distribution module 32200 is configured to distribute the input information to the chat service module 32300.

The second receiving module 32400 is configured to receive candidate responses returned by the plurality of chat service modules 32300. Among them, the candidate reply has a corresponding confidence.

The sorting module 32500 is configured to sort the responses to be selected based on the confidence, and generate chat information according to the sorting result. And provide chat information to the user.

Specifically, the ranking module 32500 can acquire features of the user's input information, and sort the responses to be selected based on the characteristics and confidence of the input information. The characteristics of the input information may include classification features, literal features, topic features, and the like. The higher the confidence, the better the quality of the candidate reply is. The responses can be sorted according to the order of confidence, and finally the chat information that meets the user's needs is provided to the user.

As shown in FIG. 36, the chat service module 32300 can include a search-based chat module 32310, a rich knowledge chat module 32320, a portrait-based chat module 32330, and a crowdsourcing-based chat module 32340.

As shown in FIG. 37, the search-based chat module 32310 may include a word-cutting sub-module 32311, a query sub-module 32312, a filtering sub-module 32313, a classification sub-module 32314, and a first re-sorting sub-module 32315. The filtering submodule 32313 may include a second computing unit 323131, a filtering unit 323132, a retaining unit 323133, the classification submodule 32314 may include a third computing unit 323141, a classification unit 323142, and the first reordering submodule 32315 may include a third acquisition. Unit 323151 and reordering unit 323152.

Specifically, the word-cutting sub-module 32311 can perform a word-cutting on the input information to generate a plurality of chat phrases, and then the query sub-module 32312 can query the chat corpus according to the plurality of chat phrases to generate a plurality of chat corpus and multiple chats. The corpus corresponds to multiple chat corpora sentences. The chat corpus is pre-established, and the source of the chat can include, but is not limited to, “posting-replying” in the forum data such as posting bar, “blog-reply” in the microblog, and “question-answer” in the question and answer community.

After that, the filtering sub-module 32313 can filter a plurality of chat corpus sentences. Specifically, the second calculating unit 323131 may calculate the similarity between the input information and the plurality of chat corpus sentences. If the similarity is less than the first preset similarity threshold, the filtering unit 323132 may filter the corresponding chat corpus upper sentence; if the similarity is greater than or equal to the first preset similarity threshold, the retaining unit 323133 may compare the corresponding chat corpus The last sentence is reserved.

After filtering the upper corpus of the chat corpus, the categorization sub-module 31314 may classify the chat corpus corresponding to the sentence in the chat corpus after filtering. Specifically, the third calculating unit 323141 may calculate the similarity between the input information and the plurality of chat corpus sentences, and the classifying unit 323142 is based on the similarity based on the GBDT (Gradient Boost Decision Tree), SVM (Support Vector) Machine learning model, such as Support Vector Machine, classifies multiple chat corpora sentences. The similarity between the input information and the sentences of the plurality of chat corpora may be a literal similarity between the input information and the sentence of the chat corpus, or may be similar to the input information and the chat corpus based on the deep neural network training. Degree can also be the similarity between the input information and the chat corpus based on the machine translation model. It should be understood that the similarity between the input information and the plurality of chat corpus in the present embodiment and the machine learning model such as GBDT and SVM are well-known technologies, and are not described herein.

The first reordering sub-module 32315 can then reorder the chat corpus in the classified category and generate a candidate reply according to the sorting result. Specifically, the third obtaining unit 323151 may acquire the user's information according to the above information of the user chat. The chat attribute, the reordering unit 323152 reorders the chat corpus after the classification according to the chat attribute based on the learning-to-rank. The chat attribute may include a chat occasion such as a time and place, a chat fun, a chat style, and the like. Of course, the chat attribute is not limited to the above information from the user chat, but also can be obtained according to the user's long-term historical chat record. It should be understood that the learning ordering model in this embodiment is a well-known technology, and details are not described herein.

As shown in FIG. 38, the rich knowledge chat module 32320 can include a second search submodule 32321, an extraction submodule 32322, a rewrite submodule 32323, and a second reorder submodule 32324. Specifically, the second search sub-module 32321 may generate a search term according to the input information, and perform a search according to the search term to generate a plurality of search results, and then the extraction sub-module 32322 performs sentence extraction on the plurality of search results to obtain the search term. A set of candidate sentences of a sentence whose similarity is greater than a second preset similarity threshold. Thereafter, the rewrite sub-module 32323 can rewrite the sentences in the set of candidate sentences to generate a candidate reply. In addition, the second reordering sub-module 32324 may reorder the sentences in the candidate sentence set according to the chat attribute of the user. For example, if the input information is "I hope to have the opportunity to travel to Mount Fuji", the input information can be parsed and the corresponding search term "Mount Fuji, Tourism" can be generated, and then multiple search results can be obtained according to the search term, and the search words can be extracted and searched. A sentence with high similarity. Among them, some sentences may include obvious texts such as “journalists”, so these sentences need to be rewritten to make them more fluid, more like the sentences of natural language chat, and the final candidate for the response is “Mt. Fuji due to Due to the weather, there is only a certain period of summer in the year to climb the mountain. Compared with the traditional reply, "I also want to go to Mount Fuji, let's go together.", with certain knowledge and timeliness, can enable users to Get useful knowledge during the chat.

As shown in FIG. 39, the portrait-based chat module 32330 may include a second acquisition sub-module 32231, a first determination sub-module 32332, a transmission sub-module 32333, and a first update sub-module 32334.

In order to better realize anthropomorphization and provide personalized services for users, the human-machine chat system can set its own attributes, status, interests, etc., that is, the system portrait model. It is also possible to set the user's attributes, status, interests, etc., that is, the user portrait model. Of course, in the case of facing different users, the system portrait model used may be the same, or a system image model corresponding to each user may be set. Both the system portrait model and the user portrait model are based on the image knowledge map. The portrait knowledge map is a hierarchical knowledge system. For example, the "family member" node may include two child nodes "brothers" and "parents", and the "parent" child nodes include two child nodes "father" and "mother". Each node corresponds to a plurality of input information template clusters, such as "Who is your father", "Who is your father", and "What is your father's name" belongs to the same input information template cluster. Each input information template cluster corresponds to one or more candidate responses. The input information template cluster and the candidate reply may include variables, such as interest, hobbies, and preferences corresponding to the same attribute "INTEREST", and the attribute values of "INTEREST" may include climbing, music, reading, sports, and the like.

Specifically, the second obtaining submodule 32231 can obtain the chat context of the user, and the first determining submodule 32332 determines whether the collection condition is satisfied according to the chat context. If it is determined that the collection condition is met, the sending sub-module 32333 can provide the user to the user. Send a question. After that, the first update sub-module 32334 can receive the user's answer information according to the question and update the user portrait model according to the answer information. For example, when talking to a user about a movie-related topic, you can send the question "What movie do you like?" or the user asks the human-machine chat system "What do you like to eat?", the human-machine chat system can ask the user "You like to eat What? After the user answers, the user portrait model can be updated based on the user's answer information, which is more in line with the user's personalized needs.

In addition, as shown in FIG. 40, the portrait-based chat module 32330 may further include a third acquisition sub-module 32335, an extraction sub-module 32336, and a second update sub-module 32337.

Specifically, the third obtaining sub-module 32335 can acquire the chat content of the user, the extracting sub-module 32336 extracts the user portrait data according to the chat content, and then the second update sub-module 32337 updates the user portrait model according to the extracted user portrait data. For example, the user said in the chat process, "I like to climb the mountain and catch fishing when I am fine." I can extract the user portrait data "hobby climbing, hobby fishing" to update the user portrait model. At the same time, an appropriate answer can be extracted based on the user's portrait data, and the appropriate answer information can be returned to the user.

Specifically, the crowdsourcing-based chat module 32340 can determine whether the input information is suitable for crowdsourcing completion, for example, if the user's mood is low and needs comfort, etc., it is suitable for crowdsourcing completion. For example, if the user's input information contains personal information such as personal identification information, password, and telephone, it is not suitable for crowdsourcing.

If it is determined that the crowdsourcing is complete, the crowdsourcing based chat module 32340 can distribute the input information to the corresponding performer. Of course, the above information can also be sent to the performer together, and the performer can reply according to the above information and input information. The crowd-based chat module can then receive the reply information of the performer and make a quality judgment on the reply information. If the quality requirement is met, the reply message is used as a candidate reply. For example, if the reply message contains vulgar, reactionary, or pornographic content, the quality is not enough. Or if the executor replies for more than the predetermined duration, the replies of the executor will not be used, and the reply information can be saved to the chat corpus.

Additionally, as shown in FIG. 41, the chat service subsystem 32000 can also include an error correction module 32600.

The error correction module 32600 is configured to perform error correction and/or rewriting of the input information after receiving the input information input by the user, for correcting the typo in the input information, rewriting the irregular colloquial expression, and the like.

Additionally, as shown in FIG. 42, the chat service subsystem 32000 can also include an analysis module 32700.

The analyzing module 32700 is configured to perform domain analysis on the input information to obtain an area corresponding to the input information after receiving the input information input by the user, and then the second distribution module 32200 may distribute the input information to have the same or phase according to the domain corresponding to the input information. Approximate domain chat service module.

In addition, as shown in FIG. 43, the chat service subsystem 32000 may further include a fifth obtaining module 32800, the first judgment Module 32900, completion module 321000.

The fifth obtaining module 32800 is configured to acquire the above information that is chatted with the user after receiving the input information input by the user, and obtain the current topic information of the user according to the above information. Then, the first determining module 32900 can determine, according to the above information, whether the dependency of the input information and the above information is greater than a preset relationship threshold. When the dependency relationship is greater than the preset relationship threshold, the completion module 321000 can complete the input information according to the above information, thereby ensuring the smoothness of the human-machine chat. Specifically, complementing the input information may include referring to the digestion. For example, if the input information is "Has he got married?", he can replace "He" in the input information with "Andy Lau" according to the above information "Andy Lau". Completing the input information may also include omitting the completion. For example, the above information "Andy Lau's wife is called Zhu Liqian.", the input information is "I don't know.", then the input information can be completed as "I don't know Zhu Liqian."

In addition, as shown in FIG. 44, the chat service subsystem 32000 may further include a second determining module 32110, a seventh obtaining module 321200, a first generating module 321300, and a second generating module 321400.

The second determining module 321100 is configured to determine whether the input information belongs to chat information without actual content, such as “hehe”, “hoho”, and the like. If it is determined that the chat information belongs to the chat information without the actual content, the seventh obtaining module 321200 may acquire the current topic, that is, calculate the current topic according to the historical chat record based on the Topic Model. After acquiring the current topic, the first generation module 321300 may generate a guidance topic according to the current topic based on the topic chat map. Among them, the topic chat map is a directed graph with a topic as a node. For example, the node "leisure" can point to the node "watching movies" and the node "listening to the song", which means that the topic "leisure" can be guided to the topic "watching movies" or the topic "listening to the songs". The topic "watching movies" and the topic "listening songs" all have a certain guiding probability, and the guiding of the topic can be realized according to the guiding probability, thereby ensuring the diversity of guiding topics. Then, the second generation module 321400 can generate a candidate reply according to the guidance topic. Specifically, the template of the candidate reply may be generated based on the Natural Language Generation, and the guide topic may be filled into the template to generate a candidate reply; or the sentence including the guide topic may be selected as a candidate reply, thereby implementing the user. Actively chat topic guide.

As shown in FIG. 45, the guidance and recommendation service subsystem 33000 may include a determination module 33100, an acquisition module 33200, an eighth acquisition module 33300, and a feedback module 33400.

The determining module 33100 is configured to receive interaction information input by the user, and determine a current topic according to the interaction information.

Specifically, the determining module 33100 may first receive the interaction information input by the user, for example: “Is the dream space look good?”, and then perform the requirement identification and the correlation calculation on the interaction information, thereby determining that the current topic is “the dream space evaluation”.

The obtaining module 33200 is configured to obtain a plurality of candidate guiding topics related to the current topic based on the topic map.

Specifically, the obtaining module 33200 may acquire a plurality of candidate guiding topics related to the current topic based on the pre-established topic map. For example, if the current topic is "Privacy Dream Space Evaluation", you can obtain a number of guiding topics related to "Pirates of Dreams" based on the topic map, such as "Nolan's film", "Leonardo's movie", etc. And they are stolen The relationship between the evaluation of dream space.

The eighth obtaining module 33300 is configured to acquire user image data of the user.

The feedback module 33400 is configured to select a guiding topic from a plurality of candidate guiding topics related to the current topic according to the user portrait data, and feed back the guiding topic to the user.

Specifically, the feedback module 33400 may determine the user's intent information according to the user portrait data and the context information of the interaction information, and then select a guiding topic from the plurality of to-be-selected guiding topics related to the current topic according to the user's intention information, and provide feedback to the user. Guide the topic.

Additionally, as shown in FIG. 46, the guidance and recommendation service subsystem 33000 can also include an establishing module 33500.

The setup module 33500 is used to build a topic map.

A node in the topic map represents a topic or a requirement put forward by the user, and each node may include a reply of the corresponding topic and a resource satisfying the user's requirement, and the associated nodes may be associated by the edge to form a network. Topographical map.

Specifically, the establishing module 33500 can include a fourth obtaining submodule 33510 and a establishing submodule 33520.

The fourth obtaining submodule 33510 can acquire topic related data.

The fourth acquisition sub-module 33510 can obtain the topic association data in two cases.

When the network text data is unstructured data, the topic association data can be obtained based on entity extraction and syntax analysis. Among them, unstructured data may include news, forums, blogs, videos, and the like. For example: for web text data "most affected The eye-catching Nobel Prize for Literature is the winner of the New Ork. Of course, Murakami Haruki, who has repeatedly missed the Nobel Prize, is still the "most recent person who has won the Nobel Prize." The Chinese poet North Island has only made the Chinese people crazy. "Based on entity extraction technology, the entity information "Nobel Prize for Literature", "French Modiano", "Murao Haruki", "Chinese poet North Island" can be extracted, and based on syntactic analysis, the existence of the above entity information is known. Further, it can be analyzed that the Frenchman Modiano is a Nobel laureate in literature, and Haruki Murakami and the Chinese poet North Island did not receive the Nobel Prize in Literature.

When the network text data is semi-structured data, the topic-related data is obtained based on page structure analysis, tag extraction, and entity recognition. Among them, the semi-structured data may include Wikipedia, Baidu Encyclopedia and other encyclopedic data, or thematic data. For example, based on page structure analysis, tag extraction, and entity recognition, the “downstairs life” of “Djokovic” is included, including “family life” and “charity activities”.

When the network text data is structured data, the topic related data is obtained from the knowledge map. Among them, the structured data may include knowledge map data. For example, the film "The Pirates of the Dream" and the film "Interstellar Cross" are directed by "Christopher Nolan."

The second case: the user's behavior data can be obtained first, and then the topic association data is generated according to the behavior data. Among them, the behavior data may include search behavior data and browsing behavior data.

Certainly, the browsing behavior data of the user may also be obtained, and the corresponding browsing object is obtained according to the browsing behavior data, and the topic related data is generated according to the browsing object. For example, multiple news or videos that are clicked when a user browses a webpage can be associated to generate topic related data.

After acquiring the topic association data, the establishing sub-module 33520 may establish a topic map according to the topic association data by one or more of a RandomWalk algorithm, an association analysis algorithm, and a collaborative filtering algorithm.

For example, as shown in Fig. 17, q1, q2, q3, and q1', q2', q3', and q4' are topics, and d1, d2, d3, and d4 are resource data. As can be seen from FIG. 13, the resource data d1 and d2 are associated with the topic q1; the resource data d1, d2, d3 are associated with the topic q2; the resource data d4 is associated with the topic q3, and the topic with the associated relationship and the resource data are used. Solid lines are connected. Based on the RandomWalk algorithm, it is possible to iteratively calculate the relationship between the topic q1 and the resource data d3, which are connected by a dotted line. The topic q1' is a topic that the user issues after the resource data d1 or d4 is browsed according to the resource data d1 or d4, and the relationship between them has a sequential relationship. Similarly, the topic q2' is a topic issued based on the resource data d2, the topic q3' is a topic issued based on the resource data d2 or d3, and the topic q4' is a topic issued based on the resource data d3 or d4. Further, it can be inferred that the topic q1 and the topic q1' have an association relationship, the topic q1 and the topic q2' have an association relationship, etc., and finally the topic map as shown in Fig. 13 is established.

In addition, as shown in FIG. 47, the guidance and recommendation service subsystem 33000 may further include a parsing module 33600, a re-establishment reminder module 33700, and a reminder module 33800.

Specifically, the parsing module 33600 can parse the interaction information and obtain a key segment in the interaction information. The keyword segment may include one or more of time information, location information, and reminder events. The establishing reminding module 33700 can then establish reminding information according to the key segment. When the time information reaches the preset time, the reminding module 33800 can send the reminding information to the user. For example, if the user's interaction information is "Remind me to write a work plan at 6 o'clock tomorrow night", the time information "18:00 on August 8, 2015" can be parsed, and the reminder event is "write work plan". When this time is reached, the user can be alerted.

The human intelligence interaction system based on artificial intelligence in the embodiment of the invention comprises the following advantages: (1) realizing the transformation of the human-computer interaction system from instrumentalization to personification, and allowing users to interact in the process of intelligent interaction through chat, search and other services. Get a relaxing and enjoyable interactive experience, not just search and question and answer. (2) From keyword-based search to natural-based search, users can express their needs in a flexible and natural language. The multi-round interaction process is closer to the interactive experience between people. (3) To realize the evolution from user active search to all-weather companion service, based on the user's personalized model, you can provide users with recommendations and other services anytime, anywhere.

In the present invention, the terms "first" and "second" are used for descriptive purposes only, and are not to be construed as indicating or implying a relative importance or implicitly indicating the number of technical features indicated. Thus, features defining "first" or "second" may include at least one of the features, either explicitly or implicitly. In the description of the present invention, the meaning of "a plurality" is at least two, such as two, three, etc., unless specifically defined otherwise.

In the present specification, the schematic representation of the above terms is not necessarily directed to the same embodiment or example. Furthermore, the particular features or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. In addition, various embodiments or examples described in the specification, as well as features of various embodiments or examples, may be combined and combined.

Although the embodiments of the present invention have been shown and described, it is understood that the above-described embodiments are illustrative and are not to be construed as limiting the scope of the invention. The embodiments are subject to variations, modifications, substitutions and variations.

Claims

A human-computer interaction method based on artificial intelligence, characterized in that it comprises the following steps:

Receiving input information input by the user through the application terminal;

Obtaining the intent information of the user according to the input information of the user, and distributing the input information to at least one interaction service subsystem according to the intent information;

Receiving a return result returned by the at least one interactive service subsystem;

Generating a user return result according to the returned result according to a preset decision policy, and providing the user return result to the user.
The method of claim 1, wherein the providing the user return result to the user comprises:

Converting the user return result into a natural language and broadcasting to the user.
The method of claim 1 or 2, further comprising:

Receiving customized user task information;

The input information is distributed to at least one interactive service subsystem based on the customized task information.
The method according to any one of claims 1 to 3, wherein the application terminal comprises a PC end, a mobile terminal or an intelligent robot.
The method of any of claims 1-4, wherein if the user returns a result comprising executing an instruction, the method further comprises:

The execution instructions are sent to a corresponding execution subsystem and executed by the execution subsystem.
The method according to any one of claims 1-5, wherein the generating a user return result according to the returned result according to a preset decision policy specifically includes:

Obtaining a demand analysis feature of the input information;

Obtaining a confidence feature of the returned result returned by the interactive service subsystem, a context feature of the user's dialog interaction information, and a personalized model feature of the user;

Determining the returned result to determine the user return result according to the requirement analysis feature, the confidence feature of the returned result, the context feature of the user's dialog interaction information, and the personalized model feature of the user.
The method according to claim 6, wherein the demand analysis feature, the confidence feature of the returned result, the contextual feature of the user's dialog interaction information, and the personalized model feature of the user respectively correspond to Their respective decision weights.
The method of claim 7 further comprising:

Training according to the user's log based on the enhanced learning model, the demand analysis feature, the confidence feature of the returned result, the contextual feature of the user's dialog interaction information, and the decision weight of the user's personalized model feature .
The method of any of claims 1-8, further comprising:

Obtaining the above information of interaction with the user;

The input information is complemented according to the interaction of the above information.
The method of any of claims 1-9, wherein the interactive services subsystem comprises one or more of a requirements fulfillment service subsystem, a bootstrap and recommendation service subsystem, and a chat service subsystem.
The method of claim 10, further comprising:

The requirement satisfies the service subsystem to obtain problem information input by the user;

The requirement satisfaction service subsystem acquires user demand information of the user according to the problem information;

The demand satisfaction service subsystem distributes the problem information to a corresponding at least one question and answer service module according to the user demand information;

The demand satisfaction service subsystem receives the question and answer results returned by the at least one question and answer service module, and makes a decision on the question and answer result to determine a final question and answer result.
The method of claim 11, wherein the question and answer service module comprises an Aladdin service module, a vertical service module, a deep question and answer service module, and an information search service module.
The method of claim 12, further comprising:

The deep question and answer service module receives the problem information;

The deep question answering service module acquires a corresponding question type according to the problem information;

The deep question and answer service module selects a corresponding question and answer mode according to the type of the question, and generates a corresponding question and answer result according to the selected answer generation mode and the problem information.
The method according to claim 13, wherein when the question type is an entity type, the generating the corresponding question and answer result according to the selected answer generation mode and the problem information specifically includes:

Generating entity class problem information according to the problem information;

Generating the entity class problem information based on the summary and historical presentation logs captured by the search engine to generate the same family entity problem information cluster, wherein the same family entity problem information cluster respectively corresponds to the candidate answer;

Extracting candidate entities from the candidate answers of the same family entity problem information cluster respectively;

Calculating the confidence of the candidate entity;

The candidate entity with the confidence greater than the preset reliability threshold is fed back as a question and answer result.
The method according to claim 13, wherein when the question type is a viewpoint type, the generating a corresponding question and answer result according to the selected answer generation mode and the problem information specifically includes:

Obtaining candidate answers corresponding to the problem information;

Segmenting the candidate answers to generate a plurality of candidate answer phrases;

Aggregating the plurality of candidate answer phrases to generate a view aggregation cluster;

Determining the type of opinion of the view aggregation cluster;

Selecting an answer point from the view aggregation cluster according to the view type, and generating a summary corresponding to the answer view;

The answer point is scored, and the answer point whose score is greater than the preset score threshold is used as a question and answer result.
The method according to claim 15, wherein said aggregating said plurality of candidate answer phrases to generate a view aggregation cluster comprises:

Extracting keywords in the plurality of candidate answer phrases;

Calculating a vector angle and/or a semantic similarity between each of the two keywords;

The candidate answers whose vector angle is smaller than a preset angle or whose semantic similarity is greater than a preset threshold are aggregated to generate a view aggregation cluster.
The method according to claim 13, wherein when the question type is a segment type, the generating a corresponding question and answer result according to the selected answer generation mode and the problem information specifically includes:

Obtaining candidate answers corresponding to the problem information;

Segmenting the candidate answers to generate a plurality of candidate answer phrases;

Performing an importance score on the plurality of candidate answer short sentences to generate a short sentence importance feature corresponding to the candidate answer short sentence;

Generating an answer summary according to the short sentence importance feature;

The quality of the answer is scored according to the short sentence importance feature of the answer summary, and the candidate answers are sorted according to the score result;

The sort result is fed back as a question and answer result.
The method according to claim 17, wherein said scoring the quality of the answer according to the short sentence importance feature of the answer summary comprises:

The quality of the answer is scored according to the short sentence importance feature of the answer summary, the authority of the answer, the relevance of the question information, and the richness of the answer.
The method of claim 17, wherein the sorting the candidate answers according to the scoring result comprises:

Get user behavior data; and

The candidate answers are sorted according to the user's behavior data and the scoring result.
The method of any of claims 12-19, further comprising:

The information search service module receives the problem information;

The information search service module performs a search according to the problem information to generate a plurality of candidate web pages;

The information search service module performs chapter analysis on the candidate webpage to generate a corresponding abstract, and feeds the abstract as a question and answer result.
The method according to any one of claims 12 to 20, wherein the performing chapter analysis on the candidate webpage to generate a corresponding abstract comprises:

Performing chapter analysis on the candidate web page to generate a corresponding candidate chapter;

Sorting sentences in the candidate chapters; and

The summary is generated according to the score sorting result.
The method of claim 21, wherein the generating the summary according to the ranking result comprises:

Obtaining the user's demand scenario information;

Generating the digest according to the demand scenario information and the scoring ranking result.
The method of claim 21, further comprising:

Aggregate information for multiple candidate chapters.
The method of claim 10, further comprising:

The chat service subsystem receives input information input by a user;

The chat service subsystem distributes the input information to a chat service module;

The chat service subsystem receives candidate responses returned by the plurality of chat service modules, wherein the candidate responses have corresponding confidence levels;

The chat service subsystem sorts the to-be-selected responses based on the confidence, and generates chat information according to the sorting result, and provides the chat information to the user.
The method of claim 24, after the receiving the input information input by the user, further comprising:

The input information is error corrected and/or overwritten.
The method of claim 24, after the receiving the input information input by the user, further comprising:

Performing domain analysis on the input information to obtain an area corresponding to the input information, wherein the input information is distributed to a chat service module having the same or similar domain according to the domain corresponding to the input information.
The method of claim 24, after the receiving the input information input by the user, further comprising:

Obtaining the above information of chatting with the user;

Determining, according to the above information, whether the dependency of the input information and the above information is greater than a preset relationship threshold;

If it is greater than the preset relationship threshold, the input information is complemented according to the above information.
The method of claim 27, further comprising:

Obtaining current topic information of the user according to the above information.
The method of claim 24, wherein the chat service module comprises one or more of a search-based chat module, a rich knowledge chat module, a portrait-based chat module, and a crowdsourcing-based chat module.
The method of claim 29, further comprising:

The search-based chat module performs a word-cutting on the input information to generate a plurality of chat phrases;

The search-based chat module queries the chat corpus according to the plurality of chat phrases to generate a plurality of chat corpus sentences, and a plurality of chat corpus sentences corresponding to the plurality of chat corpora sentences;

The search-based chat module filters the plurality of chat corpus sentences;

The search-based chat module classifies the chat corpus corresponding to the upper sentence of the chat corpus after filtering;

The search-based chat module reorders the chat corpora sentence after classification, and generates the candidate reply according to the sort result.
The method of claim 30, wherein the filtering of the plurality of chat corpus by the search-based chat module comprises:

Calculating a similarity between the input information and the plurality of chat corpus sentences;

If the similarity is less than the first preset similarity threshold, filtering the corresponding chat corpus upper sentence;

If the similarity is greater than or equal to the first preset similarity threshold, the corresponding chat corpus is retained.
The method according to claim 30, wherein the classifying the chat corpus corresponding to the upper sentence of the chat corpus after filtering comprises:

Calculating a similarity between the input information and the plurality of chat corpus sentences;

Sorting the plurality of chat corpora sentences according to the similarity.
The method of claim 32, wherein the similarity between the input information and the plurality of chat corpus sentences comprises:

a literal similarity between the input information and the sentence of the chat corpus;

Or the similarity between the input information and the chat corpus based on deep neural network training;

Alternatively, the input information is similar to the training of the chat corpus based on the machine translation model.
The method according to claim 30, wherein said chat corpus after the classification is sentenced Row reordering specifically includes:

Obtaining a chat attribute of the user according to the above information of the user chat;

And re-sorting the chat corpus sentence after the classification according to the chat attribute.
The method of claim 29, further comprising:

The rich knowledge chat module generates a search word according to the input information, and searches according to the search word to generate a plurality of search results;

The rich knowledge chat module performs sentence extraction on the plurality of search results to obtain a candidate sentence set;

The rich knowledge chat module rewrites sentences in the candidate sentence set to generate the candidate reply.
The method of claim 35, wherein the similarity between the sentence in the set of candidate sentences and the search term is greater than a second predetermined similarity threshold.
The method of claim 35, further comprising:

The sentences in the candidate sentence set are reordered according to the chat attribute of the user.
The method of claim 29, further comprising:

The portrait-based chat module acquires a chat context of the user;

The portrait-based chat module determines whether the collection condition is satisfied according to the chat context;

Sending a question to the user if it is determined that the collection condition is satisfied;

Receiving the answer information of the user according to the question, and updating the user portrait model according to the answer information.
The method of claim 29, further comprising:

The portrait-based chat module acquires chat content of the user;

The portrait-based chat module extracts user portrait data according to the chat content;

The portrait-based chat module updates the user portrait model according to the extracted user portrait data.
The method of claim 29, further comprising:

The crowdsourcing-based chat module determines whether the input information is suitable for crowdsourcing completion;

If it is determined that the crowdsourcing is completed, the input information is distributed to the corresponding performer;

Receiving reply information of the performer, and performing quality judgment on the reply information;

If the quality requirement is met, the reply message is replied as the candidate.
The method of claim 24, further comprising:

Determining whether the input information belongs to chat information without actual content;

If it is determined that the chat information belongs to the content without actual content, the current topic is obtained;

Generating a boot topic based on the current topic;

The candidate reply is generated according to the guiding topic.
The method according to claim 24, wherein the sorting the candidate responses based on the confidence level specifically comprises:

Obtaining characteristics of the input information of the user;

The candidate responses are sorted based on the characteristics of the input information and the confidence.
The method of claim 10, further comprising:

The guidance and recommendation service subsystem receives interaction information input by the user, and determines a current topic according to the interaction information;

The guidance and recommendation service subsystem obtains a plurality of candidate guidance topics related to the current topic based on the topic map, wherein the topic map includes a plurality of topics and an association relationship between the topics;

The guidance and recommendation service subsystem acquires user image data of the user;

The guidance and recommendation service subsystem selects a guidance topic from the plurality of candidate guidance topics related to the current topic according to the user portrait data, and feeds back the guidance topic to the user.
The method according to claim 43, wherein before the guiding and recommending service subsystem obtains a plurality of candidate guidance topics related to the current topic based on the topic map, the method further comprises:

Establish the topic map.
The method of claim 44, wherein the establishing the topic map specifically comprises:

Get topic association data; and

The topic map is established based on the topic association data.
The method of claim 45, wherein the obtaining the topic association data comprises:

Obtain network text data;

When the network text data is unstructured data, acquiring the topic association data based on entity extraction and syntax analysis; or

When the network text data is semi-structured data, acquiring the topic-related data based on page structure analysis, label extraction, and entity identification; or

When the network text data is structured data, the topic related data is obtained from the knowledge map.
The method of claim 45, wherein the obtaining topic association data comprises:

Obtaining search behavior data of the user, and acquiring a corresponding search object according to the search behavior data, and generating the topic association data according to the search object; or

Acquiring the browsing behavior data of the user, and acquiring a corresponding browsing object according to the browsing behavior data, and generating the topic association data according to the browsing object.
The method of claim 45, wherein said establishing said said message based on said topic association data The map, specifically including:

The topic map is established according to the topic association data by one or more of a RandomWalk algorithm, an association analysis algorithm, and a collaborative filtering algorithm.
The method according to claim 43, wherein the determining the current topic according to the interaction information comprises:

A requirement identification and a correlation calculation are performed on the interaction information to determine the current topic.
The method according to claim 43, wherein said guidance and recommendation service subsystem selects a guidance topic from said plurality of candidate guidance topics related to said current topic based on said user portrait data, and The user feedbacks the guiding topic, which specifically includes:

Determining the intent information of the user according to the user portrait data and the context information of the interaction information;

Selecting a guiding topic from the plurality of candidate guiding topics related to the current topic according to the intention information of the user, and feeding back the guiding topic to the user.
The method of claim 43 further comprising:

Parsing the interaction information, and acquiring a key field in the interaction information, where the keyword segment includes one or more of time information, location information, and reminder event;

Establishing reminder information according to the keyword segment;

When the time information reaches a preset time, the reminder information is sent to the user.
The method of claim 1 further comprising:

If the user returns a result that satisfies the user requirement in the network resource, the input information of the user is recorded;

Monitoring, by the preset period, whether the user returns a result that satisfies the user requirement in the network resource;

The user return result is provided to the user when the user returns a result.
The method of claim 12, further comprising:

The vertical service module obtains a query word input by a user;

The vertical class service module determines a vertical class to which the query term belongs;

The vertical service module performs at least one round of interaction with the user in the vertical class to which the query term belongs, and obtains a query result required by the user, wherein, in each round of interaction, the information displayed to the user includes: corresponding query word The results of the query, as well, the boot information.
The method according to claim 53, wherein the query word is expressed in a natural language, and the at least one round of interaction with the user is performed in the vertical class to which the query word belongs, to obtain a query required by the user. The results include:

Parsing the query term into structured information that can be represented by a vertical knowledge system of the vertical class to which the query term belongs;

Obtaining related information according to the structured information, the vertical knowledge system, and the vertical resource library of the vertical class to which the query term belongs, the related information includes: a query result corresponding to the query word, and , guidance information;

Presenting the query result and the guidance information to a user;

After the user inputs the query word again according to the guiding information, the process of obtaining the related information according to the query word is repeated until the query result required by the user is obtained.
The method according to claim 54, wherein the obtaining, according to the structured information, the vertical knowledge system, and the vertical class resource library of the vertical class to which the query word belongs, acquire relevant information, including :

Updating the current state information of the user according to the structured information and the previous state information of the user;

Generating a candidate action corresponding to the current state information according to the vertical knowledge system and the vertical resource library;

Selecting a predetermined number of candidate actions that match the current state information to the candidate action according to the preset model, and selecting the selected candidate action as the related information.
The method of claim 55, further comprising:

The parameters of the preset model are updated according to the feedback of the user, so that different candidate actions are selected when the parameters are different.
The method of claim 55, further comprising:

The initialization status information of the user is obtained according to the user's preference or interaction history.
The method according to claim 55, wherein the candidate action comprises: an action that satisfies a user's needs, or an action of further clarifying a user's demand, or providing horizontal or vertical guidance information for the user's needs, wherein the user The requirement is determined according to the query term, the action that satisfies the user's requirement, or the action that further clarifies the user's demand is selected as the query result, and the horizontal or vertical guidance information is provided as the guide information after being selected.
The method of claim 54 further comprising:

Obtaining a structured resource and an unstructured resource of the vertical class to which the query word belongs, and the structured resource and the unstructured resource are configured into the vertical resource library, wherein the structured resource is from a plurality of The corresponding vertical website captures the full-light data resource obtained by integrating the data, and the unstructured resource supplements or expands the information according to the structured resource obtained by the user query word or the Internet text mining.
The method according to any one of claims 53-59, wherein the obtaining a query word input by a user comprises:

Get the query words that the user enters in text, voice, or image.
The method according to any one of claims 53-60, wherein the determining the vertical class to which the query term belongs comprises:

The vertical class to which the query term belongs is determined based on a machine learning manner or based on a mode analysis manner.
A human-computer interaction system based on artificial intelligence, characterized in that it comprises:

a first receiving subsystem, configured to receive input information input by a user through the application terminal;

a distribution subsystem, configured to acquire the intent information of the user according to the input information of the user, and distribute the input information to at least one interaction service subsystem according to the intent information;

a second receiving subsystem, configured to receive a return result returned by the at least one interactive service subsystem;

Generating a subsystem for generating a user return result according to the returned result according to a preset decision policy;

A subsystem is provided for providing the user return result to the user.
The system of claim 62 wherein said providing subsystem is specifically for:

Converting the user return result into a natural language and broadcasting to the user.
The system of claim 62 or 63, further comprising:

The first receiving subsystem is further configured to receive customized task information of the user;

The distribution subsystem is further configured to distribute the input information to the at least one interactive service subsystem according to the customized task information.
A system according to any one of claims 62 to 64, wherein the application terminal comprises a PC terminal, a mobile terminal or an intelligent robot.
The system of any of claims 62-65, wherein the system further comprises:

a sending subsystem, when the user returns a result including executing an instruction, sending the execution instruction to a corresponding execution subsystem;

The execution subsystem executes the execution instructions.
The system of any one of claims 62-66, wherein the generating subsystem comprises:

a first obtaining module, configured to acquire a demand analysis feature of the input information;

a second obtaining module, configured to acquire a confidence feature of the returned result returned by the interaction service subsystem, a context feature of the user's dialog interaction information, and a personalized model feature of the user;

a first decision module, configured to determine the return result according to the requirement analysis feature, a confidence feature of the returned result, a context feature of the user's dialog interaction information, and a personalized model feature of the user Determining that the user returns a result.
The system according to claim 67, wherein said demand analysis feature, a confidence feature of said return result, a contextual feature of said user's dialog interaction information, and said user's personalized model feature respectively correspond to Their respective decision weights.
The system of claim 68, wherein the generating subsystem further comprises:

a training module, configured to analyze, according to the user's log, the requirement analysis feature, the confidence feature of the returned result, the context feature of the user's dialog interaction information, and the personalized model feature of the user based on the enhanced learning model The decision weights are trained.
The system of any of claims 62-69, wherein the system further comprises:

Completing the subsystem for acquiring the above information of the interaction with the user, and completing the input information according to the information of the interaction.
The system of any of claims 62-70, wherein the interactive services subsystem comprises one or more of a demand fulfillment service subsystem, a bootstrap and recommendation service subsystem, and a chat service subsystem.
The system of claim 71, wherein the demand meets a service subsystem, specifically comprising:

a third obtaining module, configured to obtain problem information input by the user;

a fourth obtaining module, configured to acquire user demand information of the user according to the problem information;

a first distribution module, configured to distribute the problem information to the corresponding at least one question and answer service module according to the user requirement information;

And a second decision module, configured to receive a Q&A result returned by the at least one Q&A service module, and make a decision on the Q&A result to determine a final Q&A result.
The system according to claim 72, wherein said question answering service module comprises an Aladdin service module, a vertical service module, a deep question and answer service module, and an information search service module.
The system of claim 73, wherein the deep question and answer service module comprises:

a first receiving submodule, configured to receive the problem information;

a first obtaining submodule, configured to obtain a corresponding problem type according to the problem information;

And generating a sub-module, configured to select a corresponding question and answer mode according to the type of the problem, and generate a corresponding question and answer result according to the selected answer generation mode and the problem information.
The system of claim 74, wherein when the problem type is an entity type, the generating the sub-module specifically includes:

a first generating unit, configured to generate entity class problem information according to the problem information;

And an extension unit, configured to expand the entity class problem information to generate a cluster of the same family entity problem information according to the summary and historical presentation logs captured by the search engine, where the same family entity problem information cluster respectively corresponds to the candidate answer;

And an extracting unit, configured to extract a candidate entity from the corresponding candidate answers of the same family entity problem information cluster;

a first calculating unit, configured to calculate a confidence level of the candidate entity;

The first feedback unit is configured to feed back the candidate entity whose confidence is greater than the preset reliability threshold as a question and answer result.
The system of claim 74, wherein when the question type is a view type, the generating the sub-module specifically includes:

a first obtaining unit, configured to acquire a candidate answer corresponding to the problem information;

a first sub-unit for segmenting the candidate answers to generate a plurality of candidate answer phrases;

a first aggregating unit, configured to aggregate the plurality of candidate answer phrases to generate a view aggregation cluster;

a determining unit, configured to determine a view type of the view aggregation cluster;

a selecting unit, configured to select an answer point from the view aggregation cluster according to the view type, and generate a summary corresponding to the answer view;

A score feedback unit is configured to score the answer point of view and feed back an answer point whose score is greater than a preset score threshold as a question and answer result.
The system of claim 76, wherein the first aggregating unit is specifically configured to:

Extracting keywords in the plurality of candidate answer phrases, and calculating a vector angle and/or a semantic similarity between each of the two keywords, and the angle between the vectors is less than a preset angle or semantic similarity The candidate answers whose degree is greater than a preset threshold are aggregated to generate a view aggregation cluster.
The system of claim 74, wherein when the problem type is a segment type, the generating the sub-module specifically includes:

a second obtaining unit, configured to acquire a candidate answer corresponding to the problem information;

a second segmentation unit, configured to segment the candidate answer to generate a plurality of candidate answer phrases;

a scoring unit, configured to perform an importance score on the plurality of candidate answer short sentences to generate a short sentence importance feature corresponding to the candidate answer short sentence;

a second generating unit, configured to generate an answer summary according to the short sentence importance feature;

a first sorting unit, configured to score the quality of the answer according to the short sentence importance feature of the answer summary, and sort the candidate answers according to the score result;

The second feedback unit is configured to feed back the sorting result as a question and answer result.
The system of claim 78, wherein the scoring unit is specifically configured to:

The quality of the answer is scored according to the short sentence importance feature of the answer summary, the authority of the answer, the relevance of the question information, and the richness of the answer.
The system of claim 78, wherein the first sorting unit is specifically configured to:

Obtaining behavior data of the user, and sorting the candidate answers according to the behavior data of the user and the scoring result.
The system of any one of claims 73 to 80, wherein the information search service module comprises:

a second receiving submodule, configured to receive the problem information;

a first search submodule, configured to perform a search according to the problem information to generate a plurality of candidate webpages;

The analysis feedback sub-module is configured to perform chapter analysis on the candidate webpage to generate a corresponding abstract, and feed the abstract as a question and answer result.
The system of any one of claims 73-81, wherein the analyzing the feedback sub-module comprises:

An analyzing unit, configured to perform chapter analysis on the candidate webpage to generate a corresponding candidate chapter;

a second sorting unit, configured to score the sentences in the candidate chapter; and

And a third generating unit, configured to generate the digest according to the scoring ranking result.
The system of claim 82, wherein the third generating unit is specifically configured to:

Obtaining the user's demand scenario information, and generating the summary according to the demand scenario information and the score sorting result.
The system of claim 82, wherein the analyzing the feedback sub-module further comprises:

The second aggregating unit is configured to aggregate information of the plurality of candidate chapters.
The system of claim 71, wherein the chat service subsystem comprises:

a first receiving module, configured to receive input information input by a user;

a second distribution module, configured to distribute the input information to a chat service module;

a second receiving module, configured to receive a candidate reply returned by the multiple chat service modules, where the candidate reply has a corresponding confidence level;

a sorting module, configured to sort the to-be-selected responses based on the confidence, and generate chat information according to the sorting result, and provide the chat information to the user.
The system of claim 85, further comprising:

And an error correction module, configured to perform error correction and/or rewriting on the input information after receiving the input information input by the user.
The system of claim 85, further comprising:

An analysis module, configured to perform domain analysis on the input information to obtain an area corresponding to the input information after receiving the input information input by the user;

The second distribution module is configured to distribute the input information to a chat service module having the same or similar domain according to an area corresponding to the input information.
The system of claim 85, further comprising:

a fifth obtaining module, configured to acquire, after receiving the input information input by the user, the above information that is chat with the user;

a first determining module, configured to determine, according to the foregoing information, whether a dependency relationship between the input information and the foregoing information is greater than a preset relationship threshold;

And a completion module, configured to: when the dependency of the input information and the information above is greater than the preset relationship threshold, complete the input information according to the foregoing information.
The system of claim 88, further comprising:

And a sixth acquiring module, configured to acquire current topic information of the user according to the foregoing information.
The system of claim 85, wherein the chat service module comprises one or more of a search-based chat module, a rich knowledge chat module, a portrait-based chat module, and a crowdsourcing-based chat module.
The system of claim 90, wherein the search-based chat module comprises:

a word sub-module, configured to perform a word-cutting on the input information to generate a plurality of chat phrases;

a query sub-module, configured to query a chat corpus according to the plurality of chat phrases to generate a plurality of chat corpus sentences, and a plurality of chat corpus sentences corresponding to the plurality of chat corpora sentences;

a filtering submodule, configured to filter the plurality of chat corpus sentences;

a classification sub-module for classifying the chat corpus corresponding to the upper sentence of the chat corpus after filtering;

And a first reordering sub-module, configured to reorder the chat corpora sentence after the classification, and generate the candidate reply according to the sorting result.
The system of claim 91, wherein the filtering sub-module comprises:

a second calculating unit, configured to calculate a similarity between the input information and the plurality of chat corpus sentences;

a filtering unit, configured to: if the similarity is less than the first preset similarity threshold, filter the corresponding chat corpus;

And a saving unit, configured to reserve the corresponding chat corpus if the similarity is greater than or equal to the first preset similarity threshold.
The system of claim 91, wherein the classification sub-module comprises:

a third calculating unit, configured to calculate a similarity between the input information and the plurality of chat corpus sentences;

a classifying unit, configured to classify the plurality of chat corpus sentences according to the similarity.
The system of claim 93, wherein the similarity between the input information and the plurality of chat corpus sentences comprises:

a literal similarity between the input information and the sentence of the chat corpus;

Or the similarity between the input information and the chat corpus based on deep neural network training;

Alternatively, the input information is similar to the training of the chat corpus based on the machine translation model.
The system of claim 91, wherein the first reordering sub-module comprises:

a third acquiring unit, configured to acquire, according to the foregoing information of the user chat, a chat attribute of the user;

a reordering unit, configured to reorder the chat corpus sentence after the classification according to the chat attribute.
The system of claim 90, wherein the rich knowledge chat module comprises:

a second search submodule, configured to generate a search term according to the input information, and perform a search according to the search term to generate a plurality of search results;

An extraction submodule, configured to perform sentence extraction on the plurality of search results to obtain a candidate sentence set;

Rewriting a sub-module for rewriting a sentence in the candidate sentence set to generate the candidate reply.
The system of claim 96, wherein the similarity between the sentence in the set of candidate sentences and the search term is greater than a second predetermined similarity threshold.
The system of claim 96, further comprising:

And a second reordering submodule, configured to reorder the sentences in the candidate sentence set according to the chat attribute of the user.
The system of claim 90, wherein the image-based chat module comprises:

a second obtaining submodule, configured to acquire a chat context of the user;

a first determining sub-module, configured to determine, according to the chat context, whether a collection condition is met;

a sending submodule, configured to send a question to the user if it is determined that the collection condition is met;

The first update submodule is configured to receive the answer information of the user according to the question, and update the user portrait model according to the answer information.
The system of claim 90, wherein the image-based chat module comprises:

a third obtaining submodule, configured to acquire chat content of the user;

An extraction submodule, configured to extract user portrait data according to the chat content;

And a second update submodule, configured to update the user portrait model according to the extracted user portrait data.
The system of claim 90, wherein the crowdsourcing-based chat module is specifically configured to:

Determining whether the input information is suitable for crowdsourcing completion, if it is determined that the crowdsourcing is completed, distributing the input information to a corresponding performer, and receiving the reply information of the performer, and performing quality judgment on the reply information If the quality requirement is met, the reply message is replied as the candidate.
The system of claim 85, wherein the chat service subsystem further comprises:

a second determining module, configured to determine whether the input information belongs to chat information without actual content;

a seventh obtaining module, configured to acquire a current topic if it is determined to belong to chat information without actual content;

a first generating module, configured to generate a guiding topic according to the current topic;

And a second generating module, configured to generate the candidate reply according to the guiding topic.
The system of claim 85, wherein the ranking module is specifically configured to:

Acquiring characteristics of the input information of the user, and sorting the to-be-selected responses based on characteristics of the input information and the confidence.
The system of claim 71, wherein the guiding and recommending service subsystem comprises:

a determining module, configured to receive interaction information input by the user, and determine a current topic according to the interaction information;

Obtaining a module, configured to obtain, according to a topic map, a plurality of candidate guidance topics related to the current topic, where the topic map includes a plurality of topics and an association relationship between the topics;

An eighth obtaining module, configured to acquire user image data of the user;

And a feedback module, configured to select a guiding topic from the plurality of candidate guiding topics related to the current topic according to the user portrait data, and feed back the guiding topic to the user.
The system of claim 104, further comprising:

And a establishing module, configured to establish the topic map before the guiding and recommendation service subsystem obtains a plurality of candidate guiding topics related to the current topic based on the topic map.
The system of claim 105, wherein the establishing module comprises:

a fourth obtaining submodule for acquiring topic related data;

Establishing a sub-module for establishing the topic map according to the topic association data.
The system of claim 106, wherein the fourth obtaining sub-module is specifically configured to:

Obtain network text data;

When the network text data is unstructured data, acquiring the topic association data based on entity extraction and syntax analysis; or

When the network text data is semi-structured data, acquiring the topic-related data based on page structure analysis, label extraction, and entity identification; or

When the network text data is structured data, the topic related data is obtained from the knowledge map.
The system of claim 106, wherein the fourth obtaining sub-module is specifically configured to:

Obtaining search behavior data of the user, and acquiring a corresponding search object according to the search behavior data, and generating the topic association data according to the search object; or

Acquiring the browsing behavior data of the user, and acquiring a corresponding browsing object according to the browsing behavior data, and generating the topic association data according to the browsing object.
The system of claim 106, wherein the establishing submodule is specifically configured to:

The topic map is established according to the topic association data by one or more of a RandomWalk algorithm, an association analysis algorithm, and a collaborative filtering algorithm.
The system of claim 104, wherein the determining module is specifically configured to:

A requirement identification and a correlation calculation are performed on the interaction information to determine the current topic.
The system of claim 104, wherein the feedback module is specifically configured to:

Determining the intent information of the user according to the user portrait data and the context information of the interaction information, and selecting a guiding topic from the plurality of candidate guiding topics related to the current topic according to the intent information of the user And feeding back the guiding topic to the user.
The system of claim 104, further comprising:

a parsing module, configured to parse the interaction information, and obtain a key segment in the interaction information, where the keyword segment includes one or more of time information, location information, and reminder events;

Establishing a reminding module, configured to establish reminding information according to the key segment;

The reminding module is configured to send the reminding information to the user when the time information reaches a preset time.
The system of claim 62, further comprising:

a recording subsystem, configured to record input information of the user if the user returns a result that satisfies the user requirement in the network resource;

a monitoring subsystem, configured to monitor, by using a preset period, whether the user returns a result that meets the user requirement in the network resource;

Providing a subsystem, further configured to provide the user return result to the user when the user returns a result.
The system of claim 73, wherein the service module comprises:

a fifth obtaining submodule, configured to obtain a query word input by the user;

Determining a sub-module for determining a vertical class to which the query term belongs;

The interaction sub-module is configured to perform at least one round of interaction with the user in the vertical class to which the query term belongs, to obtain a query result required by the user, wherein, in each round of interaction, the information displayed to the user includes: corresponding query word The results of the query, as well, the boot information.
The system according to claim 114, wherein the query term is expressed in a natural language, and the interaction sub-module specifically includes:

a parsing unit, configured to parse the query word into structured information that can be represented by a vertical knowledge system of the vertical class to which the query term belongs;

a fourth obtaining unit, configured to acquire related information according to the structured information, the vertical knowledge system, and a vertical resource library of the vertical class to which the query word belongs, where the related information includes: corresponding to the The query result of the query word, and the guide information;

a display unit, configured to display the query result and the guiding information to a user;

The fifth obtaining unit is configured to repeat the foregoing process of acquiring related information according to the query word after the user inputs the query word again according to the guiding information, until the query result required by the user is obtained.
The system of claim 115, wherein the fourth obtaining unit comprises:

Updating a subunit, configured to update current status information of the user according to the structured information and a previous status information of the user;

Generating a subunit, configured to generate the current state information according to the vertical knowledge system and the vertical resource library Corresponding candidate action;

And a matching sub-unit, configured to select, in the candidate action, a preset number of candidate actions that match the current state information according to the preset model, and select the selected candidate action as related information.
The system of claim 116, further comprising:

The parameter unit is updated to update the parameters of the preset model according to the feedback of the user, so as to select different candidate actions when the parameters are different.
The system of claim 116, further comprising:

The obtaining subunit is configured to obtain initialization state information of the user according to the preference of the user or the interaction history.
The system according to claim 116, wherein said candidate actions comprise: actions that satisfy user requirements, or actions that further clarify user requirements, or provide horizontal or vertical guidance information for user needs, wherein the user The requirement is determined according to the query term, the action that satisfies the user's requirement, or the action that further clarifies the user's demand is selected as the query result, and the horizontal or vertical guidance information is provided as the guide information after being selected.
The system of claim 115, further comprising:

a component unit, configured to acquire a structured resource and an unstructured resource of a vertical class to which the query term belongs, and the structured resource and the unstructured resource are configured into the vertical resource library, wherein the structure The resource is a full-light data resource obtained by fetching integrated data from a plurality of corresponding web sites, and the unstructured resource is supplemented or expanded according to a structured resource obtained by user query words or Internet text mining.
The system according to any one of claims 114 to 120, wherein the fifth obtaining submodule is specifically configured to:

Get the query words that the user enters in text, voice, or image.
The system according to any one of claims 114-121, wherein the determining sub-module is specifically configured to:

The vertical class to which the query term belongs is determined based on a machine learning manner or based on a mode analysis manner.
An apparatus, comprising:

One or more processors;

Memory

One or more programs, the one or more programs being stored in the memory, when executed by the one or more processors, performing the artificial intelligence based method of any one of claims 1-61 Human-computer interaction method.
A non-volatile computer storage medium, characterized in that the computer storage medium stores one or more programs, when the one or more programs are executed by a device, causing the device to perform The artificial intelligence based human-computer interaction method according to any one of claims 1-61.