WO2021027421A1 - 用于人机交互的人机多轮对话方法及系统、智能设备 - Google Patents

用于人机交互的人机多轮对话方法及系统、智能设备 Download PDF

Info

Publication number
WO2021027421A1
WO2021027421A1 PCT/CN2020/099423 CN2020099423W WO2021027421A1 WO 2021027421 A1 WO2021027421 A1 WO 2021027421A1 CN 2020099423 W CN2020099423 W CN 2020099423W WO 2021027421 A1 WO2021027421 A1 WO 2021027421A1
Authority
WO
WIPO (PCT)
Prior art keywords
node
knowledge graph
user
information
support
Prior art date
Application number
PCT/CN2020/099423
Other languages
English (en)
French (fr)
Inventor
黄姿荣
贾巨涛
吴伟
李禹慧
戴林
Original Assignee
珠海格力电器股份有限公司
珠海联云科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 珠海格力电器股份有限公司, 珠海联云科技有限公司 filed Critical 珠海格力电器股份有限公司
Priority to US17/623,665 priority Critical patent/US20220253710A1/en
Publication of WO2021027421A1 publication Critical patent/WO2021027421A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G06N5/022—Knowledge engineering; Knowledge acquisition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/332—Query formulation
    • G06F16/3329—Natural language query formulation or dialogue systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3344—Query execution using natural language analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367—Ontology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models

Definitions

  • the present disclosure relates to the field of human-computer interaction technology, and in particular to a human-computer multi-round dialogue method and system for human-computer interaction, and smart devices.
  • voice interaction has gradually become the standard configuration of smart homes and smart home appliances.
  • current voice interaction achieves basic task-based interaction and satisfies command control, there is still an unintelligent state in the process of multi-dimensional dialogue understanding and intelligent response, resulting in poor voice interaction experience.
  • Dialogue is based on a principle of cooperative interaction.
  • the participants in the dialogue create and agree with the expression of the language in a rich and subtle context.
  • current human-computer interaction can realize user intent recognition through multiple rounds of dialogue and contextual correlation, multiple rounds of dialogue will cause poor interaction experience. Therefore, for voice interaction between machines and humans, it is necessary to establish a cognitive background for the machines to understand the meaning that users want to express, thereby reducing multiple rounds of human-computer interaction and achieving the effect of intelligent interaction.
  • the technical problem to be solved by the present disclosure is to overcome the problem that the current smart device requires multiple rounds of dialogue to understand the user's intention, resulting in poor voice interaction experience.
  • the present disclosure provides a man-machine multi-round dialogue method for man-machine interaction.
  • the method includes the following steps:
  • Step S1 Establish a knowledge graph of user dialogue behavior information
  • Step S2 According to the information currently input by the user, the node corresponding to the information currently input and the child nodes of the node are determined in the knowledge graph;
  • Step S3 Calculate the support degree of the child node relative to the node based on the number of times the node is queried and the number of times the child node and the node are queried simultaneously in the historical query record of the knowledge graph;
  • Step S4 Determine whether to output the semantic information of the child node by judging the magnitude relationship between the support degree and a preset support degree threshold.
  • the support degree is obtained by calculating the ratio of the number of times the child node and the node are simultaneously queried to the number of times the node is queried.
  • step S4 in step S4,
  • step S4 in step S4,
  • the method further includes:
  • Step S5 According to the information input by the user in real time, the knowledge graph is updated in real time.
  • step S5 includes:
  • Step S51 Collect the information input by the user in real time and form structured data
  • Step S52 Perform entity recognition and naming on the structured data, and form a new node corresponding to the information input by the user in real time in the knowledge graph;
  • Step S53 Extract the relationship data between the new node and the original node in the knowledge graph from a preset text corpus
  • Step S54 Perform similarity calculation and matching between the new node and the original node, and combine the relationship data to form a structure of node-relation-node and node-attribute-attribute value.
  • a man-machine multi-round dialogue system for man-machine interaction including:
  • Knowledge graph establishment module which is configured to establish a knowledge graph of user dialogue behavior information
  • An index module which is configured to: according to the information currently input by the user, determine in the knowledge graph the node corresponding to the currently input information and the child nodes of the node;
  • a calculation module which calculates the support degree of the child node relative to the node according to the number of times the node is queried and the number of times the child node and the node are simultaneously queried in the historical query record of the knowledge graph;
  • the judging module is configured to determine whether to output the semantic information of the child node by judging the magnitude relationship between the support degree and a preset support degree threshold.
  • calculating the support degree of the child node relative to the node based on the number of times the node is queried and the number of times the child node and the node are queried simultaneously in the historical query record of the knowledge graph, including :
  • the support degree is obtained by calculating the ratio of the number of times the child node and the node are simultaneously queried to the number of times the node is queried.
  • the system further includes:
  • the knowledge graph update module is configured to update the knowledge graph in real time according to the information input by the user in real time.
  • a smart device including:
  • the memory has executable code stored thereon, and when the executable code is executed by the processor, the processor executes the above-mentioned man-machine multi-round dialogue method for man-machine interaction.
  • one or more embodiments of the above solutions may have the following advantages or beneficial effects:
  • the machine does not need to use multiple rounds of voice dialogue to understand user intentions, and the user's voice interaction experience is improved.
  • Fig. 1 is a flowchart of a man-machine multi-round dialogue method for man-machine interaction according to the first embodiment of the present disclosure.
  • Fig. 2 schematically shows a knowledge graph of user dialogue behavior information established in step S1 according to an embodiment of the present disclosure.
  • Fig. 3 is a flowchart of step S1 of a man-machine multi-round dialogue method for man-machine interaction according to an embodiment of the present disclosure.
  • Fig. 4 is a flowchart of a man-machine multi-round dialogue method for man-machine interaction according to the second embodiment of the present disclosure.
  • Fig. 5 is a flowchart of step S5 of the man-machine multi-round dialogue method for man-machine interaction according to the second embodiment of the present disclosure.
  • Fig. 6 schematically shows a man-machine multi-round dialogue system for man-machine interaction according to the third embodiment of the present disclosure.
  • embodiments of the present disclosure provide a human-machine multi-round dialogue method and system for human-computer interaction, and smart devices .
  • Fig. 1 is a flowchart of a man-machine multi-round dialogue method for man-machine interaction according to the first embodiment of the present disclosure. As shown in Figure 1, the method includes the following steps:
  • Step S1 Establish a knowledge graph of user dialogue behavior information
  • Step S2 According to the information currently input by the user, determine in the knowledge graph the node corresponding to the information currently input and the child nodes of the node;
  • Step S3 Calculate the support degree of the child node relative to the node based on the number of times the node is queried and the number of times the child node and the node are queried simultaneously in the historical query record of the knowledge graph;
  • Step S4 Determine whether to output the semantic information of the child node by judging the magnitude relationship between the support degree and a preset support degree threshold.
  • step S1 a knowledge graph of user dialogue behavior information is established.
  • the knowledge graph is specifically for a certain user, that is, for each user, a knowledge graph specifically for the user is established.
  • a knowledge graph specifically for the user is established.
  • multiple users share a knowledge graph, and this disclosure is not limited to this.
  • Fig. 2 schematically shows a knowledge graph of user dialogue behavior information established in step S1 according to an embodiment of the present disclosure.
  • a knowledge graph for the dialogue behavior information of the user is established.
  • the established knowledge graph includes nodes, sub-nodes, attributes and attribute values, the relationship between nodes and nodes, and the relationship between node-attribute-attribute values.
  • nodes include location nodes, information nodes, and usage scenarios nodes.
  • information nodes are further divided into at least one or more of time nodes, interest nodes, gender nodes, age nodes, and social feature nodes.
  • its child nodes include weather nodes, music nodes, and story nodes.
  • the child nodes of the weather node include traffic state nodes and protective facilities nodes.
  • the frequency is used as an attribute of each node and each child node, and correspondingly, the frequency value (specific value, for example, 5 times) is used as An attribute value for each node and each child node.
  • the frequency value refers to the number of times the node or the child node is queried in the historical query record of the knowledge graph.
  • Fig. 3 is a flowchart of step S1 of a man-machine multi-round dialogue method for man-machine interaction according to an embodiment of the present disclosure.
  • the knowledge graph is based on the basic graph template framework and is established by the user through multiple previous trainings. As shown in Figure 3, the specific process is as follows:
  • Step S11 Establish a basic graph template framework including basic nodes and relationships
  • Step S12 Collect the information input by the user during training, and form structured data
  • Step S13 Perform entity recognition and naming on the structured data, and form a new node corresponding to the information input by the user in the knowledge graph;
  • Step S14 Extract the relationship data between the new node and the original node in the knowledge graph from a preset text corpus
  • Step S15 Perform similarity calculation and matching between the new node and the original node, and combine the relationship data to form a structure of node-relation-node and node-attribute-attribute value.
  • step S11 the developer establishes a basic graph template framework including basic nodes and relationships according to the general needs of public users. For example, current users are more concerned with information such as weather, traffic, and news. At this time, in some implementations, weather, traffic, and news can be formed as basic nodes in the basic map template framework. For another example, in some embodiments, a general user first asks about the weather and then about the traffic. At this time, the traffic state node can be regarded as a child node of the weather node.
  • frequency is used as an attribute of the nodes and sub-nodes in the basic graph template framework, and correspondingly, the frequency value is taken as an attribute value of the nodes and sub-nodes in the basic graph template framework.
  • the basic atlas template framework established in step S11 may not be able to specifically satisfy a certain user. Therefore, it is necessary to train the basic atlas template framework to obtain a specific Knowledge graph of users.
  • step S12 the information input by the user during training is collected, and a top-down construction method is used to form structured data.
  • the formed structured data is data logically expressed by a two-dimensional table structure, with strict format and length specifications.
  • the semantic information input by the user is "Do you need to open an umbrella?"
  • the structured data formed is "open umbrella”.
  • step S13 entity recognition.
  • entity recognition and naming are performed on the structured data, and a new node corresponding to the information input by the user is formed in the knowledge graph.
  • the server log is used to search, and the semantic feature corresponding to the structured data is searched out to identify the structured data and name it. For example, for the structured data "Umbrella”, according to the results of the server log search, "Umbrella” is identified as an action, and it is named “Action, Umbrella”. Then, a new node "action, umbrella” is generated in the knowledge graph.
  • step S14 relation extraction.
  • the relationship data between the new node and the original node in the knowledge graph is extracted from the preset text corpus.
  • the original node refers to the existing node in the knowledge graph before the new node "action, umbrella" is formed.
  • a method of semantic rule pattern matching is used for relation extraction.
  • step S15 knowledge fusion.
  • the purpose of knowledge fusion in this step is to eliminate ambiguity so as to facilitate the fusion of nodes and nodes.
  • similarity calculations are performed on the new nodes and the original nodes in sequence, and this process includes: object alignment, entity matching, and entity synonymous processing.
  • the attribute value (frequency value) of the attribute frequency of the "weather” node is increased by 1.
  • the new node is retained in the knowledge graph, and the attribute value (frequency value) of the attribute frequency of the new node is increased by 1.
  • the relationship data obtained in step S14 is combined to form a structure of node-relation-node and node-attribute-attribute value.
  • these two structures are referred to as entity-relation-entity and entity-attribute-attribute value structures.
  • a knowledge graph specifically for the user's dialogue behavior information is obtained.
  • the user will be supplemented with related intentions. For example, if the user says “How is the weather today?", the machine will ask “Which city do you want to check the weather?” to supplement the user's intention and send the user dialogue information
  • the intention in is updated to the node of the knowledge graph or the attribute value of the node to establish a knowledge graph of the user’s dialogue behavior information.
  • step S2 to step S4 are executed next.
  • step S2 according to the information currently input by the user, the node corresponding to the information currently input and the child nodes of the node are determined in the knowledge graph.
  • the weather node As shown in Figure 2, taking the weather node as an example, if the user asks "How is the weather today?", through semantic analysis the user wants to query "weather", then the weather node and the child nodes of the weather node are determined in the knowledge graph , And select the attributes (such as frequency) under the weather node, traverse the sub-nodes such as protective measures, traffic status, and clothing matching to form a set, and read the attribute values (frequency values) of the nodes and sub-nodes as items.
  • the attributes such as frequency
  • step S3 is executed: according to the number of times the node is queried and the number of times the child nodes and the node are simultaneously queried in the historical query records of the knowledge graph, the support degree of the child nodes relative to the node is calculated. Specifically, see Table 1.
  • each user request is recorded as a transaction, and each transaction records the nodes involved, where 1 is the question and answer involved in the transaction, and 0 is the question and answer not involved in the transaction.
  • its child nodes include: protective measures B, traffic status C, and clothes matching D.
  • the weather node A was queried three times in total, that is, the frequency value is 3, and the dressing match D was queried only once when the request number was 3, that is, the dressing match D and the weather The frequency value of node A being queried simultaneously is 1.
  • the support degree can be obtained by calculating the ratio of the number of times the child nodes and nodes are simultaneously queried to the number of times the node is queried. which is,
  • the support for dressing match D is 33%. It should be noted that the transaction that A and D appear at the same time is that the dressing collocation D and the weather node A are queried at the same time (when the weather node A is queried, the dressing collocation D is also queried, as shown in the request number 3 in Table 1. In the case of), the total transaction is the frequency value of the weather node A being queried.
  • the support degree of protective measure B relative to weather node A (hereinafter referred to as the support degree of protective measure B), and the support degree of traffic state C relative to weather node A (hereinafter referred to as the support of traffic state C) are sequentially calculated. degree). According to the content shown in Table 1, it is calculated that the support degree of protective measure B is 66%, and the support degree of traffic state C is 66%.
  • step S4 is executed next: the size relationship between the support degree and the preset support degree threshold is determined to determine the information to be output.
  • step S4 it is judged whether the support degree is greater than the preset support threshold value: if the support degree is greater than the preset support threshold value, the semantic information of the child node is output while the semantic information of the node is output; otherwise only the semantic information of the node is output .
  • the preset support threshold is set to 50%.
  • the size of the preset support threshold can be flexibly set according to actual conditions, and the present disclosure is not limited to this.
  • the semantic information of dressing collocation D is not output; while the support for protective measure B and traffic state C are both greater than the preset support With a threshold of 50%, while outputting the semantic information of the weather node, it also outputs the semantic information of child node protection measures B and child node traffic status C.
  • the machine will output "The weather in Guangzhou is sunny and the ultraviolet rays are strong. Please use sunscreen appropriately, South China Express Line processing congestion status" these information.
  • the support of all child nodes is not greater than the preset support threshold of 50%, only the semantic information of the weather node will be output, that is, the machine will only output "Guangzhou Weather Sunny".
  • step S4 is performed in the following manner: determining whether the support degree is greater than the preset support degree threshold: if the support degree is greater than the preset support threshold value, the formation includes all supports The frequent set of child nodes whose degree is greater than the preset support threshold, and while outputting the semantic information of the node, the semantic information of each child node in the frequent set is output in order from the largest support to the smallest; otherwise, only the node's semantic information is output Semantic information.
  • the frequent set includes protective measures B and traffic status C, and then While outputting the semantic information of the weather node, the semantic information of frequently concentrated protective measures B and traffic status C is output in order of support degree from large to small.
  • the output voice information takes too long, which affects the user's interactive experience.
  • the semantic information of the output node only the semantic information of the top three child nodes with the highest frequently concentrated support is output.
  • the machine does not need to use multiple rounds of voice dialogue to understand the user's intentions, and the user's voice interaction experience is improved.
  • Fig. 4 is a flowchart of a man-machine multi-round dialogue method for man-machine interaction according to the second embodiment of the present disclosure.
  • the difference from the first embodiment is that the man-machine multi-round dialogue method for man-machine interaction provided in the second embodiment can update the knowledge graph in real time.
  • the method includes the following steps:
  • Step S1 Establish a knowledge graph of user dialogue behavior information
  • Step S2 According to the information currently input by the user, the node corresponding to the information currently input and the child nodes of the node are determined in the knowledge graph;
  • Step S3 Calculate the support degree of the child node relative to the node according to the number of times the node is queried and the number of times the child node and the node are queried simultaneously in the historical query record of the knowledge graph;
  • Step S4 Determine whether to output the semantic information of the child node by judging the magnitude relationship between the support degree and a preset support degree threshold;
  • Step S5 According to the information input by the user in real time, the knowledge graph is updated in real time.
  • step S1 a knowledge graph of user dialogue behavior information is established.
  • the knowledge graph is specifically for a certain user, that is, for each user, a knowledge graph specifically for the user is established.
  • Fig. 2 schematically shows a knowledge graph of user dialogue behavior information established in step S1 according to an embodiment of the present disclosure.
  • a knowledge graph for the dialogue behavior information of the user is established.
  • the established knowledge graph includes nodes, sub-nodes, attributes and attribute values, the relationship between nodes and nodes, and the relationship between node-attribute-attribute values.
  • nodes include location nodes, information nodes, and usage scenarios nodes.
  • information nodes are further divided into at least one or more of time nodes, interest nodes, gender nodes, age nodes, and social feature nodes.
  • its child nodes include weather nodes, music nodes, and story nodes.
  • the child nodes of the weather node include traffic state nodes and protective facilities nodes.
  • the frequency is used as an attribute of each node and each child node, and correspondingly, the frequency value (specific value, for example, 5 times) is used as An attribute value for each node and each child node.
  • the frequency value refers to the number of times the node or the child node is queried in the historical query record of the knowledge graph.
  • Fig. 3 is a flowchart of step S1 of a man-machine multi-round dialogue method for man-machine interaction according to an embodiment of the present disclosure.
  • the knowledge graph is based on the basic graph template framework and is established by the user through multiple previous trainings. As shown in Figure 3, the specific process is as follows:
  • Step S11 Establish a basic graph template framework including basic nodes and relationships
  • Step S12 Collect the information input by the user during training, and form structured data
  • Step S13 Perform entity recognition and naming on the structured data, and form a new node corresponding to the information input by the user in the knowledge graph;
  • Step S14 Extract the relationship data between the new node and the original node in the knowledge graph from a preset text corpus
  • Step S15 Perform similarity calculation and matching between the new node and the original node, and combine the relationship data to form a structure of node-relation-node and node-attribute-attribute value.
  • step S11 the developer establishes a basic graph template framework including basic nodes and relationships according to the general needs of public users. For example, current users are more concerned with information such as weather, traffic, and news. At this time, in some implementations, weather, traffic, and news can be formed as basic nodes in the basic map template framework. For another example, a general user first inquires about the weather and then about the traffic. In some embodiments, the traffic state node can be regarded as a child node of the weather node. In the embodiment of the present disclosure, frequency is used as an attribute of the nodes and sub-nodes in the basic graph template framework, and correspondingly, the frequency value is taken as an attribute value of the nodes and sub-nodes in the basic graph template framework.
  • the basic atlas template framework established in step S11 may not be able to specifically satisfy a certain user. Therefore, it is necessary to train the basic atlas template framework to obtain a specific Knowledge graph of users.
  • step S12 the information input by the user during training is collected, and a top-down construction method is used to form structured data.
  • the formed structured data is data logically expressed by a two-dimensional table structure, with strict format and length specifications.
  • the semantic information input by the user is "Do you need to open an umbrella?"
  • the structured data formed is "open umbrella”.
  • step S13 entity recognition.
  • entity recognition and naming are performed on the structured data, and a new node corresponding to the information input by the user is formed in the knowledge graph.
  • the server log is used to search, and the semantic feature corresponding to the structured data is searched out to identify the structured data and name it. For example, for the structured data "Umbrella”, according to the results of the server log search, "Umbrella” is identified as an action, and it is named “Action, Umbrella”. Then, a new node "action, umbrella” is generated in the knowledge graph.
  • step S14 relation extraction.
  • the relationship data between the new node and the original node in the knowledge graph is extracted from the preset text corpus.
  • the original node refers to the existing node in the knowledge graph before the new node "action, umbrella" is formed.
  • a method of semantic rule pattern matching is used for relation extraction.
  • step S15 knowledge fusion.
  • the purpose of knowledge fusion in this step is to eliminate ambiguity so as to facilitate the fusion of nodes and nodes.
  • similarity calculations are performed on the new nodes and the original nodes in sequence, and this process includes: object alignment, entity matching, and entity synonymous processing.
  • the attribute value (frequency value) of the attribute frequency of the "weather” node is increased by 1.
  • the new node is retained in the knowledge graph, and the attribute value (frequency value) of the attribute frequency of the new node is increased by 1.
  • the relationship data obtained in step S14 is combined to form a structure of node-relation-node and node-attribute-attribute value.
  • these two structures are also referred to as entity-relation-entity and entity-attribute-attribute value structures.
  • a knowledge graph specifically for the user's dialogue behavior information is obtained.
  • the user will be supplemented with related intentions. For example, if the user says “How is the weather today?", the machine will ask “Which city do you want to check the weather?” to supplement the user's intention and send the user dialogue information
  • the intention in is updated to the node of the knowledge graph or the attribute value of the node to establish a knowledge graph of the user’s dialogue behavior information.
  • step S2 to step S4 are executed next.
  • step S2 according to the information currently input by the user, the node corresponding to the information currently input and the child nodes of the node are determined in the knowledge graph.
  • the weather node As shown in Figure 2, taking the weather node as an example, if the user asks "How is the weather today?", through semantic analysis the user wants to query "weather", then the weather node and the child nodes of the weather node are determined in the knowledge graph , And select the attributes (such as frequency) under the weather node, traverse the sub-nodes such as protective measures, traffic status, and clothing matching to form a set, and read the attribute values (frequency values) of the nodes and sub-nodes as items.
  • the attributes such as frequency
  • step S3 is executed: according to the number of times the node is queried and the number of times the child nodes and the node are simultaneously queried in the historical query records of the knowledge graph, the support degree of the child nodes relative to the node is calculated. Specifically, see Table 1.
  • each user request is recorded as a transaction, and each transaction records the nodes involved, where 1 is the question and answer involved in the transaction, and 0 is the question and answer not involved in the transaction.
  • its child nodes include: protective measures B, traffic status C, and clothes matching D.
  • the weather node A was queried three times in total, that is, the frequency value is 3, and the dressing match D was queried only once when the request number was 3, that is, the dressing match D and the weather The frequency value of node A being queried at the same time is 1.
  • the support degree can be obtained by calculating the ratio of the number of times the child nodes and nodes are simultaneously queried to the number of times the node is queried. which is,
  • the support for dressing match D is 33%. It should be noted that the transaction that A and D appear at the same time is that the dressing collocation D and the weather node A are queried at the same time (when the weather node A is queried, the dressing collocation D is also queried, as shown in the request number 3 in Table 1. In the case of), the total transaction is the frequency value of the weather node A being queried.
  • the support degree of protective measure B relative to weather node A (hereinafter referred to as the support degree of protective measure B), and the support degree of traffic state C relative to weather node A (hereinafter referred to as the support of traffic state C) are sequentially calculated. degree). According to the content shown in Table 1, it is calculated that the support degree of protective measure B is 66%, and the support degree of traffic state C is 66%.
  • step S4 is executed next: the size relationship between the support degree and the preset support degree threshold is determined to determine the information to be output.
  • step S4 it is judged whether the support degree is greater than the preset support threshold value: if the support degree is greater than the preset support threshold value, the semantic information of the child node is output while the semantic information of the node is output; otherwise only the semantic information of the node is output .
  • the preset support threshold is set to 50%.
  • the size of the preset support threshold can be flexibly set according to actual conditions, and the present disclosure is not limited to this.
  • the semantic information of dressing collocation D is not output; while the support for protective measure B and traffic state C are both greater than the preset support With a threshold of 50%, while outputting the semantic information of the weather node, it also outputs the semantic information of child node protection measures B and child node traffic status C.
  • the machine will output "The weather in Guangzhou is sunny and the ultraviolet rays are strong. Please use sunscreen appropriately, South China Express Line processing congestion status" these information.
  • the support of all child nodes is not greater than the preset support threshold of 50%, only the semantic information of the weather node will be output, that is, the machine will only output "Guangzhou Weather Sunny".
  • step S4 can also be performed in the following manner: determine whether the support degree is greater than the preset support degree threshold: if the support degree is greater than the preset support threshold value, then the form includes The frequent set of all sub-nodes with support greater than the preset support threshold, and the semantic information of the node is output, and the semantic information of each sub-node in the frequent set is output in order from the largest support to the smallest; otherwise only output The semantic information of the node.
  • the frequent set includes protective measures B and traffic status C, and then While outputting the semantic information of the weather node, the semantic information of frequently concentrated protective measures B and traffic status C is output in order of support degree from large to small.
  • the output voice information takes too long, which affects the user's interactive experience.
  • the semantic information of the output node only the semantic information of the top three child nodes with the highest frequently concentrated support is output.
  • step S5 is added: real-time update according to the information input by the user in real time
  • the knowledge graph that is, after the establishment of the knowledge graph is completed, or after the user’s pre-training is completed, in order to enable the knowledge graph to be updated in real time with the user’s interest or inquiry intention and improve the user’s interactive experience, the second embodiment of the present disclosure is based on the information input by the user in real time. Update the knowledge graph in real time.
  • step S5 includes:
  • Step S51 Collect the information input by the user in real time and form structured data
  • Step S52 Perform entity recognition and naming on the structured data, and form a new node corresponding to the information input by the user in real time in the knowledge graph;
  • Step S53 Extract the relationship data between the new node and the original node in the knowledge graph from a preset text corpus
  • Step S54 Perform similarity calculation and matching between the new node and the original node, and combine the relationship data to form a structure of node-relation-node and node-attribute-attribute value.
  • step S51 the information input by the user in real time is collected, and structured data is formed using a top-down construction method.
  • the formed structured data is data logically expressed by a two-dimensional table structure, with strict format and length specifications.
  • the semantic information input by the user is "Need to bring an umbrella?"
  • the structured data formed is "with an umbrella”.
  • step S52 entity recognition.
  • entity recognition and naming are performed on the structured data, and a new node corresponding to the information input by the user is formed in the knowledge graph.
  • a server log is used to search to find out the semantic features corresponding to the structured data, so as to identify the structured data and name it. For example, for structured data "with umbrella”, according to the search results of the server log, "with umbrella” is identified as an action, and it is named “action, with umbrella”. Then, a new node "action with umbrella” is generated in the knowledge graph.
  • step S53 relation extraction.
  • the relationship data between the new node and the original node in the knowledge graph is extracted from the preset text corpus.
  • the original node refers to the existing node in the knowledge graph before the new node "action, umbrella" is formed.
  • a method of semantic rule pattern matching is used for relation extraction.
  • step S54 knowledge fusion.
  • the purpose of knowledge fusion in this step is to eliminate ambiguity so as to facilitate the fusion of nodes and nodes.
  • similarity calculations are performed on the new nodes and the original nodes in sequence, and this process includes: object alignment, entity matching, and entity synonymous processing.
  • step S53 the relationship data obtained in step S53 is combined to form a structure of node-relation-node and node-attribute-attribute value.
  • these two structures are also called entity-relation-entity and entity-attribute-attribute value structures.
  • the machine does not need to use multiple rounds of voice dialogue to understand the user's intention, and the user's voice interaction experience is improved.
  • applying the man-machine multi-round dialogue method for man-machine interaction provided in the second embodiment of the present disclosure can update the knowledge graph in real time according to the user's interest or inquiry intention, and improve the user's voice interaction experience.
  • the third embodiment of the present disclosure provides a man-machine multi-round dialogue system for man-machine interaction.
  • Fig. 6 schematically shows a man-machine multi-round dialogue system for man-machine interaction according to the third embodiment of the present disclosure.
  • the system 600 includes:
  • Knowledge graph establishment module 601 which is configured to establish a knowledge graph of user dialogue behavior information
  • Index module 602 which is configured to: according to the information currently input by the user, determine the node corresponding to the currently input information and the child nodes of the node in the knowledge graph;
  • the calculation module 603 is configured to calculate the number of the child node relative to the node according to the number of times the node is queried and the number of times the child node and the node are simultaneously queried in the historical query record of the knowledge graph Support;
  • the judging module 604 is configured to determine whether to output the semantic information of the child node by judging the magnitude relationship between the support degree and a preset support degree threshold.
  • system 600 also includes:
  • the input module 605 which is configured to receive information input by the user;
  • the output module 606 is connected to the judgment module 604, and is configured to output the information to be output according to the judgment result of the judgment module 604.
  • the knowledge graph establishing module 601 establishes a knowledge graph for the dialogue behavior information of the user.
  • the established knowledge graph includes nodes, sub-nodes, attributes and attribute values, the relationship between nodes and (child) nodes, and the relationship between node-attribute-attribute values.
  • nodes include location nodes, information nodes, and usage scenarios nodes.
  • information nodes are further divided into at least one or more of time nodes, interest nodes, gender nodes, age nodes, and social feature nodes.
  • its child nodes include weather nodes, music nodes, and story nodes.
  • the child nodes of the weather node include traffic state nodes and protective facilities nodes.
  • the frequency is used as an attribute of each node and each child node, and correspondingly, the frequency value (specific value, for example, 5 times) is used as An attribute value for each node and each child node.
  • the frequency value refers to the number of times the node or the child node is queried in the historical query record of the knowledge graph.
  • the knowledge graph establishment module 601 builds a knowledge graph based on the basic graph template framework through multiple previous trainings of the user.
  • the specific process is as follows:
  • the knowledge graph establishment module 601 establishes a basic graph template framework including basic nodes and relationships according to the general needs of the mass users. For example, in some implementations, current users are more concerned with information such as weather, traffic, and news. At this time, weather, traffic, and news are formed as basic nodes in the basic map template framework. For another example, in some implementations, a general user first asks about the weather and then about the traffic. At this time, the traffic state node is regarded as a child node of the weather node.
  • frequency is used as an attribute of the nodes and sub-nodes in the basic graph template framework, and correspondingly, the frequency value is taken as an attribute value of the nodes and sub-nodes in the basic graph template framework.
  • the basic atlas template framework established may not be able to specifically satisfy a certain user. Therefore, it is necessary to train the basic atlas template framework to obtain a specific user-specific Knowledge graph.
  • the knowledge graph establishment module 601 uses the top-down construction method to form structured data according to the information input by the user during training collected by the input module 605.
  • the formed structured data is data logically expressed by a two-dimensional table structure, with strict format and length specifications.
  • the semantic information input by the user is "Do you need to open an umbrella?"
  • the structured data formed is "open umbrella”.
  • the knowledge graph establishment module 601 performs entity recognition and naming of the structured data, and forms a new node corresponding to the information input by the user in the knowledge graph. Specifically, based on the obtained structured data, a server log is used to search, and the semantic features corresponding to the structured data are searched out to identify the structured data and name it. For example, for the structured data "Umbrella”, according to the results of the server log search, "Umbrella” is identified as an action, and it is named “Action, Umbrella”. Then, a new node "action, umbrella” is generated in the knowledge graph.
  • the knowledge graph establishment module 601 extracts the relationship data between the new node and the original node in the knowledge graph from the preset text corpus.
  • the original node refers to the existing node in the knowledge graph before the new node "action, umbrella" is formed.
  • the knowledge graph establishment module 601 performs relation extraction by running a semantic rule pattern matching algorithm stored in itself.
  • the knowledge graph establishment module 601 performs knowledge fusion between the new node and the original node.
  • the purpose of knowledge fusion is to eliminate ambiguity to facilitate the fusion of nodes and nodes.
  • similarity calculations are performed on the new nodes and the original nodes in sequence, and this process includes: object alignment, entity matching, and entity synonymous processing.
  • the attribute value (frequency value) of the attribute frequency of the "weather” node is increased by 1.
  • the new node is retained in the knowledge graph, and the attribute value (frequency value) of the attribute frequency of the new node is increased by 1.
  • these two structures are also referred to as entity-relation-entity and entity-attribute-attribute value structures.
  • the knowledge graph establishing module 601 is based on the basic graph template framework, through the user's preliminary training, that is, in the preliminary interaction process, to supplement the user's related intentions.
  • the machine will ask “Which city do you want to check the weather” to supplement the user's intention and update the intention in the user's dialogue information to the knowledge graph Among the node or the attribute value of the node, the knowledge graph of the dialogue behavior information of the user is established.
  • the index module 602 After the knowledge graph is established, the index module 602: according to the information currently input by the user, determine the node corresponding to the currently input information and the child nodes of the node in the knowledge graph.
  • the index module 602 obtains that the user wants to query "weather” through semantic analysis, and then the index module 602 determines the weather node in the knowledge graph And the child nodes of the weather node, and select the attributes (such as frequency) under the weather node, traverse the child nodes such as protective measures, traffic status and clothing matching, etc. to form a set, and read the attribute values (frequency value) of the nodes and child nodes ) As the item.
  • the calculation module 603 calculates the relative value of the child node according to the number of times the node is queried and the number of times the child node and the node are simultaneously queried in the historical query record of the knowledge graph The support of the node. Specifically, see Table 1:
  • each user request is recorded as a transaction, and each transaction records the nodes involved, where 1 is the question and answer involved in the transaction, and 0 is the question and answer not involved in the transaction.
  • its child nodes include: protective measures B, traffic status C, and clothes matching D.
  • the weather node A was queried three times in total, that is, the frequency value is 3, and the dressing match D was queried only once when the request number was 3, that is, the dressing match D and the weather The frequency value of node A being queried simultaneously is 1.
  • the calculation module 603 calculates the degree of support between the dressing collocation D and the weather node A. Specifically, the calculation module 603: calculates the ratio of the number of times the child node and the node are simultaneously queried to the number of times the node is queried to obtain Support. which is,
  • the calculation module 603 obtains that the support degree of the dressing match D is 33%. It should be noted that the transaction that A and D appear at the same time is that the dressing collocation D and the weather node A are queried at the same time (when the weather node A is queried, the dressing collocation D is also queried, as shown in the request number 3 in Table 1. In the case of), the total transaction is the frequency value of the weather node A being queried.
  • the calculation module 603 sequentially calculates the support degree of protective measure B relative to weather node A (hereinafter referred to as the support degree of protective measure B), and the support degree of traffic state C relative to weather node A (hereinafter referred to as traffic state). C's support). According to the content shown in Table 1, the calculation module 603 calculates that the support degree of the protective measure B is 66%, and the support degree of the traffic state C is 66%.
  • the judgment module 604 Determine the information to be output by judging the magnitude relationship between the support degree and the preset support degree threshold.
  • the judging module 604 judges whether the support is greater than the preset support threshold: if the support is greater than the preset support threshold, the semantic information of the child node is output while the semantic information of the node is output; otherwise, only the semantic information of the node is output.
  • the preset support threshold is set to 50%.
  • the size of the preset support threshold can be flexibly set according to actual conditions, and the present disclosure is not limited to this.
  • the output module 606 After the judgment module 604 judges that the support degree of the dressing match D is less than 50% of the preset support threshold, the semantic information of the dressing match D will not be output; while the support of the protective measure B and the support of the traffic state C are both greater than With a preset support threshold of 50%, the output module 606 outputs the semantic information of the child node protection measures B and the child node traffic status C while outputting the semantic information of the weather node. For example, the output module 606 will output “The weather in Guangzhou is clear, the ultraviolet rays are strong , Please use sunscreen appropriately, and the South China Express will handle the congestion status" information. In addition, if the support of all child nodes is not greater than the preset support threshold of 50%, the output module 606 only outputs the semantic information of the weather node, for example, the output module 606 only outputs the information of "Guangzhou weather sunny".
  • the judgment module 604 also judges by the following method: judging whether the support degree of each child node is greater than the preset support threshold: if the support is greater than the preset support threshold, Then a frequent set containing all sub-nodes with support greater than the preset support threshold is formed, and while outputting the semantic information of the node, the semantic information of each sub-node in the frequent set is output in order of support in descending order; Otherwise, only the semantic information of the node is output.
  • the judging module 604 forms a frequent set of all sub-nodes with a support greater than the preset support threshold.
  • the frequent set includes protective measures B and traffic status C.
  • the output module 606 according to the judgment result of the judgment module 604: while outputting the semantic information of the weather node, it sequentially outputs the frequently concentrated semantic information of the protective measures B and the traffic state C in order of the support degree from large to small.
  • the output module 606 In order to avoid that the output voice information takes too long, which affects the user's interactive experience.
  • it is set by setting: while outputting the semantic information of the node, the output module 606 only outputs the semantic information of the top three child nodes with the highest frequently concentrated support.
  • the system of the third embodiment of the present disclosure further includes: knowledge graph update
  • the module 607 is configured to update the knowledge graph in real time according to the information input by the user in real time.
  • the knowledge graph update module 607 uses the top-down construction method to form structured data according to the real-time input information collected by the input module 605 by the user.
  • the formed structured data is data logically expressed by a two-dimensional table structure, with strict format and length specifications.
  • the semantic information input by the user is "Need to bring an umbrella?"
  • the structured data formed is "with an umbrella”.
  • the knowledge graph update module 607 performs entity recognition and naming of the structured data, and forms a new node in the knowledge graph corresponding to the information input by the user. Specifically, based on the obtained structured data, a server log is used to search, and the semantic features corresponding to the structured data are searched out to identify the structured data and name it. For example, for structured data "with umbrella”, according to the search results of the server log, "with umbrella” is identified as an action, and it is named “action, with umbrella”. Then, a new node with "action, umbrella” is generated in the knowledge graph.
  • the knowledge graph update module 607 extracts the relationship data between the new node and the original node in the knowledge graph from the preset text corpus.
  • the original node refers to the node that already exists in the knowledge graph before the new node "action, umbrella" is formed.
  • the knowledge graph updating module 607 performs relation extraction by running an algorithm of semantic rule pattern matching stored in itself.
  • the knowledge graph update module 607 sequentially calculates the similarity between the new node and the original node.
  • it includes: object alignment, entity matching, and entity synonymous processing.
  • it can be determined whether the new node can be integrated into the original node. For example, after similarity calculation and matching, it is determined that the new node “action, umbrella” is related to the original node “weather” node, then the new node “action, umbrella” is merged with the "weather” node and deleted Add 1 to the attribute value (frequency value) of the attribute frequency of the "weather” node when the new node "action, with umbrella” is added.
  • the new node is retained in the knowledge graph, and the attribute value (frequency value) of the attribute frequency of the new node is increased by 1. Then, combine the obtained relationship data to form a structure of node-relation-node and node-attribute-attribute value.
  • these two structures are also referred to as entity-relation-entity and entity-attribute-attribute value structures.
  • the knowledge graph update module 607 updates the knowledge graph in real time.
  • the updating of the knowledge graph is also completed by the knowledge graph establishment module 601.
  • the system 600 does not include the knowledge graph update module 607. Disclosure is not limited to this.
  • an embodiment of the present disclosure also provides a smart device, including:
  • a memory which stores executable code, when the executable code is executed by the processor, causes the processor to execute the human-machine multi-round dialogue for human-computer interaction in the first or second embodiment above method.
  • the machine does not need to use multiple rounds of voice dialogue to understand the user’s intentions, which improves the user’s voice interaction.
  • the machine does not need to use multiple rounds of voice dialogue to understand the user’s intentions, which improves the user’s voice interaction.
  • applying the human-machine multi-round dialogue method and system for human-computer interaction, and smart devices provided by the embodiments of the present disclosure can update the knowledge graph in real time according to the user's interest or inquiry intention, and improve the user's voice interaction experience.
  • modules or steps of the present disclosure can be implemented by a general computing device, and they can be concentrated on a single computing device or distributed on a network composed of multiple computing devices.
  • they can be implemented with program codes executable by a computing device, so that they can be stored in a storage device to be executed by the computing device, or they can be made into individual integrated circuit modules, or many of them Each module or step is made into a single integrated circuit module to achieve.
  • the present disclosure is not limited to any specific hardware and software combination.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Animal Behavior & Ethology (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种用于人机交互的人机多轮对话方法及系统、智能设备，该方法包括：步骤S1：建立用户对话行为信息的知识图谱；步骤S2：根据用户当前输入的信息，在所述知识图谱中确定出与所述当前输入的信息对应的节点和该节点的子节点；步骤S3：根据所述知识图谱的历史查询记录中所述节点被查询的次数以及所述子节点和所述节点同时被查询的次数，计算所述子节点相对于所述节点的支持度；以及步骤S4：通过判断所述支持度与预设支持度阈值的大小关系，确定是否输出所述子节点的语义信息。

Description

用于人机交互的人机多轮对话方法及系统、智能设备
本公开以2019年8月12日递交的、申请号为201910740449.4且名称为“用于人机交互的人机多轮对话方法及系统、智能设备”的专利文件为优先权文件，其全部内容通过引用结合在本公开中。
技术领域
本公开涉及人机交互技术领域，尤其涉及一种用于人机交互的人机多轮对话方法及系统、智能设备。
背景技术
随着智能家居和智能家电的快速发展，语音交互逐渐成为智能家居和智能家电的标配。然而，虽然当前语音交互实现了基本的任务型交互，满足了指令控制，但是在多维度对话理解以及智能应答过程中，依旧存在不智能的状态，造成语音交互体验差。
从人与人对话失败的角度来看，其中原因往往是双方缺少共同的认知背景。即使在功能性对话中，形式与内容也有可能是不准确的，需要修复才可以回到正确的轨道上。对话是基于一种原则协同互动行为，对话的参与方在丰富而微妙的语境下创造并认同语言的表意。虽然当前人机交互能够通过多轮对话，上下文关联，实现用户意图识别，但多轮对话会造成交互体验差。因此，机器与人之间进行语音交互，需要对机器建立认知背景，以了解用户所要表达的含义，进而减少人机交互多轮对话，达到智能交互的效果。
发明内容
本公开所要解决的技术问题是克服目前智能设备需要多轮对话才能理解用户意图，造成语音交互体验差的问题。
为了解决上述技术问题，本公开提供了一种用于人机交互的人机多轮对话方法，所述方法包括如下步骤：
步骤S1：建立用户对话行为信息的知识图谱；
步骤S2：根据用户当前输入的信息，在所述知识图谱中确定出与所述当前输入的信息对应的节点和该节点的子节点；
步骤S3：根据所述知识图谱的历史查询记录中所述节点被查询的次数以及所述子节点和所述节点同时被查询的次数，计算所述子节点相对于所述节点的支持度；以及
步骤S4：通过判断所述支持度与预设支持度阈值的大小关系，确定是否输出所述子节点的语义信息。
优选地，在步骤S3中，通过计算所述子节点和所述节点同时被查询的次数与所述节点被查询的次数的比值，得到所述支持度。
在一实施例中，在步骤S4中，
判断所述支持度是否大于预设支持度阈值：若所述支持度大于所述预设支持度阈值，则在输出所述节点的语义信息的同时，输出所述子节点的语义信息；否则只输出所述节点的语义信息。
在另一实施例中，在步骤S4中，
判断所述支持度是否大于预设支持度阈值：若所述支持度大于所述预设支持度阈值，则形成包含所有所述支持度大于所述预设支持度阈值的子节点的频繁集，并在输出所述节点的语义信息的同时，按照所述支持度从大到小的顺序依次输出所述频繁集中的每个子节点的语义信息；否则只输出所述节点的语义信息。
优选地，所述方法还包括：
步骤S5：根据用户实时输入的信息，实时更新所述知识图谱。
优选地，步骤S5包括：
步骤S51：采集用户实时输入的信息，并形成结构化数据；
步骤S52：对所述结构化数据进行实体识别和命名，并在所述知识图谱中形成与所述用户实时输入的信息对应的新的节点；
步骤S53：从预置的文本语料库中抽取所述新的节点与所述知识图谱中原有的节点之间的关系数据；以及
步骤S54：对所述新的节点与所述原有的节点进行相似度计算和匹配，并结合所述关系数据，以形成节点-关系-节点和节点-属性-属性值的结构。
根据本公开的一个方面，提供了一种用于人机交互的人机多轮对话系统，所述系统包括：
知识图谱建立模块，其被配置为建立用户对话行为信息的知识图谱；
索引模块，其被配置为：根据用户当前输入的信息，在所述知识图谱中确定出与所述当前输入的信息对应的节点和该节点的子节点；
计算模块，根据所述知识图谱的历史查询记录中所述节点被查询的次数以及所述子节点和所述节点同时被查询的次数，计算所述子节点相对于所述节点的支持度；以及
判断模块，其被配置为通过判断所述支持度与预设支持度阈值的大小关系，确定是否输出所述子节点的语义信息。
优选地，根据所述知识图谱的历史查询记录中所述节点被查询的次数以及所述子节点和所述节点同时被查询的次数，计算所述子节点相对于所述节点的支持度，包括：通过计算所述子节点和所述节点同时被查询的次数与所述节点被查询的次数的比值，得到所述支持度。
优选地，所述系统还包括：
知识图谱更新模块，其被配置为根据用户实时输入的信息，实时更新所述知识图谱。
根据本公开的另一个方面，提供了一种智能设备，包括：
处理器；以及
存储器，其上存储有可执行代码，所述可执行代码在被所述处理器执行时，使所述处理器执行上述的用于人机交互的人机多轮对话方法。
与现有技术相比，上述方案中的一个或多个实施例可以具有如下优点或有益效果：
应用本公开实施例提供的用于人机交互的人机多轮对话方法及系统、智能设备，机器不需要通过多轮语音对话来进行了解用户意图，提高了用户语音交互的体验。
本公开的其它特征和优点将在随后的说明书中阐述，并且部分地从说明书中变得显而易见，或者通过实施本公开而了解。本公开的目的和其他优点可通过在说明书、权利要求书以及说明书附图中所特别指出的结构来实现和获得。
附图说明
附图用来提供对本公开的进一步理解，并且构成说明书的一部分，与本公开 的实施例共同用于解释本公开，并不构成对本公开的限制。
图1为根据本公开实施例一的用于人机交互的人机多轮对话方法的流程图。
图2示意性示出了根据本公开实施例的步骤S1建立的用户对话行为信息的知识图谱。
图3为根据本公开实施例的用于人机交互的人机多轮对话方法的步骤S1的流程图。
图4为根据本公开实施例二的用于人机交互的人机多轮对话方法的流程图。
图5为根据本公开实施例二的用于人机交互的人机多轮对话方法的步骤S5的流程图。
图6示意性示出了根据本公开实施例三的用于人机交互的人机多轮对话系统。
具体实施方式
以下将结合附图及实施例来详细说明本公开的实施方式，借此对本公开如何应用技术手段来解决技术问题，并达成技术效果的实现过程能充分理解并据以实施。需要说明的是，只要不构成冲突，本公开中的各个实施例以及各实施例中的各个特征可以相互结合，所形成的技术方案均在本公开的保护范围之内。
人与人之间对话过程中，重要的信息不一定通过对话获取，对话本身所包含的信息也只是占传递信息量的一小部分，更多信息的信息来源于说话人的信息，当前的时间或者地点等一系列场景信息。当前由于机器对用户缺少认知，需要通过多轮对话实现更多信息的获取，从而补充用户的意图。
为了解决现有技术中智能设备需要多轮对话才能理解用户，造成语音交互体验差的技术问题，本公开实施例提供了一种用于人机交互的人机多轮对话方法及系统、智能设备。
实施例一
图1为根据本公开实施例一的用于人机交互的人机多轮对话方法的流程图。如图1所示，该方法包括如下步骤：
步骤S1：建立用户对话行为信息的知识图谱；
步骤S2：根据用户当前输入的信息，在所述知识图谱中确定出与所述当前输 入的信息对应的节点和该节点的子节点；
步骤S3：根据所述知识图谱的历史查询记录中所述节点被查询的次数以及所述子节点和所述节点同时被查询的次数，计算所述子节点相对于所述节点的支持度；以及
步骤S4：通过判断所述支持度与预设支持度阈值的大小关系，确定是否输出所述子节点的语义信息。
在步骤S1中，建立用户对话行为信息的知识图谱。首先，需要说明的是，在一些实施方式中，该知识图谱是专门只针对于某一用户的，即对于每个用户都建立一个专门只针对于该用户的知识图谱，在另一些实施方式中，也是多名用户共用一个知识图谱，本公开不限于此。
为了便于理解与说明，下面以一个用户为例进行说明。
图2示意性示出了根据本公开实施例的步骤S1建立的用户对话行为信息的知识图谱。如图2所示，在步骤S1中，针对于某一用户，建立了针对该名用户对话行为信息的知识图谱。建立的知识图谱包括节点、子节点、属性和属性值以及节点与节点之间的关系、节点-属性-属性值之间的关系。
例如，节点包括地点节点、信息节点和使用场景节点。例如，在一些实施方式中，信息节点又分为时间节点、兴趣节点、性别节点、年龄节点和社会特征节点中的至少一种或多种。其中，以使用场景节点为例，它的子节点包括天气节点、音乐节点和故事节点等。在一些实施方式中，天气节点的子节点又包括交通状态节点和防护设施节点等。在本公开实施例中，对于知识图谱中的每个节点和每个子节点，均将频次作为每个节点和每个子节点的一个属性，对应地，将频次值(具体数值，例如5次)作为每个节点和每个子节点的一个属性值。其中，频次值指的是在知识图谱的历史查询记录中该节点或该子节点被查询的次数。
图3为根据本公开实施例的用于人机交互的人机多轮对话方法的步骤S1的流程图。在步骤S1中，在一些实施方式中，知识图谱是在基本图谱模板框架的基础上，由用户前期多次训练而建立。如图3所示，具体过程如下：
步骤S11：建立包含基本节点以及关系的基本图谱模板框架；
步骤S12：采集用户训练时输入的信息，并形成结构化数据；
步骤S13：对所述结构化数据进行实体识别和命名，并在所述知识图谱中形成与所述用户输入的信息对应的新的节点；
步骤S14：从预置的文本语料库中抽取所述新的节点与所述知识图谱中原有的节点之间的关系数据；以及
步骤S15：对所述新的节点与所述原有的节点进行相似度计算和匹配，并结合所述关系数据，以形成节点-关系-节点和节点-属性-属性值的结构。
具体地，在步骤S11中，由开发人员根据大众用户的普遍需求建立包含基本节点以及关系的基本图谱模板框架。例如，目前用户均比较关心天气、交通和新闻等信息，这时，在一些实施方式中，就能把天气、交通和新闻作为基本节点形成在基本图谱模板框架之中。再例如，在一些实施例中，一般用户都是先询问天气，再询问交通，这时，就能把交通状态节点作为天气节点的子节点。在本公开实施例中，将频次作为基本图谱模板框架中节点和子节点的一个属性，对应地，将频次值作为基本图谱模板框架中节点和子节点的一个属性值。
因用户的询问信息可能因人而异，所以步骤S11所建立的基本图谱模板框架很可能无法专门满足某一用户，所以这时就需要对基本图谱模板框架进行训练，以得到专门针对于某一名用户的知识图谱。
具体地，在步骤S12中，采集用户训练时输入的信息，运用自顶向下构建的方式，形成结构化数据。其中，所形成的结构化数据为由二维表结构来逻辑表达的数据，有着严格的格式和长度规范。例如，在一些实施方式中，用户输入的语义信息为“需要打伞吗？”，这时，经过自顶向下构建的方式，所形成的结构化数据为“打伞”。
接下来，执行步骤S13：实体识别。在步骤S13中，对结构化数据进行实体识别和命名，并在知识图谱中形成与用户输入的信息对应的新的节点。具体地，基于步骤S12得到的结构化数据，利用服务器日志进行搜索，搜索出该结构化数据对应的语义特征，以识别该结构化数据，并对其进行命名。比如，对于结构化数据“打伞”，根据服务器日志搜索的结果，识别出“打伞”为动作，并将其命名为“动作，打伞”。然后，在知识图谱中生成新的节点“动作，打伞”。
接下来，执行步骤S14：关系抽取。在步骤S14中，从预置的文本语料库中抽取新的节点与知识图谱中原有的节点之间的关系数据。其中，原有的节点指的是在形成“动作，打伞”这个新的节点之前，知识图谱中已存在的节点。优选地，在该步骤S14中，利用语义规则模式匹配的方法进行关系抽取。
接下来，执行步骤S15：知识融合。本步骤知识融合的目的在于：消除歧义， 以便于节点与节点进行融合。具体地，依次将新的节点与原有的节点进行相似度计算，在这过程中，包括：对象对齐、实体匹配和实体同义处理。在一些实施方式中，经过上述处理过程，就能确定新的节点是否能融合到原有的节点之中。例如，经过相似度计算和匹配，确定出新的节点“动作，打伞”和原有的节点“天气”节点相关，则将新的节点“动作，打伞”与“天气”节点相融合，并且在删除新的节点“动作，打伞”的同时，令“天气”节点的属性频次的属性值(频次值)加1。另外，如果新的节点与原有的节点不相关，则在知识图谱中保留该新的节点，并令新的节点的属性频次的属性值(频次值)加1。然后，结合步骤S14得到的关系数据，形成节点-关系-节点和节点-属性-属性值的结构。在一些实施方式中，这两种结构被称为实体-关系-实体和实体-属性-属性值的结构。
由此，在基本图谱模板框架的基础上，通过用户前期的训练，得到专门针对该用户对话行为信息的知识图谱。在前期的交互过程中，给予用户相关意图的补充，如用户说“今天天气怎么样？”，机器会问“请问你要查哪个城市的天气”，来补充用户的意图，并将用户对话信息中的意图更新到知识图谱的节点或者节点的属性值之中，以建立该名用户对话行为信息的知识图谱。
在知识图谱建立之后，接下来执行步骤S2至步骤S4。
在步骤S2中，根据用户当前输入的信息，在知识图谱中确定出与当前输入的信息对应的节点和该节点的子节点。如图2所示，以天气节点为例，如果用户问“今天天气怎么样？”，通过语义分析得到用户想要查询“天气”，则在知识图谱中确定出天气节点和天气节点的子节点，并选定天气节点下的属性(例如频次)，遍历子节点如防护措施、交通状态和穿衣搭配等形成集合，并读取出节点和子节点的属性值(频次值)作为项。
接下来，执行步骤S3：根据知识图谱的历史查询记录中节点被查询的次数以及子节点和节点同时被查询的次数，计算子节点相对于节点的支持度。具体地，参见表1。
表1
请求序号 天气A 防护措施B 交通状态C 穿衣搭配D
1 1 0 1 0
2 1 1 1 0
3 1 1 0 1
如表1所示，在知识图谱中，用户每次请求记作一次事务，每次事务记录涉及的节点，其中1为该次事务涉及问答，0为该次事务不涉及的问答。以天气节点A为例，它的子节点包括：防护措施B、交通状态C和穿衣搭配D。例如，在知识图谱的历史查询记录中，天气节点A总共被查询了三次，即频次值为3，而穿衣搭配D只在请求序号为3时被查询过一次，即穿衣搭配D与天气节点A同时被查询的频次值为1。现在来计算穿衣搭配D相对于天气节点A之间的支持度，可依据以下方式：通过计算子节点和节点同时被查询的次数与节点被查询的次数的比值，得到支持度。即，
求穿衣搭配D相对于天气节点A的支持度(以下简称穿衣搭配D的支持度)为：
支持度P＝A与D同时出现的事务/总的事务＝1/3＝33％，
穿衣搭配D相对于天气节点A的支持度为：P＝P(A&D)/P(A)＝1/3＝33％。
通过上述计算，得到穿衣搭配D的支持度为33％。需要说明的是，A与D同时出现的事务即穿衣搭配D与天气节点A同时被查询(在天气节点A被查询时，穿衣搭配D也被查询，例如表1中请求序号3所示的情况)的频次值，总的事务即天气节点A被查询的频次值。
依照上述同样的方式，依次计算出防护措施B相对于天气节点A的支持度(以下简称防护措施B的支持度)，交通状态C相对于天气节点A的支持度(以下简称交通状态C的支持度)。根据表1所示的内容，经过计算得出防护措施B的支持度为66％，交通状态C的支持度为66％。
在计算完各子节点的支持度之后，接下来，执行步骤S4：通过判断所述支持度与预设支持度阈值的大小关系，以确定所要输出的信息。在步骤S4中，判断支持度是否大于预设支持度阈值：若支持度大于预设支持度阈值，则在输出节点的语义信息的同时，输出子节点的语义信息；否则只输出节点的语义信息。
优选地，在本公开实施例中，预设支持度阈值设为50％。当然，预设支持度阈值的大小可根据实际情况灵活设定，本公开不限于此。
通过判断，穿衣搭配D的支持度小于预设支持度阈值50％，则不输出穿衣搭配D的语义信息；而防护措施B的支持度和交通状态C的支持度均大于预设支持度阈值50％，则在输出天气节点的语义信息的同时，输出子节点防护措施B和子节点交通状态C的语义信息，例如机器会输出“广州天气晴，紫外线强，请适 当使用防晒霜，华南快线处理拥堵状态”这些信息。另外，如果所有子节点的支持度均不大于预设支持度阈值50％，则只输出天气节点的语义信息，即机器只会输出“广州天气晴”。
作为一种更优的实施方式，步骤S4除了上述方式之外，通过以下方式执行：判断支持度是否大于预设支持度阈值：若支持度大于所述预设支持度阈值，则形成包含所有支持度大于预设支持度阈值的子节点的频繁集，并在输出节点的语义信息的同时，按照支持度从大到小的顺序依次输出频繁集中的每个子节点的语义信息；否则只输出节点的语义信息。
具体地，当支持度与预设支持度阈值大小关系判断完成后，将所有支持度大于预设支持度阈值的子节点形成一个频繁集，例如频繁集包括防护措施B和交通状态C，接下来在输出天气节点的语义信息的同时，按照支持度从大到小的顺序依次输出频繁集中的防护措施B和交通状态C的语义信息。
为了避免输出的语音信息的时间过长，影响用户的交互体验。优选地，在本公开实施例中，通过设定在输出节点的语义信息的同时，只输出频繁集中支持度最高的前三位的子节点的语义信息。
综上所述，应用本公开实施例一提供的用于人机交互的人机多轮对话的方法，机器不需要通过多轮语音对话来进行了解用户意图，提高了用户语音交互的体验。
实施例二
图4为根据本公开实施例二的用于人机交互的人机多轮对话方法的流程图。与实施例一不同的是，实施例二提供的用于人机交互的人机多轮对话方法能够实时地更新知识图谱。如图4所示，该方法包括如下步骤：
步骤S1：建立用户对话行为信息的知识图谱；
步骤S2：根据用户当前输入的信息，在所述知识图谱中确定出与所述当前输入的信息对应的节点和该节点的子节点；
步骤S3：根据所述知识图谱的历史查询记录中所述节点被查询的次数以及所述子节点和所述节点同时被查询的次数，计算所述子节点相对于所述节点的支持度；
步骤S4：通过判断所述支持度与预设支持度阈值的大小关系，确定是否输出 所述子节点的语义信息；以及
步骤S5：根据用户实时输入的信息，实时更新所述知识图谱。
在步骤S1中，建立用户对话行为信息的知识图谱。首先，需要说明的是，在一些实施方式中，该知识图谱是专门只针对于某一用户的，即对于每个用户都建立一个专门只针对于该用户的知识图谱，在另一些实施方式中，也是多名用户共用的一个知识图谱，本公开不限于此。
为了便于理解与说明，下面以一个用户为例进行说明。
图2示意性示出了根据本公开实施例的步骤S1建立的用户对话行为信息的知识图谱。如图2所示，在步骤S1中，针对于某一用户，建立了针对该名用户对话行为信息的知识图谱。建立的知识图谱包括节点、子节点、属性和属性值以及节点与节点之间的关系、节点-属性-属性值之间的关系。
例如，节点包括地点节点、信息节点和使用场景节点。例如，在一些实施方式中，信息节点又分为时间节点、兴趣节点、性别节点、年龄节点和社会特征节点中的至少一种或多种。其中，以使用场景节点为例，它的子节点包括天气节点、音乐节点和故事节点等。在一些实施方式中，天气节点的子节点又包括交通状态节点和防护设施节点等。在本公开实施例中，对于知识图谱中的每个节点和每个子节点，均将频次作为每个节点和每个子节点的一个属性，对应地，将频次值(具体数值，例如5次)作为每个节点和每个子节点的一个属性值。其中，频次值指的是在知识图谱的历史查询记录中该节点或该子节点被查询的次数。
图3为根据本公开实施例的用于人机交互的人机多轮对话方法的步骤S1的流程图。在步骤S1中，在一些实施方式中，知识图谱是在基本图谱模板框架的基础上，由用户前期多次训练而建立。如图3所示，具体过程如下：
步骤S11：建立包含基本节点以及关系的基本图谱模板框架；
步骤S12：采集用户训练时输入的信息，并形成结构化数据；
步骤S13：对所述结构化数据进行实体识别和命名，并在所述知识图谱中形成与所述用户输入的信息对应的新的节点；
步骤S14：从预置的文本语料库中抽取所述新的节点与所述知识图谱中原有的节点之间的关系数据；以及
步骤S15：对所述新的节点与所述原有的节点进行相似度计算和匹配，并结合所述关系数据，以形成节点-关系-节点和节点-属性-属性值的结构。
具体地，在步骤S11中，由开发人员根据大众用户的普遍需求建立包含基本节点以及关系的基本图谱模板框架。例如，目前用户均比较关心天气、交通和新闻等信息，这时，在一些实施方式中，就能把天气、交通和新闻作为基本节点形成在基本图谱模板框架之中。再例如，一般用户都是先询问天气，再询问交通，在一些实施例中，就能把交通状态节点作为天气节点的子节点。在本公开实施例中，将频次作为基本图谱模板框架中节点和子节点的一个属性，对应地，将频次值作为基本图谱模板框架中节点和子节点的一个属性值。
因用户的询问信息可能因人而异，所以步骤S11所建立的基本图谱模板框架很可能无法专门满足某一用户，所以这时就需要对基本图谱模板框架进行训练，以得到专门针对于某一名用户的知识图谱。
具体地，在步骤S12中，采集用户训练时输入的信息，运用自顶向下构建的方式，形成结构化数据。其中，所形成的结构化数据为由二维表结构来逻辑表达的数据，有着严格的格式和长度规范。例如，在一些实施方式中，用户输入的语义信息为“需要打伞吗？”，这时，经过自顶向下构建的方式，所形成的结构化数据为“打伞”。
接下来，执行步骤S13：实体识别。在步骤S13中，对结构化数据进行实体识别和命名，并在知识图谱中形成与用户输入的信息对应的新的节点。具体地，基于步骤S12得到的结构化数据，利用服务器日志进行搜索，搜索出该结构化数据对应的语义特征，以识别该结构化数据，并对其进行命名。比如，对于结构化数据“打伞”，根据服务器日志搜索的结果，识别出“打伞”为动作，并将其命名为“动作，打伞”。然后，在知识图谱中生成新的节点“动作，打伞”。
接下来，执行步骤S14：关系抽取。在步骤S14中，从预置的文本语料库中抽取新的节点与知识图谱中原有的节点之间的关系数据。其中，原有的节点指的是在形成“动作，打伞”这个新的节点之前，知识图谱中已存在的节点。优选地，在该步骤S14中，利用语义规则模式匹配的方法进行关系抽取。
接下来，执行步骤S15：知识融合。本步骤知识融合的目的在于：消除歧义，以便于节点与节点进行融合。具体地，依次将新的节点与原有的节点进行相似度计算，在这过程中，包括：对象对齐、实体匹配和实体同义处理。在一些实施方式中，经过上述处理过程，就能确定新的节点是否能融合到原有的节点之中。例如，经过相似度计算和匹配，确定出新的节点“动作，打伞”和原有的节点“天 气”节点相关，则将新的节点“动作，打伞”与“天气”节点相融合，并且在删除新的节点“动作，打伞”的同时，令“天气”节点的属性频次的属性值(频次值)加1。另外，如果新的节点与原有的节点不相关，则在知识图谱中保留该新的节点，并令新的节点的属性频次的属性值(频次值)加1。然后，结合步骤S14得到的关系数据，形成节点-关系-节点和节点-属性-属性值的结构。在一些实施方式中，这两种结构也被称为实体-关系-实体和实体-属性-属性值的结构。
由此，在基本图谱模板框架的基础上，通过用户前期的训练，得到专门针对该用户对话行为信息的知识图谱。在前期的交互过程中，给予用户相关意图的补充，如用户说“今天天气怎么样？”，机器会问“请问你要查哪个城市的天气”，来补充用户的意图，并将用户对话信息中的意图更新到知识图谱的节点或者节点的属性值之中，以建立该名用户对话行为信息的知识图谱。
在知识图谱建立之后，接下来执行步骤S2至步骤S4。
在步骤S2中，根据用户当前输入的信息，在知识图谱中确定出与当前输入的信息对应的节点和该节点的子节点。如图2所示，以天气节点为例，如果用户问“今天天气怎么样？”，通过语义分析得到用户想要查询“天气”，则在知识图谱中确定出天气节点和天气节点的子节点，并选定天气节点下的属性(例如频次)，遍历子节点如防护措施、交通状态和穿衣搭配等形成集合，并读取出节点和子节点的属性值(频次值)作为项。
接下来，执行步骤S3：根据知识图谱的历史查询记录中节点被查询的次数以及子节点和节点同时被查询的次数，计算子节点相对于节点的支持度。具体地，参见表1。
表1
请求序号 天气A 防护措施B 交通状态C 穿衣搭配D
1 1 0 1 0
2 1 1 1 0
3 1 1 0 1
如表1所示，在知识图谱中，用户每次请求记作一次事务，每次事务记录涉及的节点，其中1为该次事务涉及问答，0为该次事务不涉及的问答。以天气节点A为例，它的子节点包括：防护措施B、交通状态C和穿衣搭配D。例如，在知识图谱的历史查询记录中，天气节点A总共被查询了三次，即频次值为3，而 穿衣搭配D只在请求序号为3时被查询过一次，即穿衣搭配D与天气节点A同时被查询的频次值为1。现在来计算穿衣搭配D相对于天气节点A之间的支持度，可依据以下方式：通过计算子节点和节点同时被查询的次数与节点被查询的次数的比值，得到支持度。即，
求穿衣搭配D相对于天气节点A的支持度(以下简称穿衣搭配D的支持度)为：
支持度P＝A与D同时出现的事务/总的事务＝1/3＝33％，
穿衣搭配D相对于天气节点A的支持度为：P＝P(A&D)/P(A)＝1/3＝33％。
通过上述计算，得到穿衣搭配D的支持度为33％。需要说明的是，A与D同时出现的事务即穿衣搭配D与天气节点A同时被查询(在天气节点A被查询时，穿衣搭配D也被查询，例如表1中请求序号3所示的情况)的频次值，总的事务即天气节点A被查询的频次值。
依照上述同样的方式，依次计算出防护措施B相对于天气节点A的支持度(以下简称防护措施B的支持度)，交通状态C相对于天气节点A的支持度(以下简称交通状态C的支持度)。根据表1所示的内容，经过计算得出防护措施B的支持度为66％，交通状态C的支持度为66％。
在计算完各子节点的支持度之后，接下来，执行步骤S4：通过判断所述支持度与预设支持度阈值的大小关系，以确定所要输出的信息。在步骤S4中，判断支持度是否大于预设支持度阈值：若支持度大于预设支持度阈值，则在输出节点的语义信息的同时，输出子节点的语义信息；否则只输出节点的语义信息。
优选地，在本公开实施例中，预设支持度阈值设为50％。当然，预设支持度阈值的大小可根据实际情况灵活设定，本公开不限于此。
通过判断，穿衣搭配D的支持度小于预设支持度阈值50％，则不输出穿衣搭配D的语义信息；而防护措施B的支持度和交通状态C的支持度均大于预设支持度阈值50％，则在输出天气节点的语义信息的同时，输出子节点防护措施B和子节点交通状态C的语义信息，例如机器会输出“广州天气晴，紫外线强，请适当使用防晒霜，华南快线处理拥堵状态”这些信息。另外，如果所有子节点的支持度均不大于预设支持度阈值50％，则只输出天气节点的语义信息，即机器只会输出“广州天气晴”。
作为一种更优的实施方式，步骤S4除了上述方式之外，还能通过以下方式 执行：判断支持度是否大于预设支持度阈值：若支持度大于所述预设支持度阈值，则形成包含所有支持度大于预设支持度阈值的子节点的频繁集，并在输出节点的语义信息的同时，按照支持度从大到小的顺序依次输出频繁集中的每个子节点的语义信息；否则只输出节点的语义信息。
具体地，当支持度与预设支持度阈值大小关系判断完成后，将所有支持度大于预设支持度阈值的子节点形成一个频繁集，例如频繁集包括防护措施B和交通状态C，接下来在输出天气节点的语义信息的同时，按照支持度从大到小的顺序依次输出频繁集中的防护措施B和交通状态C的语义信息。
为了避免输出的语音信息的时间过长，影响用户的交互体验。优选地，在本公开实施例中，通过设定在输出节点的语义信息的同时，只输出频繁集中支持度最高的前三位的子节点的语义信息。
另外，为了使得建立的知识图谱能够实时更新，实施例二在实施例一的用于人机交互的人机多轮对话方法的基础上，增加了步骤S5：根据用户实时输入的信息，实时更新所述知识图谱。即在知识图谱建立完成后，或者说在用户前期训练完成后，为了使得知识图谱能够随用户的兴趣或询问意图实时更新，提高用户的交互体验，本公开实施例二根据用户实时输入的信息，实时更新知识图谱。
图5为根据本公开实施例二的用于人机交互的人机多轮对话方法的步骤S5的流程图。如图5所示，步骤S5包括：
步骤S51：采集用户实时输入的信息，并形成结构化数据；
步骤S52：对所述结构化数据进行实体识别和命名，并在所述知识图谱中形成与所述用户实时输入的信息对应的新的节点；
步骤S53：从预置的文本语料库中抽取所述新的节点与所述知识图谱中原有的节点之间的关系数据；以及
步骤S54：对所述新的节点与所述原有的节点进行相似度计算和匹配，并结合所述关系数据，以形成节点-关系-节点和节点-属性-属性值的结构。
在步骤S51中，采集用户实时输入的信息，运用自顶向下构建的方式，形成结构化数据。其中，所形成的结构化数据为由二维表结构来逻辑表达的数据，有着严格的格式和长度规范。例如，在一些实施方式中，用户输入的语义信息为“需要带伞吗？”，这时，经过自顶向下构建的方式，所形成的结构化数据为“带伞”。
接下来，执行步骤S52：实体识别。在步骤S52中，对结构化数据进行实体 识别和命名，并在知识图谱中形成与用户输入的信息对应的新的节点。具体地，基于步骤S51得到的结构化数据，利用服务器日志进行搜索，搜索出该结构化数据对应的语义特征，以识别该结构化数据，并对其进行命名。比如，对于结构化数据“带伞”，根据服务器日志搜索的结果，识别出“带伞”为动作，并将其命名为“动作，带伞”。然后，在知识图谱中生成新的节点“动作，带伞”。
接下来，执行步骤S53：关系抽取。在步骤S53中，从预置的文本语料库中抽取新的节点与知识图谱中原有的节点之间的关系数据。其中，原有的节点指的是在形成“动作，带伞”这个新的节点之前，知识图谱中已存在的节点。优选地，在该步骤S53中，利用语义规则模式匹配的方法进行关系抽取。
接下来，执行步骤S54：知识融合。本步骤知识融合的目的在于：消除歧义，以便于节点与节点进行融合。具体地，依次将新的节点与原有的节点进行相似度计算，在这过程中，包括：对象对齐、实体匹配和实体同义处理。在另一些实施方式中，经过上述处理过程，就能确定新的节点是否能融合到原有节点之中。例如，经过相似度计算和匹配，确定出新的节点“动作，带伞”和原有的节点“天气”节点相关，则将新的节点“动作，带伞”与“天气”节点相融合，并且在删除新的节点“动作，带伞”的同时，令“天气”节点的属性频次的属性值(频次值)加1。另外，如果新的节点与原有的节点不相关，则在知识图谱中保留该新的节点，并令新的节点的属性频次的属性值(频次值)加1。然后，结合步骤S53得到的关系数据，形成节点-关系-节点和节点-属性-属性值的结构。在另一些实施方式中，这两种结构也被称为实体-关系-实体和实体-属性-属性值的结构。
综上所述，应用本公开实施例二提供的用于人机交互的人机多轮对话方法，机器不需要通过多轮语音对话来进行了解用户意图，提高了用户语音交互的体验。
此外，应用本公开实施例二提供的用于人机交互的人机多轮对话方法，能够随用户的兴趣或询问意图实时更新知识图谱，提高用户语音交互的体验。
实施例三
对应本公开实施例一和实施例二的用于人机交互的人机多轮对话方法，本公开实施例三提供了一种用于人机交互的人机多轮对话系统。图6示意性示出了根据本公开实施例三的用于人机交互的人机多轮对话系统。如图6所示，该系统600 包括：
知识图谱建立模块601，其被配置为建立用户对话行为信息的知识图谱；
索引模块602，其被配置为：根据用户当前输入的信息，在所述知识图谱中确定出与所述当前输入的信息对应的节点和该节点的子节点；
计算模块603，其被配置为根据所述知识图谱的历史查询记录中所述节点被查询的次数以及所述子节点和所述节点同时被查询的次数，计算所述子节点相对于所述节点的支持度；以及
判断模块604，其被配置为通过判断所述支持度与预设支持度阈值的大小关系，确定是否输出所述子节点的语义信息。
此外，该系统600还包括：
输入模块605，其被配置为接收用户输入的信息；
输出模块606，与判断模块604连接，其被配置为根据判断模块604的判断结果，输出所要输出的信息。
为了便于理解与说明，下面以一个用户为例进行说明。
如图2和图6所示，针对于某一用户，知识图谱建立模块601建立了针对该名用户对话行为信息的知识图谱。建立的知识图谱包括节点、子节点、属性和属性值以及节点与(子)节点之间的关系、节点-属性-属性值之间的关系。
例如，节点包括地点节点、信息节点和使用场景节点。例如，在一些实施方式中，信息节点又分为时间节点、兴趣节点、性别节点、年龄节点和社会特征节点中的至少一种或多种。其中，以使用场景节点为例，它的子节点包括天气节点、音乐节点和故事节点等。在一些实施方式中，天气节点的子节点又包括交通状态节点和防护设施节点等。在本公开实施例中，对于知识图谱中的每个节点和每个子节点，均将频次作为每个节点和每个子节点的一个属性，对应地，将频次值(具体数值，例如5次)作为每个节点和每个子节点的一个属性值。其中，频次值指的是在知识图谱的历史查询记录中该节点或该子节点被查询的次数。
在本公开实施例中，知识图谱建立模块601在基本图谱模板框架的基础上，通过用户前期多次训练而建立知识图谱。具体过程如下：
首先，知识图谱建立模块601根据大众用户的普遍需求建立包含基本节点以及关系的基本图谱模板框架。例如，在一些实施方式中，目前用户均比较关心天气、交通和新闻等信息，这时，就把天气、交通和新闻作为基本节点形成在基本 图谱模板框架之中。再例如，在一些实施方式中，一般用户都是先询问天气，再询问交通，这时，就把交通状态节点作为天气节点的子节点。在本公开实施例中，将频次作为基本图谱模板框架中节点和子节点的一个属性，对应地，将频次值作为基本图谱模板框架中节点和子节点的一个属性值。
因用户的询问信息可能因人而异，所以建立的基本图谱模板框架很可能无法专门满足某一用户，所以这时就需要对基本图谱模板框架进行训练，以得到专门针对于某一名用户的知识图谱。
接下来，知识图谱建立模块601根据输入模块605采集的用户训练时输入的信息，运用自顶向下构建的方式，形成结构化数据。其中，所形成的结构化数据为由二维表结构来逻辑表达的数据，有着严格的格式和长度规范。例如，在一些实施方式中，用户输入的语义信息为“需要打伞吗？”，这时，经过自顶向下构建的方式，所形成的结构化数据为“打伞”。
接下来，知识图谱建立模块601对结构化数据进行实体识别和命名，并在知识图谱中形成与用户输入的信息对应的新的节点。具体地，基于得到的结构化数据，利用服务器日志进行搜索，搜索出该结构化数据对应的语义特征，以识别该结构化数据，并对其进行命名。比如，对于结构化数据“打伞”，根据服务器日志搜索的结果，识别出“打伞”为动作，并将其命名为“动作，打伞”。然后，在知识图谱中生成新的节点“动作，打伞”。
接下来，知识图谱建立模块601从预置的文本语料库中抽取新的节点与知识图谱中原有的节点之间的关系数据。其中，原有的节点指的是在形成“动作，打伞”这个新节点之前，知识图谱中已存在的节点。优选地，知识图谱建立模块601通过运行自身存储的语义规则模式匹配的算法进行关系抽取。
最后，知识图谱建立模块601对新的节点与原有的节点进行知识融合。知识融合的目的在于：消除歧义，以便于节点与节点进行融合。具体地，依次将新的节点与原有的节点进行相似度计算，在这过程中，包括：对象对齐、实体匹配和实体同义处理。经过上述处理过程，在一些实施方式中，就能确定新的节点是否能融合到原有的节点之中。例如，经过相似度计算和匹配，确定出新的节点“动作，打伞”和原有的节点“天气”节点相关，则将新的节点“动作，打伞”与“天气”节点相融合，并且在删除新的节点“动作，打伞”的同时，令“天气”节点的属性频次的属性值(频次值)加1。另外，如果新节点与原有节点不相关，则 在知识图谱中保留该新的节点，并令新的节点的属性频次的属性值(频次值)加1。然后，结合得到的关系数据，形成节点-关系-节点和节点-属性-属性值的结构。在一些实施方式中，这两种结构也被称为实体-关系-实体和实体-属性-属性值的结构。
由此，知识图谱建立模块601在基本图谱模板框架的基础上，通过用户前期的训练，即在前期的交互过程中，给予用户相关意图的补充。在多轮对话中，如用户说“今天天气怎么样？”，机器会问“请问你要查哪个城市的天气”，来补充用户的意图，并将用户对话信息中的意图更新到知识图谱的节点或者节点的属性值之中，以建立该名用户对话行为信息的知识图谱。
在知识图谱建立之后，索引模块602：根据用户当前输入的信息，在知识图谱中确定出与当前输入的信息对应的节点和该节点的子节点。
如图2所示，以天气节点为例，如果用户问“今天天气怎么样？”，索引模块602通过语义分析得到用户想要查询“天气”，随后索引模块602在知识图谱中确定出天气节点和天气节点的子节点，并选定天气节点下的属性(例如频次)，遍历子节点如防护措施、交通状态和穿衣搭配等形成集合，并读取出节点和子节点的属性值(频次值)作为项。
在得到节点、子节点、节点的属性值和子节点的属性值之后，计算模块603：根据知识图谱的历史查询记录中节点被查询的次数以及子节点和节点同时被查询的次数，计算子节点相对于节点的支持度。具体地，参见表1：
表1
请求序号 天气A 防护措施B 交通状态C 穿衣搭配D
1 1 0 1 0
2 1 1 1 0
3 1 1 0 1
如表1所示，在知识图谱中，用户每次请求记作一次事务，每次事务记录涉及的节点，其中1为该次事务涉及问答，0为该次事务不涉及的问答。以天气节点A为例，它的子节点包括：防护措施B、交通状态C和穿衣搭配D。例如，在知识图谱的历史查询记录中，天气节点A总共被查询了三次，即频次值为3，而穿衣搭配D只在请求序号为3时被查询过一次，即穿衣搭配D与天气节点A同时被查询的频次值为1。现在由计算模块603来计算穿衣搭配D相对于天气节点 A之间的支持度，具体地，计算模块603：通过计算子节点和节点同时被查询的次数与节点被查询的次数的比值，得到支持度。即，
求穿衣搭配D相对于天气节点A的支持度(以下简称穿衣搭配D的支持度)为：
支持度P＝A与D同时出现的事务/总的事务＝1/3＝33％，
穿衣搭配D相对于天气节点A的支持度为：P＝P(A&D)/P(A)＝1/3＝33％。
通过上述计算，计算模块603得出穿衣搭配D的支持度为33％。需要说明的是，A与D同时出现的事务即穿衣搭配D与天气节点A同时被查询(在天气节点A被查询时，穿衣搭配D也被查询，例如表1中请求序号3所示的情况)的频次值，总的事务即天气节点A被查询的频次值。
依照上述同样的方式，计算模块603依次计算出防护措施B相对于天气节点A的支持度(以下简称防护措施B的支持度)，交通状态C相对于天气节点A的支持度(以下简称交通状态C的支持度)。根据表1所示的内容，计算模块603经过计算得出防护措施B的支持度为66％，交通状态C的支持度为66％。
在计算完各子节点的支持度之后，接下来，判断模块604：通过判断所述支持度与预设支持度阈值的大小关系，以确定所要输出的信息。判断模块604判断支持度是否大于预设支持度阈值：若支持度大于预设支持度阈值，则在输出节点的语义信息的同时，输出子节点的语义信息；否则只输出节点的语义信息。
优选地，在本公开实施例中，预设支持度阈值设为50％。当然，预设支持度阈值的大小可根据实际情况灵活设定，本公开不限于此。
经过判断模块604的判断，穿衣搭配D的支持度小于预设支持度阈值50％，则不输出穿衣搭配D的语义信息；而防护措施B的支持度和交通状态C的支持度均大于预设支持度阈值50％，则输出模块606在输出天气节点的语义信息的同时，输出子节点防护措施B和子节点交通状态C的语义信息，例如输出模块606会输出“广州天气晴，紫外线强，请适当使用防晒霜，华南快线处理拥堵状态”的信息。另外，如果所有子节点的支持度均不大于预设支持度阈值50％，则输出模块606只输出天气节点的语义信息，例如输出模块606只输出“广州天气晴”的信息。
作为一种更优的实施方式，除了上述方式之外，判断模块604还通过以下方式进行判断：判断每个子节点的支持度是否大于预设支持度阈值：若支持度大于 预设支持度阈值，则形成包含所有支持度大于预设支持度阈值的子节点的频繁集，并在输出节点的语义信息的同时，按照支持度从大到小的顺序依次输出频繁集中的每个子节点的语义信息；否则只输出节点的语义信息。
具体地，当支持度与预设支持度阈值大小关系判断完成后，判断模块604将所有支持度大于预设支持度阈值的子节点形成一个频繁集，例如频繁集包括防护措施B和交通状态C。接下来，输出模块606根据判断模块604的判断结果：在输出天气节点的语义信息的同时，按照支持度从大到小的顺序依次输出频繁集中的防护措施B和交通状态C的语义信息。
为了避免输出的语音信息的时间过长，影响用户的交互体验。优选地，在本公开实施例中，通过设定：输出模块606在输出节点的语义信息的同时，只输出频繁集中支持度最高的前三位的子节点的语义信息。
在知识图谱建立完成后，或者说在用户前期训练完成后，为了使得知识图谱能够随用户的兴趣或询问意图实时更新，提高用户的交互体验，本公开实施例三的系统还包括：知识图谱更新模块607，其被配置为根据用户实时输入的信息，实时更新所述知识图谱。
具体地，知识图谱更新模块607根据输入模块605采集的用户实时输入的信息，运用自顶向下构建的方式，形成结构化数据。其中，所形成的结构化数据为由二维表结构来逻辑表达的数据，有着严格的格式和长度规范。例如，在一些实施方式中，用户输入的语义信息为“需要带伞吗？”，这时，经过自顶向下构建的方式，所形成的结构化数据为“带伞”。
接下来，知识图谱更新模块607对结构化数据进行实体识别和命名，并在知识图谱中形成与用户输入的信息对应的新的节点。具体地，基于得到的结构化数据，利用服务器日志进行搜索，搜索出该结构化数据对应的语义特征，以识别该结构化数据，并对其进行命名。比如，对于结构化数据“带伞”，根据服务器日志搜索的结果，识别出“带伞”为动作，并将其命名为“动作，带伞”。然后，在知识图谱中生成“动作，带伞”的新节点。
接下来，知识图谱更新模块607从预置的文本语料库中抽取新的节点与知识图谱中原有的节点之间的关系数据。其中，原有节点指的是在形成“动作，带伞”这个新节点之前，知识图谱中已存在的节点。优选地，知识图谱更新模块607通过运行自身存储的语义规则模式匹配的算法进行关系抽取。
接下来，知识图谱更新模块607依次将新的节点与原有的节点进行相似度计算，在这过程中，包括：对象对齐、实体匹配和实体同义处理。在一些实施方式中，经过上述处理过程，就能确定新的节点是否能融合到原有的节点之中。例如，经过相似度计算和匹配，确定出“动作，带伞”新节点和原有节点“天气”节点相关，则将“动作，带伞”新节点与“天气”节点相融合，并且在删除“动作，带伞”新节点的同时，令“天气”节点的属性频次的属性值(频次值)加1。另外，如果新节点与原有节点不相关，则在知识图谱中保留该新节点，并令该新节点的属性频次的属性值(频次值)加1。然后，结合得到的关系数据，形成节点-关系-节点和节点-属性-属性值的结构。在一些实施方式中，这两种结构也被称为实体-关系-实体和实体-属性-属性值的结构。
由此，知识图谱更新模块607实时更新知识图谱。
需要说明的是，在一些实施方式中，知识图谱的更新也是由知识图谱建立模块601来完成，当由知识图谱建立模块601实时更新知识图谱时，该系统600不包括知识图谱更新模块607，本公开不限于此。
相应地，本公开实施例还提供了一种智能设备，包括：
处理器；以及
存储器，其上存储有可执行代码，所述可执行代码在被所述处理器执行时，使所述处理器执行上述实施例一或实施例二的用于人机交互的人机多轮对话方法。
综上所述，应用本公开实施例提供的用于人机交互的人机多轮对话方法及系统、智能设备，机器不需要通过多轮语音对话来进行了解用户意图，提高了用户语音交互的体验。
此外，应用本公开实施例提供的用于人机交互的人机多轮对话方法及系统、智能设备，能够随用户的兴趣或询问意图实时更新知识图谱，提高用户语音交互体验。
本领域的技术人员应该明白，上述的本公开的各模块或各步骤可以用通用的计算装置来实现，它们可以集中在单个的计算装置上，或者分布在多个计算装置所组成的网络上，可选地，它们可以用计算装置可执行的程序代码来实现，从而，可以将它们存储在存储装置中由计算装置来执行，或者将它们分别制作成各个集成电路模块，或者将它们中的多个模块或步骤制作成单个集成电路模块来实现。 这样，本公开不限制于任何特定的硬件和软件结合。
虽然本公开所公开的实施方式如上，但所述的内容只是为了便于理解本公开而采用的实施方式，并非用以限定本公开。任何本公开所属技术领域内的技术人员，在不脱离本公开所公开的精神和范围的前提下，可以在实施的形式上及细节上作任何的修改与变化，但本公开的保护范围，仍须以所附的权利要求书所界定的范围为准。

Claims (10)

  1. 一种用于人机交互的人机多轮对话方法，其包括如下步骤：
    步骤S1：建立用户对话行为信息的知识图谱；
    步骤S2：根据用户当前输入的信息，在所述知识图谱中确定出与所述当前输入的信息对应的节点和该节点的子节点；
    步骤S3：根据所述知识图谱的历史查询记录中所述节点被查询的次数以及所述子节点和所述节点同时被查询的次数，计算所述子节点相对于所述节点的支持度；以及
    步骤S4：通过判断所述支持度与预设支持度阈值的大小关系，确定是否输出所述子节点的语义信息。
  2. 根据权利要求1所述的方法，其中，在步骤S3中，
    通过计算所述子节点和所述节点同时被查询的次数与所述节点被查询的次数的比值，得到所述支持度。
  3. 根据权利要求1所述的方法，其中，在步骤S4中，
    判断所述支持度是否大于预设支持度阈值：
    若所述支持度大于所述预设支持度阈值，则在输出所述节点的语义信息的同时，输出所述子节点的语义信息；
    否则只输出所述节点的语义信息。
  4. 根据权利要求1所述的方法，其中，在步骤S4中，
    判断所述支持度是否大于预设支持度阈值：
    若所述支持度大于所述预设支持度阈值，则形成包含所有所述支持度大于所述预设支持度阈值的子节点的频繁集，并在输出所述节点的语义信息的同时，按照所述支持度从大到小的顺序依次输出所述频繁集中的每个子节点的语义信息；
    否则只输出所述节点的语义信息。
  5. 根据权利要求1所述的方法，其中，所述方法还包括：
    步骤S5：根据用户实时输入的信息，实时更新所述知识图谱。
  6. 根据权利要求5所述的方法，其中，步骤S5包括：
    步骤S51：采集用户实时输入的信息，并形成结构化数据；
    步骤S52：对所述结构化数据进行实体识别和命名，并在所述知识图谱中形成与所述用户实时输入的信息对应的新的节点；
    步骤S53：从预置的文本语料库中抽取所述新的节点与所述知识图谱中原有的节点之间的关系数据；以及
    步骤S54：对所述新的节点与所述原有的节点进行相似度计算和匹配，并结合所述关系数据，以形成节点-关系-节点和节点-属性-属性值的结构。
  7. 一种用于人机交互的人机多轮对话系统，所述系统包括：
    知识图谱建立模块，其被配置为建立用户对话行为信息的知识图谱；
    索引模块，其被配置为：根据用户当前输入的信息，在所述知识图谱中确定出与所述当前输入的信息对应的节点和该节点的子节点；
    计算模块，其被配置为：根据所述知识图谱的历史查询记录中所述节点被查询的次数以及所述子节点和所述节点同时被查询的次数，计算所述子节点相对于所述节点的支持度；以及
    判断模块，其被配置为通过判断所述支持度与预设支持度阈值的大小关系，确定是否输出所述子节点的语义信息。
  8. 根据权利要求7所述的系统，其中，根据所述知识图谱的历史查询记录中所述节点被查询的次数以及所述子节点和所述节点同时被查询的次数，计算所述子节点相对于所述节点的支持度，包括：
    通过计算所述子节点和所述节点同时被查询的次数与所述节点被查询的次数的比值，得到所述支持度。
  9. 根据权利要求7所述的系统，其中，所述系统还包括：
    知识图谱更新模块，其被配置为根据用户实时输入的信息，实时更新所述知识图谱。
  10. 一种智能设备，包括：
    处理器；以及
    存储器，其上存储有可执行代码，所述可执行代码在被所述处理器执行时，使所述处理器执行根据权利要求1至6中任一项所述的用于人机交互的人机多轮对话方法。
PCT/CN2020/099423 2019-08-12 2020-06-30 用于人机交互的人机多轮对话方法及系统、智能设备 WO2021027421A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/623,665 US20220253710A1 (en) 2019-08-12 2020-06-30 Human-Machine Multi-Turn Conversation Method and System for Human-Machine Interaction, and Intelligent Apparatus

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910740449.4 2019-08-12
CN201910740449.4A CN110442700A (zh) 2019-08-12 2019-08-12 用于人机交互的人机多轮对话方法及系统、智能设备

Publications (1)

Publication Number Publication Date
WO2021027421A1 true WO2021027421A1 (zh) 2021-02-18

Family

ID=68434753

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/099423 WO2021027421A1 (zh) 2019-08-12 2020-06-30 用于人机交互的人机多轮对话方法及系统、智能设备

Country Status (3)

Country Link
US (1) US20220253710A1 (zh)
CN (1) CN110442700A (zh)
WO (1) WO2021027421A1 (zh)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110442700A (zh) * 2019-08-12 2019-11-12 珠海格力电器股份有限公司 用于人机交互的人机多轮对话方法及系统、智能设备
CN111125372A (zh) * 2019-12-12 2020-05-08 中汇信息技术（上海）有限公司 文本信息发布方法、装置、可读存储介质和电子设备
CN112214607B (zh) * 2020-09-07 2021-06-11 深圳市镜象科技有限公司 基于人工智能的交互方法、心理干预系统、终端和介质
CN112256857B (zh) * 2020-11-17 2023-02-03 北京邮电大学 一种基于图注意力机制的对话历史建模方法
CN112435669B (zh) * 2020-11-19 2024-04-02 福州灵医科技有限公司 一种机器人多轮对话语音交互方法、系统和终端设备
CN112434224A (zh) * 2020-12-08 2021-03-02 神州数码信息系统有限公司 一种基于知识图谱的税收优惠政策推荐方法及其系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105068661A (zh) * 2015-09-07 2015-11-18 百度在线网络技术（北京）有限公司 基于人工智能的人机交互方法和系统
CN107688606A (zh) * 2017-07-26 2018-02-13 北京三快在线科技有限公司 一种推荐信息的获取方法及装置，电子设备
CN108959366A (zh) * 2018-05-21 2018-12-07 宁波薄言信息技术有限公司 一种开放性问答的方法
US20190179917A1 (en) * 2017-12-08 2019-06-13 Apple Inc. Geographical knowledge graph
CN110442700A (zh) * 2019-08-12 2019-11-12 珠海格力电器股份有限公司 用于人机交互的人机多轮对话方法及系统、智能设备

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11436469B2 (en) * 2017-07-31 2022-09-06 Microsoft Technology Licensing, Llc Knowledge graph for conversational semantic search
CN107688614B (zh) * 2017-08-04 2018-08-10 平安科技（深圳）有限公司 意图获取方法、电子装置及计算机可读存储介质
CN108491443B (zh) * 2018-02-13 2021-05-25 上海好体信息科技有限公司 由计算机实施的与用户对话的方法和计算机系统
CN109087132B (zh) * 2018-07-18 2021-07-30 国家电网有限公司 一种基于知识图谱的用户问题推送方法及装置
CN109543019A (zh) * 2018-11-27 2019-03-29 苏州思必驰信息科技有限公司 用于车辆的对话服务方法及装置
CN109766445B (zh) * 2018-12-13 2024-03-26 平安科技（深圳）有限公司 一种知识图谱构建方法及数据处理装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105068661A (zh) * 2015-09-07 2015-11-18 百度在线网络技术（北京）有限公司 基于人工智能的人机交互方法和系统
CN107688606A (zh) * 2017-07-26 2018-02-13 北京三快在线科技有限公司 一种推荐信息的获取方法及装置，电子设备
US20190179917A1 (en) * 2017-12-08 2019-06-13 Apple Inc. Geographical knowledge graph
CN108959366A (zh) * 2018-05-21 2018-12-07 宁波薄言信息技术有限公司 一种开放性问答的方法
CN110442700A (zh) * 2019-08-12 2019-11-12 珠海格力电器股份有限公司 用于人机交互的人机多轮对话方法及系统、智能设备

Also Published As

Publication number Publication date
US20220253710A1 (en) 2022-08-11
CN110442700A (zh) 2019-11-12

Similar Documents

Publication Publication Date Title
WO2021027421A1 (zh) 用于人机交互的人机多轮对话方法及系统、智能设备
WO2020168725A1 (zh) 语音控制方法、装置、存储介质及智能家居系统
WO2021083239A1 (zh) 一种进行图数据查询的方法、装置、设备及存储介质
CN109241266B (zh) 人机交互中基于标准问创建扩展问的方法和装置
JP6229046B2 (ja) 地方なまりを区別する音声データ認識方法、装置及びサーバ
WO2021073298A1 (zh) 一种语音信息的处理方法、装置、智能终端以及存储介质
JP6163607B2 (ja) イベント知識データベースの構築方法および装置
CN106649825B (zh) 语音交互系统及其创建方法和装置
TW201833903A (zh) 對話生成方法、裝置及儲存介質
CN109408811B (zh) 一种数据处理方法及服务器
CN106202476A (zh) 一种基于知识图谱的人机对话的方法及装置
WO2015014122A1 (zh) 语音交互的方法、系统以及交互终端
CN105608199A (zh) 用于智能问答系统中的标准问的扩展方法和装置
US10810056B2 (en) Adding descriptive metadata to application programming interfaces for consumption by an intelligent agent
CN105354180B (zh) 一种实现开放式语义交互服务的方法及系统
WO2018166343A1 (zh) 一种数据融合方法及装置、存储介质以及电子装置
CN105095433A (zh) 实体推荐方法及装置
US20140207716A1 (en) Natural language processing method and system
WO2020233360A1 (zh) 一种产品测评模型的生成方法及设备
CN111145732A (zh) 多任务语音识别后的处理方法及系统
CN110909126A (zh) 一种信息查询方法及装置
CN105335466A (zh) 一种音频数据的检索方法与装置
CN115033661A (zh) 一种基于垂直领域知识图谱的自然语言语义理解方法和装置
CN106055641B (zh) 一种面向智能机器人的人机交互方法及装置
KR102307380B1 (ko) 자연어 처리 기반 콜센터 지원 시스템 및 방법

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20852036

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20852036

Country of ref document: EP

Kind code of ref document: A1