Intelligent outbound mapping knowledge domain construction method and system based on dialogue scene

Abstract

The invention provides an intelligent outbound mapping knowledge domain construction method and system based on a dialogue scene, and relates to the technical field of voice mapping knowledge domains, dialogue audio is converted into a text in real time through a voice recognition API, a syntactic tree is constructed by using a word segmentation tool and syntactic analysis, and part-of-speech and dependency relationships are recognized. Thirdly, inputting the syntactic labeling sequence into a BERT model for entity recognition, and generating an entity set; and establishing a rule set reasoning new relationship, and calculating the entity influence by using a graph theory algorithm. And finally, mapping the dependency relationship into an entity set, generating a hierarchical index of each entity according to the up-and-down hierarchical relationship of the entities and the influence index, and constructing a multi-level knowledge graph through hierarchical mapping.

Description

Technical Field

[0001] The present invention relates to the field of speech knowledge graph technology, and specifically to a method and system for constructing an intelligent outbound call knowledge graph based on a conversation scenario. Background Art

[0002] In intelligent outbound call systems, real-time processing of voice conversations and information extraction are the key to achieving automation and intelligence. Traditional outbound call systems often rely on preset scripts and limited keyword matching, lacking flexibility and in-depth understanding of the content of the conversation. This limitation makes it difficult for the system to dynamically adapt to different conversation scenarios, and it is impossible to effectively extract valuable information from the conversation and update the knowledge base. In addition, traditional methods are often inefficient when dealing with complex relationships and multi-level information, and it is difficult to build a knowledge graph that can fully reflect the content of the conversation.

[0003] In the prior art, publication number CN113051405B discloses a method and device for constructing an intelligent outbound call knowledge graph based on a conversation scenario: obtaining a script file related to the outbound call process, wherein the script file includes multiple conversation topics and multiple conversation contents corresponding to the multiple conversation topics, and the multiple conversation contents include conversation content related to the caller and conversation content related to the called party; according to the script file, determining the connection relationship between the multiple conversation topics; according to the script file, determining the multiple conversation contents; and constructing a knowledge graph related to the outbound call process based on the connection relationship between the determined multiple conversation topics and the multiple conversation contents.

[0004] The main problems with the above method are: based on static script files, it has limited adaptability to dialogue scenarios, and it is difficult to update or adjust the knowledge graph in a timely manner to reflect new information or changes in the dialogue, resulting in insufficient applicability and flexibility of the knowledge graph. In addition, determining the connection relationship between dialogue topics depends on manual judgment. If the judgment is inaccurate, it will directly affect the rationality of the knowledge graph. The system based on preset scripts cannot fully consider the personalized needs and characteristics of each customer, affecting the customer experience.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0006] The purpose of the present invention is to provide a method and system for constructing an intelligent outbound call knowledge graph based on dialogue scenarios to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for constructing an intelligent outbound call knowledge graph based on a conversation scenario, the specific steps include:

[0009] Step 1: Use the speech recognition API to convert the conversation audio into conversation text in real time. Use the word segmentation tool to decompose the conversation text into words and annotate the word parts of speech. Generate a part-of-speech syntactic annotation sequence for the conversation text. Generate a dependency set based on the conversation text with annotated parts of speech and build a syntax tree.

[0010] Step 2: Input the part-of-speech syntactic tag sequence of the conversation text into the BERT model to generate a context embedding vector, and perform entity recognition on the tagged words in the conversation text to generate a set of recognized entities;

[0011] Step 3: Build a rule set, input the entity set into the rule set to infer new relationships, and generate the influence of each entity based on the graph theory algorithm;

[0012] Step 4: Map the dependencies in the conversation text to the entity set, generate the hierarchical relationship of the entity set, generate the hierarchical index of each entity based on the hierarchical level of the entity set and the influence of the entity, and generate a hierarchical index set based on the hierarchical index, construct a hierarchical mapping, and generate a multi-level knowledge graph.

[0013] Furthermore, the principle for constructing the syntax tree is:

[0014] A part-of-speech tag is marked for each word according to its part of speech. In the syntactic parser tool, the syntactic components of each word in the dialogue text are identified, and the syntactic relationship between each word and other words in the sentence is expressed. The dependency relationship between words is extracted, including subject-predicate relationship, verb-object relationship, modification relationship and preposition relationship. After the dependency relationship is confirmed, the dependent words are used as the upper layer, and the words are supplemented layer by layer in the parser to generate a syntactic tree.

[0015] Furthermore, the generated entity set is expressed as follows:

[0016]

[0017] Generate a context embedding vector score for each word in the conversation text through the BERT model. If the score of a word in the conversation text is greater than the entity recognition threshold, integrate the word into the entity set as an entity.

[0018] in, Represents a collection of entities, represents the i-th entity, Represents the index of the entity and I represents the number of entities.

[0019] Furthermore, the principles on which the rule set is constructed are:

[0020] Identify the main purposes of smart outbound calls, including sales conversion, customer feedback collection, and service reminders, and set priorities in descending order;

[0021] Identify key nodes in the conversation, including opening greetings, demand confirmation, and closing, and identify the customer's possible intentions at each key node, including asking for details, expressing interest, and refusing;

[0022] At each interaction node, response instructions are determined based on the customer's likely intent, and a set of "if-then" rules is formulated.

[0023] Furthermore, the principle for generating the influence of each entity is:

[0024]

[0025] in, Representing Entities The influence of Indicates entities, represents the damping coefficient, Indicates direction The collection of other entities, Representing a collection Middle Pointing Entity, express The initial value is the inverse of the number of entities. express The number of pointers to other entities.

[0026] Furthermore, the principle for generating the hierarchical index is:

[0027]

[0028]

[0029] in, represents the average value of the entity's influence, represents the total number of entities, represents the level of the i-th entity, and , M represents the total number of levels, Indicates that the calculation result is rounded up. Represents the level index of the i-th entity.

[0030] Furthermore, the principle for constructing the hierarchical mapping is:

[0031]

[0032] in, The level index is A set of mappings, Indicates the value of the level index, represents an entity selected from a collection of entities, represents a relation selected from the set of dependencies, The level index is The entity set of The level index is The set of dependencies corresponding to the entity;

[0033] All entities and dependencies are assigned to different levels according to the level index. A subgraph is created in each level, which contains all entities and dependencies of the level. All subgraphs are spliced ​​from high to low according to the level index to generate a multi-level graph.

[0034] The present invention also provides a system for constructing an intelligent outbound call knowledge graph based on a conversation scenario, and the system is used to implement the method for constructing an intelligent outbound call knowledge graph based on a conversation scenario, specifically comprising:

[0035] The speech recognition module is used to convert the conversation audio into conversation text in real time through the speech recognition API, decompose the conversation text into words through the word segmentation tool, annotate the word parts of speech, generate the part-of-speech syntactic annotation sequence of the conversation text, generate a dependency set based on the conversation text annotated with parts of speech, and construct a syntax tree;

[0036] The entity recognition module is used to input the part-of-speech syntactic tag sequence of the conversation text into the BERT model to generate a context embedding vector, and perform entity recognition on the tagged words in the conversation text to generate a set of recognized entities;

[0037] The rule building module is used to build a rule set, input the entity set into the rule set to infer new relationships, and generate the influence of each entity based on the graph theory algorithm;

[0038] The graph construction module is used to map the dependencies in the conversation text to the entity set, generate the upper and lower hierarchical relationships of the entity set, generate the hierarchical index of each entity based on the upper and lower levels of the entity set and the size of the entity's influence, and generate a hierarchical index set based on the hierarchical index, construct a hierarchical mapping, and generate a multi-level knowledge graph.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention converts conversation audio into conversation text, ensuring the integrity and voice clarity of conversation information during transmission. The constructed syntactic tree and dependency relationships provide important contextual basis for subsequent entity recognition and relationship reasoning, thereby improving the updating efficiency of the knowledge base and the rationality of the hierarchical structure of the graph. The BERT model's powerful context understanding ability and deep learning technology are used to achieve deep semantic analysis of the conversation text. By generating context embedding vectors, the complex semantic relationships in the conversation text can be effectively understood, thereby improving the accuracy of entity recognition.

[0041] The present invention also combines rule set reasoning and graph theory algorithms to automatically infer potential new relationships from the identified entity set and calculate the influence of each entity. Automatic reasoning and influence analysis enable the system to learn and optimize autonomously, enhance the decision-making ability of the intelligent outbound call system, and improve the intelligence and practicality of the graph. By mapping the dependencies in the conversation text to the entity set, the hierarchical relationship of the entities can be clearly constructed, and the influence of the entities can be comprehensively considered to create a multi-level knowledge graph. By combining the hierarchical relationship and influence index of the entities, the importance of the entities in the entire system can be more comprehensively evaluated, thereby improving the practicality of the graph. Based on the dynamic adjustment capability of the hierarchical index, the graph can be quickly updated when the information or relationship changes to maintain its timeliness, so that the system can capture deep information in complex conversations and provide strong support for further business decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a schematic diagram of a method flow chart of an embodiment of the present invention;

[0043] Figure 2 Schematic diagram of system modules according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0045] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0046] Example:

[0047] See also Figure 1 , the present invention provides a technical solution:

[0048] A method for constructing an intelligent outbound call knowledge graph based on a conversation scenario, the specific steps include:

[0049] Step 1: Use the speech recognition API to convert the conversation audio into conversation text in real time. Use the word segmentation tool to decompose the conversation text into words and annotate the word parts of speech. Generate a part-of-speech syntactic annotation sequence for the conversation text. Generate a dependency set based on the conversation text with annotated parts of speech and build a syntax tree.

[0050] In this embodiment, the principle for constructing the syntax tree is:

[0051] A part-of-speech tag is marked for each word according to its part of speech. In the syntactic parser tool, the syntactic components of each word in the dialogue text are identified, and the syntactic relationship between each word and other words in the sentence is expressed. The dependency relationship between words is extracted, including subject-predicate relationship, verb-object relationship, modification relationship and preposition relationship. After the dependency relationship is confirmed, the dependent words are used as the upper layer, and the words are supplemented layer by layer in the parser to generate a syntactic tree.

[0052] The conversation audio is converted into text in real time through Google Speech-to-Text, and the text is decomposed into independent words through the natural language processing tool SpaCy. A corresponding grammatical category, such as name, verb, adjective, etc., is assigned to each word to generate a dependency graph. The nodes of the graph are words, and the edges of the graph represent the dependency relationship between words. The dependency graph is converted into a tree structure, which is a syntactic tree. The root node of the syntactic tree is the core verb or predicate word of the sentence, and other words are connected in sequence as child nodes to form a tree hierarchy.

[0053] Step 2: Input the part-of-speech syntactic tag sequence of the conversation text into the BERT model to generate a context embedding vector, and perform entity recognition on the tagged words in the conversation text to generate a set of recognized entities;

[0054] The entity set generated in this embodiment is expressed as follows:

[0055]

[0056] Generate a context embedding vector score for each word in the conversation text through the BERT model. If the score of a word in the conversation text is greater than the entity recognition threshold, integrate the word into the entity set as an entity.

[0057] in, Represents a collection of entities, represents the i-th entity, Represents the index of the entity and I represents the number of entities.

[0058] Only when the context embedding vector score is greater than the entity recognition threshold When , the word is recognized as an entity; Generate entity recognition threshold The principles are:

[0059] Compute the mean and standard deviation of the context embedding scores for each word in the conversation text:

[0060]

[0061]

[0062]

[0063] in, represents the average of the context embedding vector scores, represents the index of the context embedding vector score, represents the number of context embedding vector scores, Indicates context embedding vector scores, represents the standard deviation of the context embedding vector score, Represents the entity recognition threshold.

[0064] Step 3: Build a rule set, input the entity set into the rule set to infer new relationships, and generate the influence of each entity based on the graph theory algorithm;

[0065] In this embodiment, the principle for constructing the rule set is:

[0066] Identify the main purposes of smart outbound calls, including sales conversion, customer feedback collection, and service reminders, and set priorities in descending order;

[0067] Identify key nodes in the conversation, including opening greetings, demand confirmation, and closing, and identify the customer's possible intentions at each key node, including asking for details, expressing interest, and refusing;

[0068] At each interaction node, response instructions are determined based on the customer's likely intent, and a set of "if-then" rules is formulated.

[0069] Define clear goals, identify the types of outbound calls, including sales, surveys, reminders, etc., divide goals into short-term goals and long-term goals, and set evaluation criteria such as sales volume, survey completion rate, customer satisfaction, etc., assign weights to different goals to guide resource allocation; integrate multi-channel data including call records, CRM system and customer feedback to ensure data consistency; create customer interaction flow charts, identify key nodes in the conversation, including opening greetings, demand confirmation, and ending, etc. The main needs include product introduction and question answering, identify customer reactions and potential needs at each node, establish a customer intent library covering common problems and needs based on historical data, and establish an "if + condition, then + action" rule library based on the customer intent library. Set dynamic priorities for the rule set based on the completion of the evaluation criteria of the business goals. Conversation categories with higher dynamic priorities are executed first. The formula for setting dynamic priorities is:

[0070]

[0071] in, Indicates the dynamic priority of the target, Indicates the completion criteria of the goal. Indicates the actual achievement of the goal;

[0072] The more the actual completion amount of the target, the lower the dynamic priority, the later the execution will be, and other targets with less actual completion will be executed first.

[0073] The influence of each entity is generated based on the following principles:

[0074]

[0075] in, Representing Entities The influence of Indicates entities, represents the damping coefficient, Indicates direction The collection of other entities, Representing a collection Middle Pointing Entity, express The initial value is the inverse of the number of entities. express The number of pointers to other entities.

[0076] In graph theory algorithms, each entity corresponds to a node. Reflects the entity node The importance of a node in the entire link structure is measured The degree to which a node is referenced or pointed to by other nodes and the importance of the nodes themselves. If a node is pointed to by multiple important nodes, the nodes pointed to are also important nodes with relatively high influence. The damping coefficient , used to simulate the probability of randomly jumping to other entity nodes;

[0077] Step 4: Map the dependencies in the conversation text to the entity set, generate the hierarchical relationship of the entity set, generate the hierarchical index of each entity based on the hierarchical relationship of the entity set and the influence of the entity, generate the hierarchical set based on the hierarchical index, construct the hierarchical mapping, and generate a multi-level knowledge graph.

[0078] The principle on which the hierarchical index is generated is:

[0079]

[0080]

[0081] in, represents the average value of the entity's influence, represents the total number of entities, represents the level of the i-th entity, and , M represents the total number of levels, Indicates that the calculation result is rounded up. Represents the level index of the i-th entity.

[0082] The hierarchy index comprehensively reflects the level and importance of the entity. The higher the level of the entity, the higher its influence, which means the more important the entity is and the higher the corresponding hierarchy index is. The hierarchy index is proportional to the level and influence of the entity. In the upper and lower hierarchical relationship, the top entity corresponds to level M, and the bottom entity corresponds to level 1.

[0083] The principles for constructing hierarchical mapping are:

[0084]

[0085] in, The level index is A set of mappings, Indicates the value of the level index, represents an entity selected from a collection of entities, represents a relation selected from the set of dependencies, The level index is The entity set of The level index is The dependency set corresponding to the entity.

[0086] In this embodiment, the hierarchy represents the up-down relationship between entities determined based on the dependency relationship, and the level represents the up-down relationship determined based on the level index of the entity;

[0087] The purpose of constructing a hierarchical mapping is to assign entities with the same hierarchical index to the same level. The higher the hierarchical index of an entity, the higher the corresponding influence and the more important the entity. Entities with higher hierarchical indices are placed at higher levels.

[0088] All entities and dependencies are assigned to different levels according to the level index. A subgraph is created in each level, which contains all entities and dependencies of the level. All subgraphs are spliced ​​from high to low according to the level index to generate a multi-level graph.

[0089] See also Figure 2 The present invention also provides a system for constructing an intelligent outbound call knowledge graph based on a conversation scenario, and the system is used to implement the method for constructing an intelligent outbound call knowledge graph based on a conversation scenario, specifically comprising:

[0090] The speech recognition module is used to convert the conversation audio into conversation text in real time through the speech recognition API, decompose the conversation text into words through the word segmentation tool, annotate the word parts of speech, generate the part-of-speech syntactic annotation sequence of the conversation text, generate a dependency set based on the conversation text annotated with parts of speech, and construct a syntax tree;

[0091] The entity recognition module is used to input the part-of-speech syntactic tag sequence of the conversation text into the BERT model to generate a context embedding vector, and perform entity recognition on the tagged words in the conversation text to generate a set of recognized entities;

[0092] The rule building module is used to build a rule set, input the entity set into the rule set to infer new relationships, and generate the influence of each entity based on the graph theory algorithm;

[0093] The graph construction module is used to map the dependencies in the conversation text to the entity set, generate the upper and lower hierarchical relationships of the entity set, generate the hierarchical index of each entity based on the upper and lower levels of the entity set and the size of the entity's influence, and generate a hierarchical index set based on the hierarchical index, construct a hierarchical mapping, and generate a multi-level knowledge graph.

[0094] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0095] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0096] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0097] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be covered by the protection scope of the present application.

Claims

1. A method for constructing an intelligent outbound call knowledge graph based on a conversation scenario, characterized in that: The specific steps include: Step 1: Use the speech recognition API to convert the conversation audio into conversation text in real time. Use the word segmentation tool to decompose the conversation text into words and annotate the word parts of speech. Generate a part-of-speech syntactic annotation sequence for the conversation text. Generate a dependency set based on the conversation text with annotated parts of speech and build a syntax tree. Step 2: Input the part-of-speech syntactic tag sequence of the conversation text into the BERT model to generate a context embedding vector, and perform entity recognition on the tagged words in the conversation text to generate a set of recognized entities; Step 3: Build a rule set, input the entity set into the rule set to infer new relationships, and generate the influence of each entity based on the graph theory algorithm; Step 4: Map the dependencies in the conversation text to the entity set, generate the hierarchical relationship of the entity set, generate the hierarchical index of each entity based on the hierarchical level of the entity set and the influence of the entity, and generate a hierarchical index set based on the hierarchical index, construct a hierarchical mapping, and generate a multi-level knowledge graph.

2. According to claim 1, a method for constructing an intelligent outbound call knowledge graph based on a conversation scenario is characterized in that: The principle for constructing the syntax tree in step 1 is: A part-of-speech tag is marked for each word according to its part of speech. In the syntactic parser tool, the syntactic components of each word in the dialogue text are identified, and the syntactic relationship between each word and other words in the sentence is expressed. The dependency relationship between words is extracted, including subject-predicate relationship, verb-object relationship, modification relationship and preposition relationship. After the dependency relationship is confirmed, the dependent words are used as the upper layer, and the words are supplemented layer by layer in the parser to generate a syntactic tree.

3. According to claim 1, a method for constructing an intelligent outbound call knowledge graph based on a conversation scenario is characterized in that: The entity set generated in step 2 is expressed as follows: Generate a context embedding vector score for each word in the conversation text through the BERT model. If the score of a word in the conversation text is greater than the entity recognition threshold, integrate the word into the entity set as an entity. in, Represents a collection of entities, represents the i-th entity, represents the index of the entity and I represents the number of entities.

4. According to claim 1, a method for constructing an intelligent outbound call knowledge graph based on a conversation scenario is characterized in that: The principle for constructing the rule set in step 3 is: Identify the main purposes of smart outbound calls, including sales conversion, customer feedback collection, and service reminders, and set priorities in descending order; Identify key nodes in the conversation, including opening greetings, demand confirmation, and closing, and identify the customer's possible intentions at each key node, including asking for details, expressing interest, and refusing; At each interaction node, response instructions are determined based on the customer's likely intent, and a set of "if-then" rules is developed.

5. According to claim 1, a method for constructing an intelligent outbound call knowledge graph based on a conversation scenario is characterized in that: The principle for generating the influence of each entity in step 3 is: in, Representing Entities The influence of Indicates entities, represents the damping coefficient, Indicates direction The collection of other entities, Representing a collection Middle Pointing Entity, express The initial value is the inverse of the number of entities. express The number of pointers to other entities.

6. According to claim 5, a method for constructing an intelligent outbound call knowledge graph based on a conversation scenario is characterized in that: The principle for generating the hierarchical index in step 4 is: in, represents the average value of the entity's influence, represents the total number of entities, represents the level of the i-th entity, and , M represents the total number of levels, Indicates that the calculation result is rounded up. Represents the level index of the i-th entity.

7. According to claim 1, a method for constructing an intelligent outbound call knowledge graph based on a conversation scenario is characterized in that: The principle for constructing the hierarchical mapping in step 4 is: in, The level index is A set of mappings, Indicates the value of the level index, represents an entity selected from a collection of entities, represents a relation selected from the set of dependencies, The level index is The entity set of The level index is The dependency set corresponding to the entity; All entities and dependencies are assigned to different levels according to the level index. A subgraph is created in each level, which contains all entities and dependencies of the level. All subgraphs are spliced ​​from high to low according to the level index to generate a multi-level graph.

8. An intelligent outbound call knowledge graph construction system based on conversation scenarios, characterized by: The system is used to implement the method for constructing an intelligent outbound call knowledge graph based on a conversation scenario as described in any one of claims 1 to 7, specifically comprising: The speech recognition module is used to convert the conversation audio into conversation text in real time through the speech recognition API, decompose the conversation text into words through the word segmentation tool, annotate the word parts of speech, generate the part-of-speech syntactic annotation sequence of the conversation text, generate a dependency set based on the conversation text annotated with parts of speech, and construct a syntax tree; The entity recognition module is used to input the part-of-speech syntactic tag sequence of the conversation text into the BERT model to generate a context embedding vector, and perform entity recognition on the tagged words in the conversation text to generate a set of recognized entities; The rule building module is used to build a rule set, input the entity set into the rule set to infer new relationships, and generate the influence of each entity based on the graph theory algorithm; The graph construction module is used to map the dependencies in the conversation text to the entity set, generate the upper and lower hierarchical relationships of the entity set, generate the hierarchical index of each entity based on the upper and lower levels of the entity set and the size of the entity's influence, and generate a hierarchical index set based on the hierarchical index, construct a hierarchical mapping, and generate a multi-level knowledge graph.

Images