Intelligent insurance customer service dialogue generation method and system based on big model

Abstract

The present application provides a method and system for generating intelligent insurance customer service dialogues based on a large model, which effectively improves the depth and breadth of dialogue understanding by accurately acquiring and processing the local and global semantic feature spaces of the target user's insurance consultation dialogue. By summarizing and encoding the global semantic features and combining them with enhanced processing of the local semantic features, a comprehensive and detailed grasp of the dialogue content is achieved. The local and global semantic coding data are integrated and calculated to generate accurate target semantic coding, laying a solid foundation for subsequent knowledge matching. Furthermore, by calculating the feature distance between the target semantic coding and the knowledge fragments in the reference insurance knowledge base, the present invention can efficiently identify and select knowledge fragments that are highly relevant to the target dialogue, thereby generating accurate and appropriate reply statements, thereby significantly improving the intelligence level of insurance customer service dialogues, improving the accuracy and satisfaction of replies, and optimizing user experience.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a method and system for generating intelligent insurance customer service dialogues based on a large model. Background Art

[0002] In the insurance industry, the efficiency and accuracy of customer service conversations are crucial to improving customer satisfaction and business processing efficiency. Traditional insurance customer service systems often rely on manual customer service or rule-based methods to handle user inquiries. However, this approach has problems such as slow response speed, limited knowledge coverage, and difficulty in dealing with complex and changing user issues.

[0003] With the rapid development of artificial intelligence technology, especially the continuous advancement of natural language processing (NLP) technology, intelligent dialogue systems based on large models have gradually become an effective means to solve the above problems. Large models can capture rich language features, understand complex semantic relationships, and demonstrate strong dialogue generation capabilities in multiple fields. However, in the professional field of insurance customer service, how to effectively use large models to generate accurate and appropriate response sentences still faces many challenges.

[0004] On the one hand, insurance consultation dialogues often involve semantic information at multiple levels, including specific insurance terms, claims settlement procedures, product features, etc. This requires the dialogue generation system to fully understand and grasp the local and global semantic features of the dialogue. On the other hand, the professionalism and complexity of insurance knowledge also place higher demands on the knowledge matching and reasoning capabilities of the dialogue generation system. Summary of the invention

[0005] In view of this, the purpose of this application is to provide a method and system for generating intelligent insurance customer service dialogue based on a large model.

[0006] According to a first aspect of the present application, a method for generating an intelligent insurance customer service dialogue based on a large model is provided, the method comprising:

[0007] Obtaining the local semantic feature space and the global semantic feature space of the target insurance consultation dialogue of the target user;

[0008] Performing semantic induction encoding on the global semantic feature space to generate global semantic encoding data, and performing semantic enhancement operation on the local semantic feature space and the global semantic feature space to generate local semantic encoding data;

[0009] The local semantic coding data and the global semantic coding data are fused and calculated to generate target semantic coding data;

[0010] Determine a first characteristic distance of each reference knowledge fragment in the reference insurance knowledge base based on the target semantic coding data and the knowledge semantic coding data of the reference insurance knowledge base; the first characteristic distance represents the knowledge relevance between the reference knowledge fragment in the reference insurance knowledge base and the target insurance consultation dialogue;

[0011] In the reference insurance knowledge base, a target knowledge segment is selected according to the first feature distance, and a reply statement result of the target insurance consultation dialogue is generated based on the target knowledge segment.

[0012] For example, in a possible implementation of the first aspect, the step of generating a reply statement result of the target insurance consultation dialogue based on the target knowledge segment includes:

[0013] Identify key information elements and associated information elements contained in the target knowledge segment; wherein the key information elements represent the core semantic content in the target knowledge segment for answering the target insurance consultation dialogue, and the associated information elements represent the auxiliary semantic content associated with the key information elements;

[0014] Constructing a preliminary framework of a reply statement based on the key information element set; the preliminary framework includes at least one key information placeholder for inserting specific key information content;

[0015] Based on the semantic similarity between the associated information elements and the key information elements, the importance of the associated information elements in the insurance field, and the historical interests of the target user, the associated information element set is screened and sorted to select the associated information elements that match the target insurance consultation dialogue, and generate a screened and sorted sequence of associated information elements;

[0016] Inserting the associated information elements in the filtered and sorted associated information element sequence into corresponding positions in the preliminary framework of the reply statement one by one, generating an intermediate version of the reply statement containing the associated information elements;

[0017] Performing syntax verification and semantic verification on the intermediate version of the reply statement to generate a verified optimized version of the reply statement;

[0018] According to the dialogue context characteristics of the target insurance consultation dialogue and the user portrait characteristics of the target user, the optimized version of the reply statement is personalized adjusted, and the final version of the reply statement after the personalized adjustment is output as the reply statement result of the target insurance consultation dialogue, wherein the personalized adjustment includes adjusting the language style and adding emotional elements.

[0019] According to the second aspect of the present application, an insurance service system is provided, which includes a machine-readable storage medium and a processor, wherein the machine-readable storage medium stores machine-executable instructions, and when the processor executes the machine-executable instructions, the insurance service system implements the aforementioned large-model-based intelligent insurance customer service dialogue generation method.

[0020] According to the third aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed, the aforementioned large-model-based intelligent insurance customer service dialogue generation method is implemented.

[0021] According to any one of the above aspects, the technical effect of the present application is:

[0022] The embodiments of the present application effectively improve the depth and breadth of dialogue understanding by accurately acquiring and processing the local and global semantic feature spaces of the target user's insurance consultation dialogue. By summarizing and encoding the global semantic features and combining them with enhanced processing of the local semantic features, a comprehensive and detailed grasp of the dialogue content is achieved. The local and global semantic coding data are integrated and calculated to generate accurate target semantic coding, laying a solid foundation for subsequent knowledge matching. Furthermore, by calculating the feature distance between the target semantic coding and the knowledge fragments in the reference insurance knowledge base, the present invention can efficiently identify and select knowledge fragments that are highly relevant to the target dialogue, thereby generating accurate and appropriate reply statements, thereby significantly improving the intelligence level of insurance customer service dialogues, improving the accuracy and satisfaction of replies, and optimizing user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 A schematic diagram of the process of generating a smart insurance customer service dialogue based on a large model provided in an embodiment of the present application is shown;

[0025] Figure 2 A schematic diagram of the component structure of an insurance service system provided in an embodiment of the present application for implementing the above-mentioned large model-based intelligent insurance customer service dialogue generation method is shown. DETAILED DESCRIPTION

[0026] The embodiments of the present application are described below in conjunction with the drawings in the present application. It should be understood that the implementation methods described below in conjunction with the drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions of the embodiments of the present application.

[0027] It will be understood by those skilled in the art that, unless specifically stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the terms "including" and "comprising" used in the embodiments of the present application refer to that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements and / or components, but do not exclude the implementation as other features, information, data, steps, operations, elements, components and / or combinations thereof supported by the technical field. It should be understood that when an element is "connected" or "coupled" to another element, the element may be directly connected or coupled to the other element, or may refer to the element and the other element establishing a connection relationship through an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling, and the term "and / or" used herein indicates at least one of the items defined by the term, such as "A and / or B" may be implemented as "A", or as "B", or as "A and B".

[0028] In order to make the purpose, technical solution and advantages of the present application clearer, the implementation mode of the present application will be further described in detail with reference to the accompanying drawings. The technical solution of the embodiment of the present application and the technical effect produced by the technical solution of the present application are explained below by describing several exemplary implementation modes. It should be pointed out that the following implementation modes can refer to, draw lessons from or combine with each other, and the same terms, similar features and similar implementation steps in different implementation modes are not described repeatedly.

[0029] Figure 1 The flowchart of the method and system for generating a smart insurance customer service dialogue based on a big model provided by the embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps in the method for generating a smart insurance customer service dialogue based on a big model of this embodiment can be shared with each other according to actual needs, or some steps can be omitted or maintained. The detailed steps of the method for generating a smart insurance customer service dialogue based on a big model include:

[0030] Step S110, obtaining the local semantic feature space and the global semantic feature space of the target insurance consultation dialogue of the target user.

[0031] In this embodiment, the server receives a target insurance consultation dialogue from a target user, the content of which is "I want to know about the major disease insurance. If I am diagnosed with cancer, what is the insurance compensation process? Also, is the compensation ratio of this insurance the same for different types of cancer?"

[0032] First, multiple semantic features corresponding to the target insurance consultation dialogue are obtained. These semantic features may include the semantic vectors of each word, such as the semantic vectors corresponding to words such as "major disease insurance", "cancer", "compensation process", and "compensation ratio".

[0033] When determining the semantic complexity of each semantic vector in multiple semantic features, the server operates as follows. For example, for the semantic vector "critical disease insurance", the server determines the number of associated concepts corresponding to this semantic vector when it corresponds to different vector values ​​in all semantic features of the target insurance consultation dialogue. Assume that when the vector value of "critical disease insurance" is a specific value, the concepts associated with it include "disease type", "insurance terms", "insured population" and other concepts, and the number of associated concepts is 3. Then the server counts the number of global semantic concepts of the target insurance consultation dialogue, assuming that a total of 10 different semantic concepts are involved in the entire dialogue. Based on the number of associated concepts and the number of global semantic concepts, the first associated parameters of the semantic vector "critical disease insurance" corresponding to different vector values ​​are determined respectively. For example, the calculated first associated parameter is 3 / 10 = 0.3. Then, based on this associated parameter corresponding to different vector values ​​and the corresponding number of associated concepts, its semantic complexity is determined. Here, it is assumed that the semantic complexity of the semantic vector "critical disease insurance" is 0.6 (this is just a hypothetical calculation result) obtained by a certain specific calculation method (such as multiplying by a preset weight factor, etc.).

[0034] After the server determines the semantic complexity of each semantic vector in the same way, it clusters the semantic feature space of each semantic vector in multiple semantic features based on the semantic complexity. For example, semantic vectors with higher semantic complexity, such as "cancer" and "payment process", may be clustered into the local semantic feature space because they are the more core and semantically closely related parts of this consultation dialogue. Some semantic vectors with relatively low semantic complexity but related to the entire insurance topic, such as some common insurance terms, may be clustered into the global semantic feature space. Through such clustering operations, the server successfully generated the local semantic feature space and the global semantic feature space of the target insurance consultation dialogue.

[0035] Another possible scenario is that the server determines the local-global semantic range ratio. For example, the local semantic range ratio is set to 0.6 and the global semantic range ratio is set to 0.4. Then, the semantic feature space clustering is performed on each semantic vector in multiple semantic features based on this ratio. Semantic vectors are divided into local semantic feature space and global semantic feature space according to this ratio based on factors such as their importance or frequency of occurrence in the conversation. For example, semantic vectors that often appear together and have close semantic associations, such as "cancer" and "compensation ratio", may be divided into local semantic feature space according to the ratio; while some semantic vectors that appear alone and have relatively loose semantic associations may be divided into global semantic feature space.

[0036] Another case is based on the semantic feature space clustering instruction. The server receives a pre-set semantic feature space clustering instruction, which may be formulated based on a large amount of insurance consultation dialogue data and related analysis results. The server then clusters the semantic vectors in multiple semantic features according to this instruction and outputs them into local semantic feature space and global semantic feature space. For example, the instruction stipulates that the semantic vectors related to the name of a specific disease are clustered into the local semantic feature space, and the semantic vectors related to the general concept of insurance are clustered into the global semantic feature space. The server performs clustering operations according to this instruction.

[0037] Step S120, performing semantic induction encoding on the global semantic feature space to generate global semantic encoding data, and performing semantic enhancement operation on the local semantic feature space and the global semantic feature space to generate local semantic encoding data.

[0038] For the global semantic feature space, the server first fuses the global semantic feature spaces belonging to the same target insurance consultation dialogue. For example, the global semantic feature space obtained by clustering previously contains multiple semantic vectors, and the server fuses these semantic vectors according to certain rules. Assuming that there are semantic vectors A, B, C, etc. in the global semantic feature space, the fusion method may be to add the corresponding elements of these semantic vectors or add them in a certain weighted manner. Through this fusion operation, the first fused semantic feature space of each target insurance consultation dialogue is generated.

[0039] Then, the server loads each first fused semantic feature space into the global semantic sub-model of the customer service dialogue model. This global semantic sub-model contains multi-level semantic coding nodes. For example, this sub-model has three layers of coding nodes. The first layer of nodes may preliminarily encode the basic semantic information in the first fused semantic feature space, the second layer of nodes further abstract and encode the first layer of encoding results, and the third layer of nodes perform higher-level encoding representation. The first fused semantic feature space is encoded and represented through these multi-level semantic coding nodes, and finally the global semantic coding data is generated.

[0040] For the semantic enhancement operation of the local semantic feature space and the global semantic feature space, the server first fuses the local semantic feature space and the global semantic feature space to generate a second fused semantic feature space. Assuming that there are semantic vectors X and Y in the local semantic feature space and semantic vectors A and B in the global semantic feature space (this is just a simple example), the fusion method can be to combine these semantic vectors together to form a new semantic feature space structure. Then, this second fused semantic feature space is loaded into the local semantic sub-model of the customer service dialogue model. This local semantic sub-model also has multi-level semantic encoding nodes. For example, the first-level nodes may focus on encoding the local semantic feature part in the second fused semantic feature space to highlight the characteristics of local semantics; the second-level nodes comprehensively consider the relationship between local and global semantics for encoding; and the third-level nodes optimize the encoding of the whole. The second fused semantic feature space is encoded and represented by these multi-level semantic encoding nodes to generate local semantic encoding data.

[0041] Step S130, performing fusion calculation on the local semantic coding data and the global semantic coding data to generate target semantic coding data.

[0042] In this embodiment, the server first determines the vector length or semantic content of the semantic mapping feature of the target insurance consultation dialogue. Assume that the server determines the vector length of the semantic mapping feature. For example, after analysis and calculation, the length of the semantic mapping feature vector corresponding to the local semantic encoding data is 10, and the length of the semantic mapping feature vector corresponding to the global semantic encoding data is 15.

[0043] Then, based on the local semantic feature space, the global semantic feature space and the vector length, the influence coefficients corresponding to the local semantic coding data and the global semantic coding data are determined. Assume that according to a certain preset calculation rule, the influence coefficient corresponding to the local semantic coding data is 0.4, and the influence coefficient corresponding to the global semantic coding data is 0.6 (this coefficient is obtained based on the vector length and the preset calculation method).

[0044] Finally, based on the influence coefficient, the local semantic coding data and the global semantic coding data are fused and calculated. For example, assuming that the local semantic coding data is L and the global semantic coding data is G, the fusion calculation method can be performed in a weighted average manner, that is, the target semantic coding data = 0.4 * L + 0.6 * G. Through such calculation, the server generates the target semantic coding data.

[0045] Step S140, determining a first characteristic distance of each reference knowledge fragment in the reference insurance knowledge base based on the target semantic encoding data and the knowledge semantic encoding data of the reference insurance knowledge base. The first characteristic distance represents the knowledge relevance between the reference knowledge fragment in the reference insurance knowledge base and the target insurance consultation dialogue.

[0046] In this embodiment, the server pre-configures a reference insurance knowledge base, which contains a large number of reference knowledge fragments, each of which has corresponding knowledge semantic coding data. For example, a knowledge fragment in the reference insurance knowledge base is a detailed description of the major disease insurance compensation details, and its knowledge semantic coding data is K1.

[0047] The server compares and calculates the target semantic coding data with the knowledge semantic coding data of each reference knowledge fragment in the reference insurance knowledge base. The method of calculating the first feature distance may be based on a certain distance measurement algorithm, such as the Euclidean distance algorithm. Assuming that the target semantic coding data is T, for the knowledge semantic coding data K1 of the knowledge fragment, calculate the Euclidean distance between the two. Assuming T = (t1, t2, t3), K1 = (k1, k2, k3), then the first feature distance d1 = √((t1 - k1)²+(t2 - k2)²+(t3 - k3)²). In the same way, the server calculates the first feature distance between each reference knowledge fragment in the reference insurance knowledge base and the target insurance consultation dialogue. The smaller the value of this first feature distance, the higher the knowledge correlation between the reference knowledge fragment and the target insurance consultation dialogue.

[0048] Step S150, in the reference insurance knowledge base, select a target knowledge segment according to the first feature distance, and generate a reply statement result of the target insurance consultation dialogue based on the target knowledge segment.

[0049] In this embodiment, the target knowledge fragment is selected from the reference insurance knowledge base according to the calculated first characteristic distance. For example, there are multiple knowledge fragments about different insurance compensation situations in the reference insurance knowledge base. According to the calculation result of the first characteristic distance, it is found that one of the knowledge fragments about cancer compensation in major disease insurance has the smallest first characteristic distance with the target insurance consultation dialogue, and then this knowledge fragment is selected as the target knowledge fragment.

[0050] Identify the key information elements and associated information elements contained in the target knowledge fragment. For example, in this target knowledge fragment, the key information elements may be "the specific process of cancer compensation", "the calculation method of compensation ratio for different types of cancer", etc. These are the core semantic content that directly answers the target insurance consultation dialogue. The associated information elements may be "the update history of insurance terms related to cancer compensation", "reference cases of compensation for other similar diseases", etc. These are auxiliary semantic content associated with the key information elements.

[0051] Based on the set of key information elements, construct a preliminary framework for the response statement. For example, the preliminary framework may be "Regarding the question you asked about cancer compensation in major disease insurance, the compensation process is as follows: [compensation process key information placeholder], and the calculation method of the compensation ratio for different types of cancer is: [compensation ratio key information placeholder]".

[0052] Based on the semantic similarity between the associated information elements and the key information elements, the importance of the associated information elements in the insurance field, and the historical interests of the target user, the associated information element set is screened and sorted to select the associated information elements that match the target insurance consultation dialogue, and generate a screened and sorted sequence of associated information elements. For example, if the target user was previously interested in insurance clause updates, then the importance of the associated information element "insurance clause update history related to cancer compensation" will be relatively high and will be ranked higher; if a certain associated information element has a low semantic similarity with the key information element, it may be screened out.

[0053] The associated information elements in the filtered and sorted associated information element sequence are inserted one by one into the corresponding positions in the preliminary framework of the response statement to generate an intermediate version of the response statement containing the associated information elements. For example, "the updated history of insurance clauses related to cancer compensation" is inserted into the appropriate position in the preliminary framework of the response statement to form a more detailed intermediate version of the response statement.

[0054] Perform grammatical and semantic checks on the intermediate version of the reply statement to generate an optimized version of the reply statement after verification. Grammar checking may check whether the sentence structure is correct, such as whether the subject, predicate, and object are complete; semantic checking may check whether the reply content is logical and whether it accurately answers the question. For example, if a grammatical error or unclear semantics is found in the intermediate version of the reply statement, it will be corrected and optimized.

[0055] According to the dialogue context characteristics of the target insurance consultation dialogue and the user profile characteristics of the target user, the optimized version of the reply statement is personalized and adjusted, and the final version of the personalized adjusted reply statement is output as the reply statement result of the target insurance consultation dialogue, where the personalized adjustment includes adjusting the language style and adding emotional elements. For example, if the user profile of the target user shows that he is relatively young and likes a concise and clear language style, then the optimized version of the reply statement is adjusted to a more concise and lively language style; if the target user shows a more anxious emotion in the previous conversation, some soothing emotional elements can be added to the reply statement, such as "Don't worry, we will give you a detailed answer", etc., and finally this personalized adjusted reply statement is output to the target user as the reply statement result of the target insurance consultation dialogue.

[0056] Based on the above steps, the embodiment of the present application effectively improves the depth and breadth of dialogue understanding by accurately acquiring and processing the local and global semantic feature space of the target user's insurance consultation dialogue. By summarizing and encoding the global semantic features and combining them with the enhanced processing of the local semantic features, a comprehensive and detailed grasp of the dialogue content is achieved. The local and global semantic coding data are integrated and calculated to generate accurate target semantic coding, laying a solid foundation for subsequent knowledge matching. Furthermore, by calculating the characteristic distance between the target semantic coding and the knowledge fragments in the reference insurance knowledge base, the present invention can efficiently identify and select knowledge fragments that are highly relevant to the target dialogue, thereby generating accurate and appropriate reply statements, thereby significantly improving the intelligence level of insurance customer service dialogues, improving the accuracy and satisfaction of replies, and optimizing user experience.

[0057] In a possible implementation, step S110 includes:

[0058] Step S111, obtaining multiple semantic features corresponding to the target insurance consultation dialogue, and determining the semantic complexity of each semantic vector in the multiple semantic features.

[0059] Step S112, performing semantic feature space clustering on each semantic vector in the multiple semantic features based on the semantic complexity, to generate a local semantic feature space and a global semantic feature space of the target insurance consultation dialogue.

[0060] Wherein, step S111 includes:

[0061] A. Determine, among the multiple semantic features, the number of associated concepts corresponding to each semantic vector when it corresponds to different vector values; determine, based on the number of associated concepts and the number of global semantic concepts of the target insurance consulting dialogue, the first associated parameters of each semantic vector in the multiple semantic features when it corresponds to different vector values; determine the semantic complexity of each semantic vector in the multiple semantic features based on the associated parameters corresponding to different vector values ​​and the corresponding number of associated concepts.

[0062] Alternatively, B. Determine, among the multiple semantic features, the number of concept associations corresponding to each semantic vector when it corresponds to different vector values; determine, based on the number of concept associations and the number of global concept associations of the target insurance consulting dialogue, the second association parameters of each semantic vector in the multiple semantic features when it corresponds to different vector values; determine the semantic complexity of each semantic vector in the multiple semantic features based on the second association parameters corresponding to different vector values ​​and the corresponding number of concept associations.

[0063] In this embodiment, first, multiple semantic features corresponding to the target insurance consultation dialogue are obtained. These semantic features are meaningful elements obtained after analyzing the dialogue content and exist in the form of semantic vectors. For example, if the target insurance consultation dialogue is "I want to know what the insurance compensation process is like if I am diagnosed with cancer in major disease insurance, and whether the compensation ratio of this insurance is the same for different types of cancer?", the words "major disease insurance", "cancer", "compensation process", "compensation ratio" and so on will be converted into corresponding semantic vectors.

[0064] Next, the semantic complexity of each semantic vector in these multiple semantic features is determined. There are two ways to determine the semantic complexity. Let's look at the first way first. The server will determine the number of associated concepts corresponding to each semantic vector when it corresponds to different vector values ​​in multiple semantic features. For example, for the semantic vector "critical disease insurance", when it takes a specific vector value, the concepts associated with it are "disease type", "insurance clause", "insured population", etc. This is the number of associated concepts corresponding to it, assuming that there are 3 here. At the same time, the server will count the number of global semantic concepts of the target insurance consultation dialogue. In this example, the entire dialogue involves concepts such as "critical disease insurance", "cancer", "compensation process", "compensation ratio", "disease type", "insurance clause", "insured population", etc., assuming that there are 10 different semantic concepts in total. Then, based on the number of associated concepts and the number of global semantic concepts, the first associated parameters of each semantic vector corresponding to different vector values ​​are determined respectively. For the semantic vector "critical disease insurance", its first associated parameter is 3 divided by 10, which is equal to 0.3. Finally, the semantic complexity is determined based on the associated parameters corresponding to different vector values ​​and the number of associated concepts. Assume that the semantic complexity of the semantic vector of "critical illness insurance" is 0.6 through a specific calculation method (such as multiplying by a preset weight factor, etc.). In the same way, the server will perform such calculations on each semantic vector in the conversation to determine its semantic complexity.

[0065] Let's look at the second way to determine semantic complexity. Among multiple semantic features, the server determines the number of concept associations corresponding to different vector values ​​for each semantic vector. For example, for the semantic vector "payment process", when it takes a certain vector value, other concepts that are conceptually associated with it include "claim information" and "claim time", etc. Assume that the number of its concept associations is 2. Then, based on this number of concept associations and the number of global concept associations of the target insurance consultation dialogue, determine the second association parameter of each semantic vector corresponding to different vector values. Assuming that the number of global concept associations is 8, the second association parameter of the semantic vector "payment process" is 2 divided by 8, which equals 0.25. Then, based on the second association parameter corresponding to different vector values ​​and the corresponding number of concept associations, determine its semantic complexity. Assume that the semantic complexity of the semantic vector "payment process" obtained through specific calculations is 0.5. Similarly, the server will perform this operation on all semantic vectors.

[0066] After determining the semantic complexity of each semantic vector, the server performs semantic feature space clustering on each semantic vector in multiple semantic features based on these semantic complexities, thereby generating a local semantic feature space and a global semantic feature space of the target insurance consultation dialogue. Semantic vectors with higher semantic complexity are often closely related to the core content of the target insurance consultation dialogue and will be clustered into the local semantic feature space. For example, in the above example, semantic vectors such as "cancer", "compensation process", and "compensation ratio" may be clustered into the local semantic feature space because they are directly related to the key issues of user consultation and have higher semantic complexity. However, those semantic vectors with relatively low semantic complexity but related to the entire insurance topic, such as some common insurance terms or semantic vectors corresponding to concepts that only play an auxiliary explanatory role in the dialogue, will be clustered into the global semantic feature space. For example, some semantic vectors corresponding to the concept of "risk assessment" commonly used in the insurance industry (if mentioned in the dialogue) may be clustered into the global semantic feature space. Through such meticulous clustering operations, the server successfully constructed the local semantic feature space and global semantic feature space of the target insurance consultation dialogue. These two semantic feature spaces will provide an important data basis for subsequent operations, so as to more accurately process the target insurance consultation dialogue and generate appropriate responses.

[0067] In a possible implementation, step S110 may further include:

[0068] Obtain multiple semantic features corresponding to the target insurance consultation dialogue.

[0069] Determine the local-global semantic range ratio, and perform semantic feature space clustering on each semantic vector in the plurality of semantic features according to the local-global semantic range ratio to generate a local semantic feature space and a global semantic feature space of the target insurance consultation dialogue. Or,

[0070] Based on the semantic feature space clustering instruction, the semantic vectors of the plurality of semantic features are clustered and output into a local semantic feature space and a global semantic feature space.

[0071] In this embodiment, first, multiple semantic features corresponding to the target insurance consultation dialogue can be obtained. For example, the target insurance consultation dialogue is "I want to know what the insurance compensation process is like if I am diagnosed with cancer in major disease insurance, and whether the compensation ratio of this insurance is the same for different types of cancer?" In this dialogue, words such as "major disease insurance", "cancer", "compensation process", and "compensation ratio" will be parsed into corresponding semantic features. These semantic features exist in the form of vectors, which contain various information about these words at the semantic level.

[0072] Then one way is to determine the local-global semantic range ratio, and cluster the semantic feature space of each semantic vector in multiple semantic features according to the ratio, so as to generate the local semantic feature space and global semantic feature space of the target insurance consultation dialogue. Assume that the server determines the local semantic range ratio to be 0.6 and the global semantic range ratio to be 0.4. The server will cluster the semantic vectors according to this ratio. For the semantic vectors in the target insurance consultation dialogue, those semantic vectors that are closely related to the core issues in the dialogue, have a high frequency of occurrence, or have a high semantic importance will be divided into the local semantic feature space according to this ratio. For example, semantic vectors such as "cancer" and "payment process" are directly related to the key content of the user's inquiry. According to the local semantic range ratio of 0.6, they are likely to be divided into the local semantic feature space. And those semantic vectors that are related to insurance topics but are relatively broad or auxiliary, such as the semantic vectors corresponding to some general insurance concepts, will be divided into the global semantic feature space according to the global semantic range ratio of 0.4. This clustering method based on the ratio of local to global semantic range can make the local semantic feature space focus on the core semantic part of the conversation, while the global semantic feature space covers a wider range of insurance-related semantics, providing a reasonable semantic structure for subsequent processing.

[0073] Another way is to cluster and output each semantic vector in multiple semantic features into a local semantic feature space and a global semantic feature space based on a semantic feature space clustering instruction. This semantic feature space clustering instruction is a pre-set rule or algorithm. For example, the instruction stipulates that the semantic vectors related to the name of a specific disease are clustered into the local semantic feature space, and the semantic vectors related to the general concept of insurance are clustered into the global semantic feature space. In the target insurance consultation dialogue, the semantic vector of "cancer" related to the name of a specific disease will be clustered into the local semantic feature space according to the instruction; and the semantic vector corresponding to the general concept of insurance such as "insurance contract" will be clustered into the global semantic feature space. Through this instruction-based clustering method, the server can quickly and accurately construct the local semantic feature space and the global semantic feature space of the target insurance consultation dialogue. These two semantic feature spaces will serve as the data basis for subsequent processing steps, which will help to make appropriate responses to the target insurance consultation dialogue more efficiently and accurately.

[0074] In a possible implementation, step S120 includes:

[0075] Step S121, respectively fuse the global semantic feature spaces belonging to the same target insurance consultation dialogue to generate a first fused semantic feature space for each target insurance consultation dialogue.

[0076] Step S122, respectively load each of the first fused semantic feature spaces into the global semantic sub-model of the customer service dialogue model, so that the multi-level semantic coding nodes in the global semantic sub-model encode and represent the first fused semantic feature space to generate global semantic coding data.

[0077] Step S123: Fusing the local semantic feature space and the global semantic feature space to generate a second fused semantic feature space.

[0078] Step S124, loading the second fused semantic feature space into the local semantic sub-model of the customer service dialogue large model respectively, so that the multi-level semantic coding nodes in the local semantic sub-model encode and represent the second fused semantic feature space to generate local semantic coding data.

[0079] In this embodiment, the global semantic feature spaces belonging to the same target insurance consultation dialogue are first fused to generate a first fused semantic feature space. For example, the target insurance consultation dialogue is "I want to know what the insurance compensation process is like if I am diagnosed with cancer in major disease insurance, and whether the compensation ratio of this insurance is the same for different types of cancer?", and its global semantic feature space contains relevant semantic vectors such as "insurance" and "major disease". The server will fuse these semantic vectors according to specific fusion rules. This fusion rule may be based on vector addition or weighted summation operations. For example, for the semantic vector "insurance", its elements are added to the corresponding elements of the semantic vector "major disease" according to certain weights. By performing such operations on all semantic vectors in the global semantic feature space, the first fused semantic feature space of each target insurance consultation dialogue is generated.

[0080] Then, each first fused semantic feature space is loaded into the global semantic sub-model of the customer service dialogue model. This global semantic sub-model contains multiple levels of semantic coding nodes, for example, there may be three layers of semantic coding nodes. After the first fused semantic feature space is loaded into the global semantic sub-model, the first layer of semantic coding nodes will perform preliminary coding processing on the basic semantic information in the first fused semantic feature space. For example, for the basic concept elements related to insurance in the first fused semantic feature space, the first layer of nodes will convert them into a coding form that is more suitable for subsequent processing, which may be to integrate some discrete semantic information into a more compact coding representation. Then the second layer of semantic coding nodes will further abstract and encode the results of the first layer of coding, and can further refine the broader representation of insurance concepts in the first layer of coding, and extract semantic features that are more relevant to the target insurance consultation dialogue, such as focusing the general concept of insurance on the semantic features related to major disease insurance. Finally, the third layer of semantic coding nodes will perform a higher-level coding representation on the second layer of coding results, and can associate the previous semantic features with some general rules or knowledge of the entire insurance industry, thereby generating global semantic coding data. This global semantic coding data contains highly abstracted and encoded information on the global semantic feature space of the target insurance consultation dialogue, and can provide an effective data basis for subsequent operations such as determining the relevance with the reference insurance knowledge base.

[0081] The process of performing semantic enhancement operations on the local semantic feature space and the global semantic feature space to generate local semantic encoding data is as follows. First, the local semantic feature space and the global semantic feature space are fused to generate a second fused semantic feature space. Still taking the above-mentioned target insurance consultation dialogue as an example, the local semantic feature space may contain semantic vectors such as "cancer" and "payment process", and the global semantic feature space contains semantic vectors such as "insurance" and "major disease". During fusion, the server will combine these semantic vectors together, possibly splicing them into a new semantic structure according to a certain order or rule, thereby generating a second fused semantic feature space.

[0082] Then the second fused semantic feature space is loaded into the local semantic sub-model of the customer service dialogue model. This local semantic sub-model also contains multi-level semantic coding nodes, for example, there are three layers of semantic coding nodes. When the second fused semantic feature space is loaded into the local semantic sub-model, the first layer of semantic coding nodes will focus on encoding the local semantic feature part in the second fused semantic feature space. For example, for the core semantic vectors such as "cancer" and "compensation process" in the local semantic feature space, the first layer of nodes will perform special encoding processing on them to highlight their importance in local semantics. The second layer of semantic coding nodes will comprehensively consider the relationship between local and global semantics for encoding. "Cancer" in the local semantic features can be associated with "major disease insurance" in the global semantic features to ensure the consistency and coherence of local semantics and global semantics. Finally, the third layer of semantic coding nodes will optimize the encoding of the whole. According to the encoding results of the previous two layers and some predefined rules of the entire customer service dialogue model, the second fused semantic feature space can be finally encoded and represented, thereby generating local semantic encoding data. This local semantic coding data contains local and global semantic information after semantic enhancement operations, and can play an important role in subsequent operations such as fusion calculation with global semantic coding data.

[0083] In a possible implementation, the method further includes:

[0084] Step A110, determining the vector length or semantic content of the semantic mapping feature of the target insurance consultation dialogue.

[0085] Step A120, based on the local semantic feature space, the global semantic feature space and the vector length, determine the influence coefficients corresponding to the local semantic coding data and the global semantic coding data, respectively. Alternatively, based on the local semantic feature space, the global semantic feature space and the semantic content, determine the influence coefficients corresponding to the local semantic coding data and the global semantic coding data, respectively.

[0086] Step S130 includes: performing a fusion calculation on the local semantic coding data and the global semantic coding data according to the influence coefficient.

[0087] In this embodiment, taking the target insurance consultation dialogue "I want to know what the insurance compensation process is like if I am diagnosed with cancer in major disease insurance, and whether the compensation ratio of this insurance is the same for different types of cancer?" as an example, the determination of the vector length or semantic content of the semantic mapping feature is based on the semantic information in the entire dialogue. If the vector length of the semantic mapping feature is determined, the server will analyze the composition of the semantic vector in the local semantic feature space and the global semantic feature space. In the local semantic feature space, semantic vectors such as "cancer" and "compensation process" have certain dimensions, and semantic vectors such as "major disease insurance" in the global semantic feature space also have their dimensions. Through the comprehensive statistics and analysis of these semantic vector dimensions, the vector length of the semantic mapping feature is obtained.

[0088] Or if the semantic content is determined, the server will quantify the amount of semantic information contained in each semantic vector. For example, the semantic vector "cancer" may be associated with semantic information such as multiple cancer types and incidence conditions. This information is quantified and counted. At the same time, the semantic vectors in the global semantic feature space, such as the insurance terms and coverage of "major disease insurance", are quantified to comprehensively derive the semantic content of the entire target insurance consultation dialogue.

[0089] Based on the local semantic feature space, global semantic feature space, and vector length or semantic content determined above, the server further determines the influence coefficients corresponding to the local semantic coding data and the global semantic coding data, respectively. If the influence coefficient is determined based on the vector length, it is assumed that the sum of the semantic vector lengths in the local semantic feature space is L1, the sum of the semantic vector lengths in the global semantic feature space is L2, and the vector length of the previously obtained semantic mapping feature is L. Then the influence coefficient corresponding to the local semantic coding data may be calculated by the formula C1 = L1 / L, and the influence coefficient corresponding to the global semantic coding data is C2 = L2 / L. If the influence coefficient is determined based on semantic content, it is assumed that the semantic content in the local semantic feature space is S1, the semantic content in the global semantic feature space is S2, and the total semantic content is S, then the influence coefficient C1 corresponding to the local semantic coding data = S1 / S, and the influence coefficient C2 corresponding to the global semantic coding data = S2 / S.

[0090] After determining the influence coefficient, the server performs a fusion calculation of the local semantic coding data and the global semantic coding data. Based on the previously determined influence coefficient, assuming that the local semantic coding data is D1 and the global semantic coding data is D2, a fusion calculation is performed. If the fusion calculation is performed by weighted summation, the fused target semantic coding data D = C1 * D1 + C2 * D2. For example, C1 is 0.4, C2 is 0.6, D1 is a specific local semantic coding representation, and D2 is a specific global semantic coding representation. The target semantic coding data D is obtained by calculating 0.4 * D1 + 0.6 * D2. This fusion calculation method fully considers the different influence levels of the local semantic feature space and the global semantic feature space in the target insurance consultation dialogue, and reasonably fuses the local semantic coding data and the global semantic coding data through the influence coefficient, so that the generated target semantic coding data can better integrate local and global semantic information, and provide an accurate data basis for subsequent operations such as determining feature distances based on the target semantic coding data and the knowledge semantic coding data of the reference insurance knowledge base.

[0091] In a possible implementation, the global semantic coding data, the local semantic coding data and the target semantic coding data are generated by training using a deep learning network. The method further includes:

[0092] Step S101, obtaining a sample local semantic feature space and a sample global semantic feature space of a sample insurance consultation dialogue.

[0093] In this embodiment, taking one of the many sample insurance consultation dialogues as an example, assume that the sample insurance consultation dialogue is "I want to know how the amount of compensation for life insurance is determined when the insured dies accidentally, and whether the insured's health condition will be taken into consideration." For this sample insurance consultation dialogue, the server will obtain the sample local semantic feature space and the sample global semantic feature space in a similar manner to the previous processing of the target insurance consultation dialogue. For example, the semantic vectors corresponding to words such as "life insurance", "accidental death", "compensation amount", and "health condition" will be divided into the sample local semantic feature space or the sample global semantic feature space based on factors such as their semantic association in the dialogue.

[0094] Step S102, using the first encoding structure of the deep learning network, performs semantic induction encoding on the sample global semantic feature space to generate sample global semantic encoding data.

[0095] In this embodiment, the first coding structure of the deep learning network includes multiple levels and nodes. The semantic vector in the sample global semantic feature space is input into the first coding structure, and in the first layer of nodes, the semantic vector may be subjected to preliminary feature extraction and coding conversion. For example, for the semantic vector of "life insurance", the first layer of nodes may convert it into an initial coding form based on the internal representation of the neural network. This coding form will highlight some basic features of the concept of "life insurance" in the global semantics of the entire sample insurance consultation dialogue. As the data is passed through the levels in the first coding structure, subsequent nodes will further abstract and optimize the previous coding results. For example, the second layer of nodes may associate "life insurance" with some common concepts in the entire insurance field for coding, and the third layer of nodes may generate a higher level of abstract coding based on the previous coding results, and finally generate sample global semantic coding data through the processing of the entire first coding structure.

[0096] Step S103, using the second encoding structure of the deep learning network, performing semantic enhancement operations on the sample local semantic feature space and the sample global semantic feature space to generate sample local semantic encoding data.

[0097] In this embodiment, the second coding structure also has multiple levels and nodes. After the data of the sample local semantic feature space and the sample global semantic feature space are input into the second coding structure, the key semantic vectors in the sample local semantic feature space may be emphatically encoded at the first layer of nodes. For example, for "accidental death", a relatively core semantic vector in the sample local semantic feature space, the first layer of nodes will perform special encoding processing on it to enhance its semantic representation. The second layer of nodes will comprehensively consider the relationship between the sample local semantic feature space and the sample global semantic feature space for encoding. For example, "accidental death" is associated with "life insurance" in the sample global semantic feature space to ensure the consistency and coherence of local and global semantics. Subsequent nodes will continue to optimize the encoding results, and finally generate sample local semantic encoding data through the second encoding structure.

[0098] Step S104, utilizing the interactive structure of the deep learning network to fuse the sample local semantic coding data and the sample global semantic coding data to generate target sample semantic coding data.

[0099] In this embodiment, the interactive structure of the deep learning network will perform fusion calculations according to preset algorithms and rules. For example, the weights may be determined based on the importance of the sample local semantic coding data and the sample global semantic coding data in the entire sample insurance consultation dialogue, and then fused in a weighted summation manner. Assuming that the sample local semantic coding data is D1, the sample global semantic coding data is D2, and the weights determined by the interactive structure are W1 and W2 respectively, then the target sample semantic coding data D = W1 * D1 + W2 * D2.

[0100] Step S105 , determining a second characteristic distance of the sample reference insurance knowledge base according to the target sample semantic encoding data and the sample training knowledge semantic encoding data of each sample knowledge fragment in the sample reference insurance knowledge base.

[0101] In this embodiment, the sample reference insurance knowledge base contains a large number of sample knowledge fragments, and each sample knowledge fragment has corresponding sample training knowledge semantic coding data. For example, there is a sample knowledge fragment in the sample reference insurance knowledge base that is a detailed provision on the details of life insurance claims, and its sample training knowledge semantic coding data is K1. The server will use a specific distance measurement algorithm, such as the Euclidean distance algorithm, to calculate the second feature distance between the target sample semantic coding data D and the sample training knowledge semantic coding data K1. Assuming D = (d1, d2, d3), K1 = (k1, k2, k3), then the second feature distance d = √((d1 - k1)²+(d2 - k2)²+(d3 - k3)²). In the same way, the server will calculate the second feature distance between the target sample semantic coding data and each sample knowledge fragment in the sample reference insurance knowledge base.

[0102] Step S106: training the deep learning network based on the error between the second feature distance and the training supervision data.

[0103] In this embodiment, the training supervision data is pre-set standard result data. For example, for a certain sample insurance consultation dialogue, there is an ideal second feature distance result as training supervision data based on professional knowledge or manual annotation. The server will calculate the error between the second feature distance actually calculated and the training supervision data. If the error is large, it means that the current parameter setting of the deep learning network is not appropriate and needs to be adjusted. Based on this error, the server will use technical means such as back propagation algorithm to adjust the various parameters in the deep learning network, such as adjusting the weights, thresholds and other parameters in the first coding structure, the second coding structure and the interaction structure, so that the deep learning network can more accurately generate sample global semantic coding data, sample local semantic coding data and target sample semantic coding data in subsequent processing, thereby continuously improving the entire system's processing capabilities for insurance consultation dialogues.

[0104] In a possible implementation, the method further includes:

[0105] The third characteristic distance of the sample reference insurance knowledge base is determined based on the sample global semantic encoding data and the sample training knowledge semantic encoding data of the sample reference insurance knowledge base.

[0106] The fourth characteristic distance of the sample reference insurance knowledge base is determined based on the sample local semantic encoding data and the sample training knowledge semantic encoding data of the sample reference insurance knowledge base.

[0107] Step S106 may include: training the deep learning network based on the error between the second feature distance and the training supervision data, the error between the third feature distance and the training supervision data, and the error between the fourth feature distance and the training supervision data.

[0108] In this embodiment, taking the previously mentioned sample insurance consultation dialogue "I want to know how the compensation amount of life insurance is determined when the insured dies accidentally, and whether the insured's health status will be considered" as an example, after the sample global semantic coding data is generated by the first coding structure of the deep learning network, it is compared with the sample training knowledge semantic coding data in the sample reference insurance knowledge base. The sample reference insurance knowledge base contains many sample knowledge fragments about insurance knowledge, and each fragment has corresponding sample training knowledge semantic coding data. For example, one of the sample knowledge fragments is a detailed provision on the details of life insurance compensation, and its sample training knowledge semantic coding data is K1. Assuming that the sample global semantic coding data is G, when calculating the third feature distance, a specific distance measurement algorithm, such as the Euclidean distance algorithm, is used. If G = (g1, g2, g3), K1 = (k1, k2, k3), then the third feature distance d3 = √((g1 - k1)²+(g2 -k2)²+(g3 - k3)²). In the same way, the server calculates the third feature distance between the sample global semantic encoding data and each sample knowledge fragment in the sample reference insurance knowledge base.

[0109] Next, after the sample local semantic coding data is generated through the second coding structure of the deep learning network, it is compared and calculated with the sample training knowledge semantic coding data in the sample reference insurance knowledge base. For example, the sample local semantic coding data is L, and for the sample training knowledge semantic coding data K1 corresponding to the sample knowledge fragment in the sample reference insurance knowledge base, the Euclidean distance algorithm is also used (assuming that), if L = (l1, l2, l3), then the fourth characteristic distance d4 = √((l1 - k1)²+(l2 - k2)²+(l3 - k3)²). Following the same process, the server will calculate the fourth characteristic distance between the sample local semantic coding data and each sample knowledge fragment in the sample reference insurance knowledge base.

[0110] Then, based on the error between the second feature distance and the training supervision data, the error between the third feature distance and the training supervision data, and the error between the fourth feature distance and the training supervision data, the deep learning network is trained. The training supervision data is the pre-set ideal result data, which reflects the feature distance relationship that the sample insurance consultation dialogue should have with the sample reference insurance knowledge base under ideal conditions. For example, for this sample insurance consultation dialogue, the ideal value corresponding to each feature distance is set as the training supervision data based on professional knowledge and experience. The server calculates the error E2 between the actual second feature distance and the training supervision data, the error E3 between the third feature distance and the training supervision data, and the error E4 between the fourth feature distance and the training supervision data. Then, the parameters of the deep learning network are adjusted by synthesizing these error information. For example, the back propagation algorithm is used to adjust the weights, thresholds and other parameters in the first coding structure, the second coding structure and the interaction structure in the deep learning network according to a certain proportional relationship based on these errors. If E2 is large, it means that there may be problems in the fusion calculation or the previous encoding operation, and the parameters in the relevant structure need to be adjusted to reduce the error; similarly, if E3 or E4 is large, the parameters in the relevant structure also need to be adjusted in a targeted manner to improve the deep learning network's processing accuracy of the sample insurance consultation dialogue, so that it can more accurately generate semantic encoding data and determine the feature distance between the sample reference insurance knowledge base.

[0111] In a possible implementation, the method further includes:

[0112] Determine the network learning weight of the sample insurance consultation dialogue. The network learning weight represents the training frequency of the sample insurance consultation dialogue being called for network training.

[0113] Based on the network learning weights, the sample local semantic feature space and the sample global semantic feature space, weighting factors corresponding to the sample local semantic coding data and the sample global semantic coding data are determined.

[0114] Alternatively, the method further comprises:

[0115] Determine the vector length or semantic content of the semantic mapping feature of the sample insurance consultation dialogue.

[0116] Based on the sample local semantic feature space, the sample global semantic feature space and the vector length, the influence coefficients corresponding to the sample local semantic coding data and the sample global semantic coding data are determined. Alternatively, based on the sample local semantic feature space, the sample global semantic feature space and the semantic content, the influence coefficients corresponding to the sample local semantic coding data and the sample global semantic coding data are determined.

[0117] The fusing calculation of the sample local semantic coding data and the sample global semantic coding data comprises:

[0118] According to the weighting factor, the sample local semantic coding data and the sample global semantic coding data are fused and calculated.

[0119] In a possible implementation, the server also needs to determine the network learning weight of the sample insurance consultation dialogue. This network learning weight represents the training frequency of the sample insurance consultation dialogue being called for network training. For example, some sample insurance consultation dialogues may be frequently used for network training due to their representativeness or complexity, and their network learning weight will be higher; while some simpler or less common sample insurance consultation dialogues may be called less frequently, and their network learning weight will be lower. Assume that for the above sample insurance consultation dialogue, its network learning weight is determined to be W based on factors such as its importance and frequency of occurrence in the entire insurance consultation scenario.

[0120] Based on the network learning weights, the sample local semantic feature space, and the sample global semantic feature space, the weighting factors corresponding to the sample local semantic coding data and the sample global semantic coding data are determined. For example, the weighting factors are determined by a specific calculation method based on the network learning weight W, the semantic vector information in the sample local semantic feature space, and the semantic vector information in the sample global semantic feature space. Assuming that by analyzing the semantic vector in the sample local semantic feature space, a certain metric value related to training is obtained as M1, and the corresponding metric value in the sample global semantic feature space is M2, the weighting factor F1 corresponding to the sample local semantic coding data and the weighting factor F2 corresponding to the sample global semantic coding data are determined according to a pre-set calculation rule (for example, weighting factor F1 = W * M1 / ​​(M1 + M2), F2 = W * M2 / (M1 + M2)).

[0121] When performing the fusion calculation of the sample local semantic coding data and the sample global semantic coding data, the sample local semantic coding data and the sample global semantic coding data are fused according to the weighting factor. Assuming that the sample local semantic coding data is L and the sample global semantic coding data is G, the fusion calculation is performed in a weighted summation manner, and the fusion result D = F1 * L + F2 * G. This fusion calculation method takes into account the network learning weights of the sample insurance consultation dialogue and the relevant information of the sample local semantic feature space and the sample global semantic feature space, so that the fusion result can better reflect the semantic characteristics of the sample insurance consultation dialogue and is compatible with the training process of the deep learning network.

[0122] In addition, the server can also determine the vector length or semantic content of the semantic mapping feature of the sample insurance consultation dialogue. Taking the vector length as an example, the server analyzes the dimensional information of the semantic vector in the sample local semantic feature space and the sample global semantic feature space, and comprehensively calculates the vector length of the semantic mapping feature. Based on the sample local semantic feature space, the sample global semantic feature space and the vector length, the influence coefficients corresponding to the sample local semantic coding data and the sample global semantic coding data are determined. Assuming that the sum of the semantic vector lengths in the sample local semantic feature space is L1, the sum of the semantic vector lengths in the sample global semantic feature space is L2, and the vector length of the semantic mapping feature is L, the influence coefficient is determined according to the calculation rule (for example, the influence coefficient C1 corresponding to the sample local semantic coding data = L1 / L, and the influence coefficient C2 corresponding to the sample global semantic coding data = L2 / L).

[0123] When the sample local semantic coding data and the sample global semantic coding data are fused and calculated, the sample local semantic coding data and the sample global semantic coding data are fused and calculated based on the weighting factor (the weighting factor here can be determined based on the network learning weight as mentioned above, or can be determined based on the influence coefficient, depending on the specific application scenario), so as to obtain a more satisfactory fusion result, and provide more effective data support for the training of the deep learning network and the subsequent accurate processing of the insurance consultation dialogue. Through such a series of complex and orderly operations, the server can continuously optimize the deep learning network's processing capabilities for insurance consultation dialogues, improve the accuracy of the semantic coding data and the accuracy of the feature distance calculation between the reference knowledge base.

[0124] For example, in a possible implementation, step S150 includes:

[0125] Step S151, identifying key information elements and associated information elements contained in the target knowledge segment, wherein the key information elements represent the core semantic content in the target knowledge segment for answering the target insurance consultation dialogue, and the associated information elements represent the auxiliary semantic content associated with the key information elements.

[0126] In this embodiment, for example, the target knowledge segment is the detailed rules for determining the amount of compensation for life insurance when the insured dies unexpectedly, where the key information elements may include "the calculation basis of the compensation amount (such as the insured amount, payment period, etc.)" and "whether the compensation amount is adjusted according to the insured's health status". These elements are the core semantic content that directly answers the target insurance consultation dialogue "I want to know how the compensation amount of life insurance is determined when the insured dies unexpectedly, and whether the insured's health status is considered". The associated information elements may be "compensation in similar cases" and "historical changes in accidental death compensation in insurance terms", which are auxiliary semantic content associated with the key information elements.

[0127] Step S152: construct a preliminary framework of a reply statement based on the key information element set. The preliminary framework includes at least one key information placeholder for inserting specific key information content.

[0128] In this embodiment, this preliminary framework will contain at least one key information placeholder for inserting specific key information content. For example, the preliminary framework may be "Regarding the method of determining the amount of compensation for the accidental death of the life insurance policyholder you inquired about, the basis for calculating the amount of compensation is [the key information placeholder for calculating the basis of compensation], and whether to consider the policyholder's health status is [the key information placeholder for whether to consider health status]". The construction of this framework is based on the structure and logical relationship of the key information elements, and is prepared for the subsequent insertion of specific content.

[0129] Step S153, based on the semantic similarity between the associated information elements and the key information elements, the importance of the associated information elements in the insurance field, and the historical interests of the target user, the set of associated information elements is screened and sorted to select the associated information elements that match the target insurance consultation dialogue, and generate a screened and sorted sequence of associated information elements.

[0130] For example, the historical interests of the target user show that they are more concerned about the historical changes in insurance terms, so the importance of the associated information element "Historical changes in accidental death compensation in insurance terms" is relatively high. From the perspective of semantic similarity, if a related information element has a strong semantic association with the key information element, for example, "compensation in similar cases" and "basis for calculating the compensation amount" have certain similarities and auxiliary functions in explaining the method of determining the compensation amount, then it will also be ranked higher in the sorting. After such screening and sorting, a sequence of related information elements that meets the needs of the target insurance consultation dialogue is generated.

[0131] Step S154, inserting the associated information elements in the filtered and sorted associated information element sequence into corresponding positions in the preliminary framework of the reply statement one by one, and generating an intermediate version of the reply statement containing the associated information elements.

[0132] For example, inserting "historical changes in insurance terms regarding accidental death compensation" into the appropriate position in the preliminary framework of the response statement may result in "Regarding the method of determining the amount of compensation when the life insurance policyholder dies accidentally, the basis for calculating the compensation amount is [placeholder for key information on the basis for calculating the compensation amount], whether the policyholder's health status is considered is [placeholder for key information on whether to consider health status], and the historical changes in insurance terms regarding accidental death compensation are [specific historical changes]", thereby forming an intermediate version of the response statement containing related information elements.

[0133] Step S155, performing syntax verification and semantic verification on the intermediate version of the reply statement to generate a verified optimized version of the reply statement.

[0134] In detail, in terms of grammatical verification, the server will check whether the structure of the sentence is correct, such as whether the subject, predicate and object are complete, whether the collocation between the sentence components is reasonable, etc. For example, if there is a statement such as "the basis for calculating the compensation amount is [the key information placeholder for the basis for calculating the compensation amount] is" in the intermediate version of the reply sentence, this is a grammatical error and the server will correct it. Semantic verification will check whether the reply content is logical and whether it accurately answers the question. For example, if the statement about the basis for calculating the compensation amount in the intermediate version of the reply sentence does not match the content in the target knowledge fragment, or there is semantic ambiguity, the server will make adjustments to generate an optimized version of the reply sentence after verification.

[0135] Step S156, based on the dialogue context characteristics of the target insurance consultation dialogue and the user portrait characteristics of the target user, the optimized version of the reply statement is personalized adjusted, and the final version of the reply statement after the personalized adjustment is output as the reply statement result of the target insurance consultation dialogue, wherein the personalized adjustment includes adjusting the language style and adding emotional elements.

[0136] Finally, from the perspective of dialogue context characteristics, if the target insurance consultation dialogue is conducted in a more formal business scenario, the language style of the optimized version of the reply statement may be adjusted to a more formal and professional style; if it is in a more relaxed personal consultation scenario, the language style may become more easy to understand. In terms of the user portrait characteristics of the target user, if the target user is a young person, he or she may prefer a concise and clear language style, so the optimized version of the reply statement will be simplified accordingly. At the same time, if the target user shows anxiety in the previous conversation, emotional elements will be added to the reply statement, such as "Don't worry, we will accurately determine the amount of compensation for you based on detailed rules", etc., and finally this personalized adjusted reply statement will be output to the target user as the reply statement result of the target insurance consultation dialogue, so as to provide a response that is both accurate and meets user needs.

[0137] Figure 2 An insurance service system 100 provided in an embodiment of the present application is shown, including a processor 1001 and a memory 1003 and program code stored in the memory 1003, and the processor 1001 executes the above program code to implement the steps of the intelligent insurance customer service dialogue generation method based on a large model.

[0138] Figure 2 The insurance service system 100 shown includes: a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the insurance service system 100 may also include a transceiver 1004, which can be used for data interaction between the insurance service system and other insurance service systems, such as data transmission and / or data reception. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of the insurance service system 100 does not constitute a limitation on the embodiments of the present application.

[0139] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0140] The bus 1002 may include a path to transmit information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 1002 may be divided into an address bus, a data bus, a control bus, and the like.

[0141] The memory 1003 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium, other magnetic storage devices, or any other medium that can be used to have or store program code and can be read by a computer, without limitation herein.

[0142] The memory 1003 is used to store program codes for executing the embodiments of the present application, and the execution is controlled by the processor 1001. The processor 1001 is used to execute the program codes stored in the memory 1003 to implement the steps shown in the above method embodiments.

[0143] An embodiment of the present application provides a computer-readable storage medium having program code stored thereon. When the program code is executed by a processor, the steps and corresponding contents of the aforementioned method embodiment can be implemented.

[0144] It should be understood that, although each operation step is indicated by arrows in the flowchart of the embodiment of the present application, the implementation order of these steps is not limited to the order indicated by the arrows. Unless there is clear explanation in this article, in some implementation scenarios of the embodiment of the present application, the implementation steps in each flowchart can be performed in other orders based on demand. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages according to the actual implementation scenario, and some or all of these sub-steps or stages may be executed at the same time, and each sub-step or stage in these sub-steps or stages may also be executed at different times respectively. In different scenarios at the execution time, the execution order of these sub-steps or stages can be flexibly configured based on demand, and the embodiment of the present application does not limit this.

[0145] The above is only an optional implementation method for some implementation scenarios of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present application, other similar implementation methods based on the technical ideas of the present application are also within the protection scope of the embodiments of the present application.

Claims

1. A method for generating intelligent insurance customer service dialogue based on a large model, characterized in that: The method comprises: Obtaining the local semantic feature space and the global semantic feature space of the target insurance consultation dialogue of the target user; Performing semantic induction encoding on the global semantic feature space to generate global semantic encoding data, and performing semantic enhancement operation on the local semantic feature space and the global semantic feature space to generate local semantic encoding data; The local semantic coding data and the global semantic coding data are fused and calculated to generate target semantic coding data; Determine a first characteristic distance of each reference knowledge fragment in the reference insurance knowledge base based on the target semantic coding data and the knowledge semantic coding data of the reference insurance knowledge base; the first characteristic distance represents the knowledge relevance between the reference knowledge fragment in the reference insurance knowledge base and the target insurance consultation dialogue; In the reference insurance knowledge base, a target knowledge segment is selected according to the first feature distance, and a reply statement result of the target insurance consultation dialogue is generated based on the target knowledge segment.

2. The method for generating intelligent insurance customer service dialogue based on a large model according to claim 1, characterized in that: The local semantic feature space and the global semantic feature space of the target insurance consultation dialogue of the target user are obtained by: Acquire multiple semantic features corresponding to the target insurance consultation dialogue, and determine the semantic complexity of each semantic vector in the multiple semantic features; Based on the semantic complexity, each semantic vector in the plurality of semantic features is clustered in a semantic feature space to generate a local semantic feature space and a global semantic feature space of the target insurance consultation dialogue; The step of determining the semantic complexity of each semantic vector in the plurality of semantic features comprises: Among the multiple semantic features, determine the number of associated concepts corresponding to each semantic vector when it corresponds to different vector values, and determine the first associated parameters of each semantic vector in the multiple semantic features when it corresponds to different vector values ​​based on the associated parameters when it corresponds to different vector values ​​and the corresponding number of associated concepts, and determine the semantic complexity of each semantic vector in the multiple semantic features based on the associated parameters when it corresponds to different vector values ​​and the corresponding number of associated concepts; Alternatively, among the multiple semantic features, determine the number of concept associations corresponding to each semantic vector when it corresponds to different vector values, and based on the number of concept associations and the number of global concept associations of the target insurance consultation dialogue, respectively determine the second association parameters of each semantic vector in the multiple semantic features when it corresponds to different vector values, and determine the semantic complexity of each semantic vector in the multiple semantic features based on the second association parameters corresponding to different vector values ​​and the corresponding number of concept associations.

3. The method for generating intelligent insurance customer service dialogue based on a large model according to claim 1, characterized in that: The local semantic feature space and the global semantic feature space of the target insurance consultation dialogue of the target user are obtained by: Obtain multiple semantic features corresponding to the target insurance consultation dialogue; Determine the local-global semantic range ratio, and perform semantic feature space clustering on each semantic vector in the plurality of semantic features according to the local-global semantic range ratio to generate a local semantic feature space and a global semantic feature space of the target insurance consultation dialogue; or, Based on the semantic feature space clustering instruction, the semantic vectors of the plurality of semantic features are clustered and output into a local semantic feature space and a global semantic feature space.

4. The method for generating intelligent insurance customer service dialogue based on a large model according to claim 1, characterized in that: The performing semantic inductive coding on the global semantic feature space to generate global semantic coding data comprises: fusing the global semantic feature spaces belonging to the same target insurance consultation dialogue respectively to generate a first fused semantic feature space for each target insurance consultation dialogue; Each of the first fused semantic feature spaces is loaded into the global semantic sub-model of the customer service dialogue model respectively, so that the multi-level semantic coding nodes in the global semantic sub-model encode and represent the first fused semantic feature space to generate global semantic coding data.

5. The method for generating intelligent insurance customer service dialogue based on a large model according to claim 1, characterized in that: The performing semantic enhancement operation on the local semantic feature space and the global semantic feature space to generate local semantic coding data comprises: Fusing the local semantic feature space and the global semantic feature space to generate a second fused semantic feature space; The second fused semantic feature space is loaded into the local semantic sub-model of the customer service dialogue model respectively, so that the multi-level semantic coding nodes in the local semantic sub-model encode and represent the second fused semantic feature space to generate local semantic coding data.

6. The method for generating intelligent insurance customer service dialogue based on a large model according to claim 1, characterized in that: The method further comprises: Determining a vector length or semantic content of a semantic mapping feature of the target insurance consultation dialogue; Determine the influence coefficients corresponding to the local semantic coding data and the global semantic coding data respectively based on the local semantic feature space, the global semantic feature space and the vector length; or determine the influence coefficients corresponding to the local semantic coding data and the global semantic coding data respectively based on the local semantic feature space, the global semantic feature space and the semantic content; The fusing calculation of the local semantic coding data and the global semantic coding data comprises: According to the influence coefficient, the local semantic coding data and the global semantic coding data are fused and calculated.

7. The method for generating intelligent insurance customer service dialogue based on a large model according to any one of claims 1 to 6, characterized in that: The global semantic coding data, the local semantic coding data and the target semantic coding data are generated by training using a deep learning network; the method further includes: Obtaining a sample local semantic feature space and a sample global semantic feature space of a sample insurance consultation dialogue; Using the first coding structure of the deep learning network, semantic inductive coding is performed on the global semantic feature space of the sample to generate global semantic coding data of the sample; Using the second encoding structure of the deep learning network, a semantic enhancement operation is performed on the sample local semantic feature space and the sample global semantic feature space to generate sample local semantic encoding data; Utilizing the interactive structure of the deep learning network, the sample local semantic coding data and the sample global semantic coding data are fused and calculated to generate target sample semantic coding data; Determining a second characteristic distance of the sample reference insurance knowledge base according to the target sample semantic encoding data and the sample training knowledge semantic encoding data of each sample knowledge fragment in the sample reference insurance knowledge base; The deep learning network is trained based on the error between the second feature distance and the training supervision data.

8. The method for generating intelligent insurance customer service dialogue based on a large model according to claim 7, characterized in that: The method further comprises: Determining a third characteristic distance of the sample reference insurance knowledge base according to the sample global semantic encoding data and the sample training knowledge semantic encoding data of the sample reference insurance knowledge base; Determining a fourth characteristic distance of the sample reference insurance knowledge base based on the sample local semantic coding data and the sample training knowledge semantic coding data of the sample reference insurance knowledge base; The training of the deep learning network based on the error between the second feature distance and the training supervision data includes: The deep learning network is trained based on the error between the second feature distance and the training supervision data, the error between the third feature distance and the training supervision data, and the error between the fourth feature distance and the training supervision data.

9. The method for generating intelligent insurance customer service dialogue based on a large model according to claim 7, characterized in that: The method further comprises: Determining a network learning weight of the sample insurance consultation dialogue; the network learning weight represents the training frequency of the sample insurance consultation dialogue being called for network training; Determine weighting factors corresponding to the sample local semantic coding data and the sample global semantic coding data respectively based on the network learning weight, the sample local semantic feature space and the sample global semantic feature space; Alternatively, the method further comprises: Determining the vector length or semantic content of the semantic mapping feature of the sample insurance consultation dialogue; Determine the influence coefficients corresponding to the sample local semantic coding data and the sample global semantic coding data respectively based on the sample local semantic feature space, the sample global semantic feature space and the vector length; or determine the influence coefficients corresponding to the sample local semantic coding data and the sample global semantic coding data respectively based on the sample local semantic feature space, the sample global semantic feature space and the semantic content; The fusing calculation of the sample local semantic coding data and the sample global semantic coding data comprises: According to the weighting factor, the sample local semantic coding data and the sample global semantic coding data are fused and calculated.

10. An insurance service system, characterized in that: It includes a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the method for generating intelligent insurance customer service dialogue based on a large model as described in any one of claims 1 to 9 is implemented.

Images