AI Semantic Understanding-Based Debt Collection Conversation Generation Method and System

Abstract

The present invention provides a method and system for generating collection conversations based on AI semantic understanding. By obtaining and utilizing the digital human style portrait vector and the initial collection conversation flow of the first virtual collection digital human, precise control and conversion of the collection conversation style are achieved. By extracting the style flow conversion vector from the collection template sub-conversation sequence of each collection style template conversation and combining it with the digital human style portrait vector of the first virtual collection digital human, the first collection style conversion is performed through the collection style conversion network, and finally a target collection conversation flow matching the target collection style is generated, which not only improves the generation efficiency of the collection conversation, but also ensures the accuracy and consistency of the collection conversation style, thereby enhancing the collection effect and user satisfaction.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology. Specifically, it relates to a method and system for generating collection conversations based on AI semantic understanding. Background Art

[0002] In the financial service industry, debt collection is an important link to ensure the recovery of funds and reduce the risk of bad debts. Traditional debt collection methods mostly rely on manual labor, which is not only costly, but also the collection effect is affected by multiple factors such as personnel quality and emotion management, making it difficult to achieve efficient and consistent collection results. With the rapid development of artificial intelligence technology, especially the breakthroughs in natural language processing (NLP) and semantic understanding technology, new solutions have been brought to the field of debt collection.

[0003] In the prior art, there have been attempts to use AI technology to assist in debt collection, such as through automated voice response systems, SMS debt collection, etc. However, most of these methods are based on fixed scripts or templates, lacking flexibility and personalization, and it is difficult to adapt to the actual situations and psychological states of different debtors. Summary of the Invention

[0004] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for generating collection conversations based on AI semantic understanding, the method comprising:

[0005] Obtain the digital human style portrait vector and the initial collection conversation flow of the first virtual collection digital human. The digital human style portrait vector of the first virtual collection digital human is generated by extracting the collection style portrait features from historical digital human collection conversations. The historical digital human collection conversations contain the collection style templates of the first virtual collection digital human. The digital human style portrait vector is a collection style semantic vector used to identify the collection style portrait of the virtual collection digital human; the initial collection conversation flow includes X collection style template conversations, and each collection style template conversation contains the collection style templates of the second virtual collection digital human;

[0006] Extract the style flow turning vectors for the collection template sub-conversation sequences corresponding to each collection style template conversation, and generate the style flow turning vectors corresponding to each collection style template conversation. The collection template sub-conversation sequence corresponding to any one collection style template conversation includes Y collection template sub-conversations, and Y is less than X;

[0007] Perform the first collection style conversion on the style flow turning vector corresponding to each collection style template conversation, the collection template sub-conversation sequence corresponding to this collection style template conversation, and the digital human style portrait vector of the first virtual collection digital human through a collection style conversion network, and generate the collection style conversion results of each collection style template conversation;

[0008] Generate a target collection conversation flow corresponding to the initial collection conversation flow according to the collection style conversion result of the X collection style template conversations.

[0009] In a possible implementation manner of the first aspect, the step of performing a first collection style conversion on the style flow steering vector corresponding to each collection style template conversation, the collection template sub-conversation sequence corresponding to the collection style template conversation, and the digital human style portrait vector of the first virtual collection digital human through a collection style conversion network to generate a collection style conversion result for each collection style template conversation includes:

[0010] Encode and represent the collection template sub-conversation sequence corresponding to the a-th collection style template conversation to generate a potential style vector corresponding to the a-th collection style template conversation, where a is a positive integer not greater than X;

[0011] Integrate the style flow steering vector corresponding to the a-th collection style template conversation with the digital human style portrait vector of the first virtual collection digital human to generate an integration vector for the a-th collection style template conversation;

[0012] Aggregate the integration vector of the a-th collection style template conversation and the potential style vector corresponding to the a-th collection style template conversation to generate an aggregation vector for the a-th collection style template conversation;

[0013] Decode and represent the aggregation vector of the a-th collection style template conversation to generate a collection style conversion result for the a-th collection style template conversation.

[0014] In a possible implementation manner of the first aspect, the step of extracting a style flow steering vector from the collection template sub-conversation sequence corresponding to each collection style template conversation to generate a style flow steering vector corresponding to each collection style template conversation includes:

[0015] Obtain the collection template sub-conversation sequence corresponding to the a-th collection style template conversation, where a is not greater than X;

[0016] Extract context semantic nodes from each collection template sub-conversation in the collection template sub-conversation sequence corresponding to the a-th collection style template conversation to generate a semantic connection feature of the context semantic nodes of each collection template sub-conversation;

[0017] Extract a style flow steering vector from each context semantic node according to the semantic jump order of the collection template sub-conversations corresponding to the a-th collection style template conversation and the semantic connection feature of the context semantic nodes of each collection template sub-conversation to generate a style flow steering vector corresponding to the a-th collection style template conversation.

[0018] In a possible implementation of the first aspect, extracting style flow vectors for each context semantic node according to the semantic jump order of the collection template sub-dialogues corresponding to the a-th collection style template dialogue and the semantic connection characteristics of the context semantic nodes of each collection template sub-dialogue, and generating the style flow vector corresponding to the a-th collection style template dialogue includes:

[0019] Extracting style flow vectors for each context semantic node according to the semantic jump order of the collection template sub-dialogues corresponding to the a-th collection style template dialogue and the semantic connection characteristics of the context semantic nodes of each collection template sub-dialogue, and generating an initial style flow vector;

[0020] Performing feature space transformation on the initial style flow vector to generate the style flow vector corresponding to the a-th collection style template dialogue;

[0021] Among them, the step of extracting style flow vectors for each context semantic node according to the semantic jump order of the collection template sub-dialogues corresponding to the a-th collection style template dialogue and the semantic connection characteristics of the context semantic nodes of each collection template sub-dialogue, and generating an initial style flow vector includes:

[0022] For each collection template sub-dialogue corresponding to the a-th collection style template dialogue, regarding its context semantic nodes as nodes in a semantic relationship graph, establishing edges between the nodes according to the semantic connection characteristics of the context semantic nodes, and arranging each of the nodes in the semantic jump order of the collection template sub-dialogue to construct the semantic relationship graph, which is used to reflect the structural relationship at the semantic level of the collection template sub-dialogues corresponding to the a-th collection style template dialogue;

[0023] Performing node clustering analysis on the constructed semantic relationship graph, specifically starting from a target node in the semantic relationship graph, searching for associated nodes whose edge weights with the target node are greater than a set weight and are adjacent in the semantic jump order, marking the target node and the corresponding associated nodes as a potential cluster, and generating a clustered semantic node group;

[0024] For each clustered semantic node group, constructing its corresponding within-group semantic vector, performing between-group relationship analysis on the within-group semantic vectors corresponding to all semantic node groups, and forming a between-group relationship matrix with the relationship measurement values between every two within-group semantic vectors of the semantic node groups, where the between-group relationship matrix is used to reflect the semantic similarity or difference relationship between different semantic node groups;

[0025] Arranging the elements in the between-group relationship matrix in a set order, where the set order is based on the original order of the nodes in the semantic relationship graph or a predefined order;

[0026] Encode the elements in the arranged inter-group relationship matrix to generate an initial style flow vector, which is used to reflect the overall style flow characteristics of the collection template sub-dialogue corresponding to the a-th collection style template dialogue under the semantic jump sequence and semantic connection characteristics, and contains the relationship information between each semantic node group.

[0027] In a possible implementation manner of the first aspect, the X collection style template dialogues are consecutive collection dialogues in the initial collection dialogue flow; the obtaining of the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue includes:

[0028] If a is less than Y, generate the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue according to the first a collection style template dialogues, and there are multiple identical collection style template dialogues in the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue;

[0029] If a is not less than Y, load the (a - Y + 1)-th to the a-th collection style template dialogues in the X collection style template dialogues into the same sequence to generate the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue.

[0030] In a possible implementation manner of the first aspect, the method includes:

[0031] Obtain the sample digital human style portrait vector and the sample collection dialogue flow of the first sample virtual collection digital human. The sample digital human style portrait vector of the first sample virtual collection digital human is generated by extracting the collection style portrait features from the sample historical digital human collection dialogues. The sample historical digital human collection dialogues include the sample collection style templates of the first virtual collection digital human. The digital human style portrait vector is a sample collection style semantic vector for identifying the sample collection style portrait of the virtual collection digital human; the sample collection dialogue flow includes X sample collection style template dialogues, and each sample collection style template dialogue includes the sample collection style template of the first sample virtual collection digital human;

[0032] Extract the style flow vector for each sample collection template sub-dialogue sequence corresponding to each sample collection style template dialogue to generate the sample style flow vector corresponding to each sample collection style template dialogue. Any sample collection template sub-dialogue sequence corresponding to a sample collection style template dialogue includes Y sample collection template sub-dialogues, and each sample collection template sub-dialogue includes the sample collection style template of the second sample virtual collection digital human, where Y is less than X;

[0033] According to the initialized collection style conversion network, perform a first collection style conversion on the example style flow steering vector corresponding to each example collection style template dialogue, the example collection template sub-dialogue sequence corresponding to the example collection style template dialogue, and the example digital person style portrait vector of the first virtual collection digital person, and generate an example collection style conversion result for each example collection style template dialogue;

[0034] Train the initialized collection style conversion network according to the example collection style conversion results of each example collection style template dialogue, and generate a trained collection style conversion network.

[0035] In a possible implementation manner of the first aspect, the training the initialized collection style conversion network according to the example collection style conversion results of each example collection style template dialogue to generate a trained collection style conversion network includes:

[0036] Perform discrimination processing on the example collection style conversion results of each example collection style template dialogue according to the discrimination network, and generate discrimination results of X example collection style conversion results; The discrimination result of any one example collection style conversion result represents whether this collection style conversion result is a conversion dialogue;

[0037] Train the GAN network according to the discrimination results of the X example collection style conversion results and the X example collection style conversion results, and generate a trained collection style conversion network. The GAN network includes the discrimination network and the initialized collection style conversion network.

[0038] In a possible implementation manner of the first aspect, the step of training the GAN network according to the discrimination results of the X example collection style conversion results and the X example collection style conversion results to generate a trained collection style conversion network includes:

[0039] Determine the discrimination cost parameter according to the error between the discrimination result of the b-th collection style conversion result and the discrimination result of the b-th example collection style template dialogue, where b is a positive integer not greater than X;

[0040] Determine the collection style conversion cost parameter according to the error between the b-th example collection style conversion result and the b-th example collection style template dialogue;

[0041] Train the GAN network based on the discrimination cost parameter and the collection style conversion cost parameter.

[0042] In a possible implementation manner of the first aspect, the determining the collection style conversion cost parameter according to the error between the b-th example collection style conversion result and the b-th example collection style template dialogue includes:

[0043] Respectively extract the semantic understanding vectors of the b-th sample collection style conversion result and the b-th sample collection style template dialogue at multiple semantic understanding depths. The semantic understanding vectors at different semantic understanding depths are generated by different semantic understanding units in the semantic understanding model;

[0044] Fuse and calculate the semantic understanding vector errors of the b-th sample collection style conversion result and the b-th sample collection style template dialogue at each semantic understanding depth to generate the semantic understanding cost parameter between the b-th sample collection style conversion result and the b-th sample collection style template dialogue;

[0045] Determine the style semantic cost parameter between the b-th sample collection style conversion result and the b-th sample collection style template dialogue according to the style vector error between the b-th sample collection style conversion result and the b-th sample collection style template dialogue;

[0046] Calculate the collection style portrait cost parameter of the b-th collection style conversion result according to the feature distance between the sample digital human style portrait vector of the b-th sample collection style conversion result and the b-th sample collection style template dialogue;

[0047] Fuse and calculate the semantic understanding cost parameter, the style semantic cost parameter, the collection style portrait cost parameter and the game cost parameter to generate the collection style conversion cost parameter; the game cost parameter is determined according to the discrimination result of the b-th collection style conversion result.

[0048] On the other hand, an embodiment of the present invention further provides a collection dialogue generation system based on AI semantic understanding, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0049] Based on the above aspects, the embodiments of the present application realize the precise control and conversion of the collection dialogue style by obtaining and using the digital human style portrait vector and the initial collection dialogue flow of the first virtual collection digital human. By extracting the style flow steering vector for each collection style template dialogue's collection template sub-dialogue sequence and combining it with the digital human style portrait vector of the first virtual collection digital human, and performing the first collection style conversion through the collection style conversion network, finally generating the target collection dialogue flow that matches the target collection style, which not only improves the generation efficiency of the collection dialogue, but also ensures the accuracy and consistency of the collection dialogue style, thus improving the collection effect and user satisfaction. Brief Description of the Drawings

[0050] Figure 1 is a schematic flowchart of the execution process of the collection dialogue generation method based on AI semantic understanding provided by an embodiment of the present invention.

[0051] Figure 2 is a schematic diagram of the hardware architecture of the collection dialogue generation system based on AI semantic understanding provided by an embodiment of the present invention. Detailed Embodiments

[0052] The present invention will be specifically described below in conjunction with the drawings of the specification. Figure 1 is a schematic flowchart of the collection dialogue generation method based on AI semantic understanding provided by an embodiment of the present invention. The collection dialogue generation method based on AI semantic understanding will be introduced in detail below.

[0053] Step S110: Obtain the digital human style portrait vector and the initial collection dialogue flow of the first virtual collection digital human. The digital human style portrait vector of the first virtual collection digital human is generated by extracting collection style portrait features from historical digital human collection dialogues. The historical digital human collection dialogues include the collection style templates of the first virtual collection digital human. The digital human style portrait vector is a collection style semantic vector used to identify the collection style portrait of the virtual collection digital human. The initial collection dialogue flow includes X collection style template dialogues, and each collection style template dialogue includes the collection style template of the second virtual collection digital human.

[0054] In this embodiment, there is a first virtual collection digital human, which has accumulated a large number of historical digital human collection dialogues in past collection work. The server first needs to obtain the digital human style portrait vector of the first virtual collection digital human. For example, in the historical digital human collection dialogues, this first virtual collection digital human has a unique collection style template. For example, at the beginning of the dialogue, it always politely introduces itself and the purpose of the collection, and maintains a gentle but firm attitude during the communication. It always patiently asks about the reasons for the debtor's overdue payment and expresses understanding. The server will conduct in-depth extraction of collection style portrait features from these historical digital human collection dialogues. It may use natural language processing technology to analyze many elements in the dialogue, such as the words used, the tone, the way of asking questions, etc. For example, if this digital human often uses sentences like "Hello, we noticed that your payment is overdue. Could you please tell us the reason?" which are polite and inquiry-based, this is a style feature. By analyzing a large number of such sentences and their distribution and frequency in the dialogue, the server finally generates a digital human style portrait vector. This vector is like the "style fingerprint" of this first virtual collection digital human and is a collection style semantic vector that can identify its collection style portrait.

[0055] Meanwhile, the server also needs to obtain the initial collection dialogue flow. Assume that this initial collection dialogue flow contains X collection style template dialogues. Each of these collection style template dialogues contains the collection style template of the second virtual collection digital human. For example, these X collection style template dialogues may be divided according to different collection stages or different types of arrears. For example, for debtors with small and short-term overdue payments, there may be a set of collection style template dialogues, and the dialogue content may focus more on the convenience of reminding repayment and the minor impact of overdue; while for debtors with large and long-term overdue payments, there is another set of collection style template dialogues, which will emphasize more on legal risks and credit impacts, etc. These different collection style template dialogues together constitute the initial collection dialogue flow, and the server obtains it completely to prepare for subsequent processing.

[0056] Step S120: Extract the style flow direction vectors for the collection template sub-dialogue sequences corresponding to each collection style template dialogue, generate the style flow direction vectors corresponding to each collection style template dialogue. Any collection template sub-dialogue sequence corresponding to a collection style template dialogue includes Y collection template sub-dialogues, and Y is less than X.

[0057] In this embodiment, the server starts to process each collection style template dialogue in the obtained initial collection dialogue flow. Taking one of the collection style template dialogues as an example, the collection template sub-dialogue sequence corresponding to this dialogue includes Y collection template sub-dialogues (Y is less than X). Assume that this collection style template dialogue is a collection dialogue for customers with medium overdue amount and short overdue time.

[0058] The server first obtains the collection template sub-dialogue sequence corresponding to this collection style template dialogue. For example, the first collection template sub-dialogue in this sequence may be a message sent by the digital human to the debtor to remind repayment, with the content "Dear customer, your repayment date has passed. Please repay as soon as possible to avoid unnecessary fees." The second collection template sub-dialogue may be the debtor's reply, expressing the situation that their funds are temporarily tight.

[0059] Then the server extracts the context semantic nodes for each collection template sub-dialogue in this sequence respectively. For the message sent by the digital human to remind repayment, the server will analyze the semantic nodes in it, such as "repayment date has passed", "repay as soon as possible", "unnecessary fees", etc. There are certain semantic connection characteristics between these nodes. For example, there is a causal relationship between "repayment date has passed" and "repay as soon as possible", because the repayment date has passed so it is necessary to repay as soon as possible; there is also a causal relationship between "repay as soon as possible" and "unnecessary fees", repaying as soon as possible can avoid unnecessary fees. For the debtor's reply, similar semantic node extraction will also be carried out, such as the semantic node of "funds are tight".

[0060] Next, based on the semantic jump order of the collection template sub-dialogues corresponding to this collection style template dialogue and the semantic connection characteristics of the context semantic nodes of each collection template sub-dialogue, the server extracts the style flow direction vectors for each context semantic node. The server first constructs a semantic relationship graph, regarding the context semantic nodes of each collection template sub-dialogue as nodes in the graph, establishing edges between the nodes according to the semantic connection characteristics, and arranging the nodes in the semantic jump order of the collection template sub-dialogues. For example, from the digital human's reminder message for repayment to the debtor's reply, this is a semantic jump order. After constructing the semantic relationship graph, the server performs node clustering analysis. Starting from a target node in the semantic relationship graph, such as starting from the node "The repayment date has passed", searching for associated nodes whose edge weights with it are greater than the set weight and are adjacent in the semantic jump order. Suppose the node "Repay as soon as possible" meets the conditions, then "The repayment date has passed" and "Repay as soon as possible" are marked as a potential cluster, thus generating a group of semantic nodes after clustering.

[0061] For each group of semantic nodes after clustering, the server constructs its corresponding intra-group semantic vector. For example, for the group of semantic nodes containing "The repayment date has passed" and "Repay as soon as possible", an intra-group semantic vector that can reflect the semantic relationship between them is constructed. Then, an inter-group relationship analysis is performed on the intra-group semantic vectors corresponding to all groups of semantic nodes, and the relationship measurement values between every two intra-group semantic vectors of the groups of semantic nodes are formed into an inter-group relationship matrix. This matrix reflects the semantic similarity or difference relationship between different groups of semantic nodes.

[0062] After that, the server arranges the elements in the inter-group relationship matrix in a set order, which can be based on the original order of the nodes in the semantic relationship graph or a predefined order. Finally, an encoding operation is performed on the elements in the arranged inter-group relationship matrix to generate an initial style flow direction vector. This initial style flow direction vector reflects the overall style flow characteristics of the collection template sub-dialogues corresponding to this collection style template dialogue under the semantic jump order and semantic connection characteristics, and contains the relationship information between each group of semantic nodes. However, this initial style flow direction vector may still need further processing. The server performs a feature space transformation on this initial style flow direction vector, and finally generates the style flow direction vector corresponding to this collection style template dialogue. For each collection style template dialogue in the initial collection dialogue flow, the server repeats this process to obtain the style flow direction vector corresponding to each collection style template dialogue.

[0063] Step S130: Perform a first collection style conversion on the style flow transition vector corresponding to each collection style template dialogue, the collection template sub-dialogue sequence corresponding to the collection style template dialogue, and the digital human style portrait vector of the first virtual collection digital human, to generate the collection style conversion result of each collection style template dialogue.

[0064] After the server obtains the style flow transition vector corresponding to each collection style template dialogue, it begins to perform the first collection style conversion through the collection style conversion network. Taking one of the collection style template dialogues as an example, assume that this collection style template dialogue is for those debtors who are resistant to collection.

[0065] First, the server encodes and represents the collection template sub-dialogue sequence corresponding to this collection style template dialogue. For example, this collection template sub-dialogue sequence contains the content of the digital human's multiple attempts to communicate with the debtor about repayment matters, but the debtor always makes excuses to shirk. The server will convert each sub-dialogue content in this sequence into a potential style vector through a specific encoding method. This encoding process may involve converting the text content into digital representations and considering factors such as the position and semantics of each sub-dialogue in the entire sequence. This potential style vector can reflect the internal style characteristics of the collection template sub-dialogue sequence corresponding to this collection style template dialogue to a certain extent.

[0066] Then, the server integrates the style flow transition vector corresponding to this collection style template dialogue with the digital human style portrait vector of the first virtual collection digital human. This integration process is like fusing the style flow characteristics of this collection style template dialogue with the overall style characteristics of the first virtual collection digital human. For example, the style of the first virtual collection digital human is relatively gentle and good at guiding, while the style flow transition vector of this collection style template dialogue shows a relatively tough but lack of guiding style. Through the integration operation, the characteristics of both can complement and adjust each other.

[0067] Next, the server aggregates the integrated vector of this collection style template dialogue and the previously generated potential style vector. This aggregation process synthesizes the multi-faceted characteristic information of this collection style template dialogue to form a more comprehensive aggregated vector.

[0068] Finally, the server decodes and represents this aggregated vector. This decoding process is the inverse operation of the encoding process, converting the aggregated vector into a debt collection style conversion result that can be understood by humans. This result may be an adjusted debt collection conversation content. For example, in the original debt collection conversation, the digital human may simply ask the debtor to repay the debt. After the style conversion, the debt collection style conversion result may be that while asking the debtor to repay the debt, the digital human also provides some feasible debt repayment plan suggestions and is more gentle and guiding in tone to better deal with debtors who resist debt collection. For each debt collection style template conversation in the initial debt collection conversation flow, the server repeats this process to obtain the debt collection style conversion result of each debt collection style template conversation.

[0069] Step S140: Generate a target debt collection conversation flow corresponding to the initial debt collection conversation flow based on the debt collection style conversion results of the X debt collection style template conversations.

[0070] After obtaining the debt collection style conversion results of the X debt collection style template conversations, the server begins to generate the target debt collection conversation flow. Suppose these X debt collection style template conversations cover various debt collection conversation types from the initial debt collection reminder to multiple debt collection follow-ups and for different debtor situations.

[0071] The server combines these X debt collection style conversion results in a certain logical order. For example, in chronological order of debt collection, first place the debt collection style conversion result for debtors with the first overdue payment at the front, and then successively place the debt collection style conversion results for debtors with multiple overdue payments and increasing debt amounts. This combination method makes the entire debt collection conversation flow more coherent and reasonable logically.

[0072] At the same time, the server also considers the connection between each debt collection style conversion result. For example, if the previous debt collection style conversion result mentions that the debtor will be contacted again at a certain time, then the opening of the next debt collection style conversion result should echo this commitment. In this way, the server organically combines each debt collection style conversion result to form a complete target debt collection conversation flow. Compared with the initial debt collection conversation flow, this target debt collection conversation flow is more unified and effective in debt collection style, can better adapt to different debtor situations, and thus improves the success rate of debt collection.

[0073] In this process, the server will also perform a final check and optimization on the target collection dialogue flow. For example, it checks for semantic incoherence or unclear expressions. If it is found that the collection style conversion result in a certain place is not very coordinated with the overall style, the server may make fine-tuning to ensure the consistency and effectiveness of the overall target collection dialogue flow in terms of style. The finally generated target collection dialogue flow can be used in actual collection work, and the virtual collection digital human interacts with the debtor according to this dialogue flow, improving the efficiency and success rate of the collection work.

[0074] Based on the above steps, the embodiment of this application realizes the accurate control and conversion of the collection dialogue style by obtaining and using the digital human style portrait vector of the first virtual collection digital human and the initial collection dialogue flow. By extracting the style flow steering vector for each collection template sub-dialogue sequence of the collection style template dialogue and combining it with the digital human style portrait vector of the first virtual collection digital human, the first collection style conversion is performed through the collection style conversion network, and finally a target collection dialogue flow matching the target collection style is generated. This not only improves the generation efficiency of the collection dialogue, but also ensures the accuracy and consistency of the collection dialogue style, thus enhancing the collection effect and user satisfaction.

[0075] In a possible implementation manner, step S130 includes:

[0076] Step S131, encoding and representing the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue to generate the potential style vector corresponding to the a-th collection style template dialogue, where a is a positive integer not greater than X.

[0077] Step S132, integrating the style flow steering vector corresponding to the a-th collection style template dialogue with the digital human style portrait vector of the first virtual collection digital human to generate the integrated vector of the a-th collection style template dialogue.

[0078] Step S133, aggregating the integrated vector of the a-th collection style template dialogue and the potential style vector corresponding to the a-th collection style template dialogue to generate the aggregated vector of the a-th collection style template dialogue.

[0079] Step S134, decoding and representing the aggregated vector of the a-th collection style template dialogue to generate the collection style conversion result of the a-th collection style template dialogue.

[0080] Taking the processing of the a-th collection style template dialogue as an example, where a is a positive integer not greater than X.

[0081] First, the server encodes the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue to generate the corresponding potential style vector. Suppose the a-th collection style template dialogue is a collection dialogue for those debtors with a long overdue period and a large amount of debt. The corresponding collection template sub-dialogue sequence contains multiple collection interaction contents. The server will conduct a detailed analysis and digital processing on each sub-dialogue content in this sequence. This process is like converting each dialogue into a special digital code according to certain rules. For example, factors such as the debt amount number mentioned in the sub-dialogue, the description of the overdue period, and the tone words of the repayment requirements are all given specific digital representations. Also, the order of the sub-dialogues in the entire sequence will be taken into account because the impact of sub-dialogues in different orders on the overall style is different. Through such a complex processing process, the server finally generates the potential style vector corresponding to the a-th collection style template dialogue, which reflects the style characteristics of this collection template sub-dialogue sequence at the digital level.

[0082] Next, the server integrates the style flow vector corresponding to the a-th collection style template dialogue with the digital human style portrait vector of the first virtual collection digital human. Still taking this collection style template dialogue for large overdue debtors as an example, the style flow vector of the a-th collection style template dialogue may contain the unique communication style trend of this dialogue. For example, during multiple collections, it gradually changes from a gentle reminder to a more serious warning. The digital human style portrait vector of the first virtual collection digital human represents its overall collection style characteristics. For example, this digital human is usually relatively rational and good at persuading debtors with data. The server integrates these two vectors, which is like integrating the style trend of this specific collection style template dialogue with the overall style characteristics of the digital human. This integration process is not a simple addition but an organic combination according to the relationships between the elements in the vectors, thus generating the integrated vector of the a-th collection style template dialogue. This integrated vector combines the style trend of the specific dialogue and the overall style characteristics of the digital human.

[0083] The server then aggregates the integrated vector of the ath collection style template conversation and the previously generated latent style vector. Continuing with the example of the collection conversation with a large overdue debtor, the integrated vector contains information after style fusion, while the latent style vector reflects the stylistic characteristics of the collection template sub-conversation sequence itself. The server aggregates these two vectors, which is like comprehensively summarizing the style information obtained from analyzing this conversation from different angles. During the aggregation process, the server calculates and combines the corresponding elements in the two vectors according to a pre-defined algorithm, fully considering the meaning and importance of each element, and ultimately generates an aggregate vector for the ath collection style template conversation. This aggregate vector contains the most comprehensive style information obtained from analyzing this collection style template conversation from all aspects.

[0084] Finally, the server decodes the aggregate vector of the ath collection style template dialogue to generate the collection style conversion result. Using this collection scenario as an example, the server reverses the encoding process according to the rules followed earlier, converting the numerical information in the aggregate vector into human-readable collection dialogue content. This conversion process considers how the various elements in the vector correspond to the actual collection dialogue's wording, tone, and logical structure. For example, if an element in the aggregate vector indicates a more serious collection attitude, the decoded collection style conversion result will reflect the use of relatively strong language to demand repayment from the debtor. After this decoding process, the server ultimately generates the collection style conversion result for the ath collection style template dialogue. This result represents the collection dialogue content after style conversion, which is more suitable for the specific debtor's situation and improves the effectiveness of collection. The server repeats this process for each collection style template dialogue, thus obtaining the collection style conversion result for each dialogue.

[0085] In a possible implementation, step S120 includes:

[0086] Step S121: Obtain the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue, where a is not greater than X.

[0087] Step S122 , extracting contextual semantic nodes from each collection template sub-dialogue in the collection template sub-dialogue sequence corresponding to the ath collection style template dialogue, and generating semantic connection features of the contextual semantic nodes of each collection template sub-dialogue.

[0088] Step S123: Based on the semantic jump order of the collection template sub-dialogue corresponding to the a-th collection style template dialogue and the semantic connection features of the context semantic nodes of each collection template sub-dialogue, style flow vectors are extracted from each context semantic node to generate the style flow vector corresponding to the a-th collection style template dialogue.

[0089] In a possible implementation, step S123 includes:

[0090] Step S1231, according to the semantic jump sequence of the collection template sub-dialogue corresponding to the a-th collection style template dialogue and the semantic connection characteristics of the context semantic nodes of each collection template sub-dialogue, extract the style flow direction vectors for each context semantic node to generate initial style flow direction vectors.

[0091] Step S1232, perform a feature space transformation on the initial style flow direction vectors to generate the style flow direction vectors corresponding to the a-th collection style template dialogue.

[0092] Among them, step S1231 includes:

[0093] Step S1231-1, for each collection template sub-dialogue corresponding to the a-th collection style template dialogue, regard its context semantic nodes as nodes in the semantic relationship graph, establish edges between the nodes according to the semantic connection characteristics of the context semantic nodes, and arrange each of the nodes in the semantic jump sequence of the collection template sub-dialogue to construct the semantic relationship graph, which is used to reflect the structural relationship at the semantic level of the collection template sub-dialogue corresponding to the a-th collection style template dialogue.

[0094] Step S1231-2, perform node clustering analysis on the constructed semantic relationship graph. Specifically, start from a target node in the semantic relationship graph, search for associated nodes whose edge weights with the target node are greater than the set weight and are adjacent in the semantic jump sequence, mark the target node and the corresponding associated nodes as a potential cluster, and generate a clustered group of semantic nodes.

[0095] Step S1231-3, for each clustered group of semantic nodes, construct its corresponding within-group semantic vector, perform an inter-group relationship analysis on the within-group semantic vectors corresponding to all groups of semantic nodes, and form a matrix of inter-group relationship measures from the relationship measure values between every two within-group semantic vectors of the groups of semantic nodes. The matrix of inter-group relationship measures is used to reflect the semantic similarity or difference relationship between different groups of semantic nodes.

[0096] Step S1231-4, arrange the elements in the matrix of inter-group relationship measures in a set order, and the set order is based on the original order of the nodes in the semantic relationship graph or a predefined order.

[0097] Step S1231-5: Encoding the elements in the arranged inter-group relationship matrix to generate an initial style flow vector, which is used to reflect the overall style flow characteristics of the collection template sub-dialogue corresponding to the a-th collection style template dialogue under the semantic jump sequence and semantic connection characteristics, and contains the relationship information between each semantic node group.

[0098] In this embodiment, taking the processing of the a-th collection style template dialogue as an example, where a is not greater than X. First, the server needs to obtain the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue. Suppose the a-th collection style template dialogue is a collection dialogue for customers with medium overdue time and medium amount of arrears. The collection template sub-dialogue sequence corresponding to this collection style template dialogue contains a series of interaction processes, such as from the initial repayment reminder, to the customer's feedback of financial difficulties, and then to the collection party's proposal of some repayment suggestions and other sub-dialogues. These sub-dialogues form the entire collection template sub-dialogue sequence in a certain order, and the server accurately obtains this sequence to prepare for subsequent analysis.

[0099] Next, the server extracts the context semantic nodes of each collection template sub-dialogue in the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue, and generates the semantic connection characteristics of the context semantic nodes of each collection template sub-dialogue. Still taking the above collection dialogue for customers with medium overdue and medium amount of arrears as an example, for the collection template sub-dialogue of the initial repayment reminder, the server will identify the key semantic elements therein as semantic nodes, such as "repayment date", "overdue situation", "amount due for repayment", etc. There are semantic connection characteristics between these semantic nodes. For example, there is a temporal correlation between "repayment date" and "overdue situation" because the overdue situation occurs because the repayment date has passed; there is a causal relationship between "overdue situation" and "amount due for repayment" because due to the overdue, the amount to be repaid includes additional parts such as overdue fees. For the sub-dialogue where the customer feedbacks financial difficulties, the server will also identify semantic nodes such as "financial difficulties" and "temporarily unable to repay", and analyze the semantic connection characteristics between them and the semantic nodes of the previous sub-dialogues. For example, "financial difficulties" is an explanation for the inability to pay the "amount due for repayment" on time.

[0100] Then, based on the semantic jump sequence of the collection template sub-dialogue corresponding to the a-th collection style template dialogue and the semantic connection characteristics of the context semantic nodes of each collection template sub-dialogue, the server extracts the style flow vector for each context semantic node to generate the style flow vector corresponding to the a-th collection style template dialogue. This process is relatively complex and is as follows:

[0101] First, according to the semantic jump sequence of the collection template sub-dialogues corresponding to the a-th collection style template dialogue and the semantic connection characteristics of the context semantic nodes of each collection template sub-dialogue, extract the style flow turning vectors for each context semantic node to generate the initial style flow turning vectors. For each collection template sub-dialogue corresponding to the a-th collection style template dialogue, regard its context semantic nodes as nodes in the semantic relationship graph, establish edges between the nodes according to the semantic connection characteristics of the context semantic nodes, and arrange each node according to the semantic jump sequence of the collection template sub-dialogue to construct the semantic relationship graph. Taking the above collection dialogue as an example, in the sub-dialogue of repayment reminder, semantic nodes such as "repayment date", "overdue situation", and "amount due for repayment" form a semantic relationship graph. The edge between "repayment date" and "overdue situation" represents their association relationship and is arranged according to the order of this sub-dialogue. For the sub-dialogue where the customer feedbacks financial difficulties, semantic nodes such as "financial difficulties" and "temporarily unable to repay" also form the corresponding semantic relationship graph, and the order of this sub-dialogue in the entire collection template sub-dialogue sequence determines the order relationship between its semantic relationship graph and the semantic relationship graph of the previous repayment reminder sub-dialogue. This semantic relationship graph reflects the structural relationship of the collection template sub-dialogues corresponding to the a-th collection style template dialogue at the semantic level.

[0102] Next, perform node clustering analysis on the constructed semantic relationship graph. Specifically, start from a target node in the semantic relationship graph, search for associated nodes whose edge weights with the target node are greater than the set weight and are adjacent in the semantic jump sequence, and mark the target node and the corresponding associated nodes as a potential cluster to generate the clustered semantic node group. Taking the semantic relationship graph of the repayment reminder sub-dialogue as an example, starting from the target node "overdue situation", if the edge weight between the associated node "amount due for repayment" and "overdue situation" is greater than the set weight (this weight may be set according to the tightness of semantic association, for example, if the semantic association between two nodes is very direct, the weight is high), and they are adjacent in the semantic jump sequence (in this sub-dialogue, the overdue situation is mentioned first, and then the amount due for repayment will be involved), then mark "overdue situation" and "amount due for repayment" as a potential cluster, thus generating the clustered semantic node group. The same operation is also performed on the semantic relationship graph of the sub-dialogue where the customer feedbacks financial difficulties.

[0103] Then, for each clustered semantic node group, construct its corresponding intra-group semantic vector. For the clustered semantic node group containing "overdue situation" and "amount due", the server constructs an intra-group semantic vector that can reflect the semantic relationship between these two nodes. This vector may contain some numerical values, which represent certain quantitative features of these two nodes in the semantic relationship, such as the tightness of semantics, the importance weight of semantics, etc. Conduct an inter-group relationship analysis on the intra-group semantic vectors corresponding to all semantic node groups, and form an inter-group relationship matrix by combining the relationship measurement values between the intra-group semantic vectors of every two semantic node groups. For example, conduct a relationship measurement between the intra-group semantic vector of the semantic node group in the repayment reminder sub-dialogue and the intra-group semantic vector of the semantic node group in the customer feedback on financial difficulties sub-dialogue. This measurement may be to calculate the semantic similarity or difference between them, and then form a matrix of these measurement values into an inter-group relationship matrix. This inter-group relationship matrix is used to reflect the semantic similarity or difference relationship between different semantic node groups.

[0104] After that, arrange the elements in the inter-group relationship matrix in a set order. This set order is based on the original order of the nodes in the semantic relationship graph or a predefined order. Taking the above example, if it is based on the original order of the nodes in the semantic relationship graph, then arrange the elements in the inter-group relationship matrix in the order where the semantic nodes in the repayment reminder sub-dialogue come first and the semantic nodes in the customer feedback on financial difficulties sub-dialogue come later.

[0105] Finally, perform an encoding operation on the elements in the arranged inter-group relationship matrix to generate an initial style flow steering vector. This initial style flow steering vector is used to reflect the overall style flow characteristics of the collection template sub-dialogue corresponding to the a-th collection style template dialogue under the semantic jump order and semantic connection characteristics, and contains the relationship information between each semantic node group. This encoding operation may be to convert the elements in the matrix into a vector form according to certain rules. This vector can represent the style flow of the entire collection style template dialogue at the semantic level, such as the style flow information from initially emphasizing repayment facts (repayment date, amount due, etc.) to the coping style after the customer feedback on difficulties (how to handle the customer's financial difficulties situation).

[0106] After generating the initial style transition vector, the server also needs to perform a feature space transformation on this initial style transition vector to generate the style transition vector corresponding to the a-th collection style template dialogue. This feature space transformation process may involve readjusting, mapping, or transforming the elements in the initial style transition vector to make it more compliant with the requirements of the entire collection style analysis system. For example, based on the analysis results of a large number of collection conversations before, some features that are more representative or differentiating under a specific style may be determined, and then the elements in the initial style transition vector are transformed into this new feature space, finally obtaining the style transition vector corresponding to the a-th collection style template dialogue. This style transition vector can accurately reflect the semantic and style characteristics of the a-th collection style template dialogue, providing an important basis for subsequent operations such as collection style conversion. For each collection style template dialogue, the server will repeat the entire above operation process to obtain the style transition vector corresponding to each collection style template dialogue.

[0107] In a possible implementation manner, the X collection style template dialogues are consecutive collection dialogues in the initial collection dialogue flow. Step S121 includes:

[0108] Step S1211, if a is less than Y, generate the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue according to the first a collection style template dialogues, and there are multiple identical collection style template dialogues in the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue.

[0109] Step S1212, if a is not less than Y, load the (a - Y + 1)-th collection style template dialogue to the a-th collection style template dialogue among the X collection style template dialogues into the same sequence to generate the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue. [[ID=...]]

[0110] Taking the conversation of the a-th collection style template as an example, when a is less than Y, the server generates the collection template sub-conversation sequence corresponding to the a-th collection style template based on the first a collection style template conversations. Suppose this is the collection process for debtors with a relatively small overdue amount and a relatively short overdue time. The consecutive collection style template conversations in the initial collection conversation flow have a certain logical order and association. For example, the first 3 (assuming a = 3 and Y is greater than 3) collection style template conversations may all be relatively gentle reminder-type conversations. The server integrates these 3 collection style template conversations to construct the collection template sub-conversation sequence corresponding to the 3rd collection style template conversation. Since it is constructed based on the first a conversations, there may be multiple identical collection style template conversations in this sequence. In this sequence, there will be repeated gentle reminder content, such as mentioning the repayment date, the number of overdue days, and the benefits of repaying on time each time. These repeated parts reflect the strategy of continuously gently reminding such debtors in the initial stage of collection. The server generates the collection template sub-conversation sequence corresponding to the a-th collection style template in this way.

[0111] When a is not less than Y, the server loads the (a - Y + 1)-th collection style template conversation to the a-th collection style template conversation among the X collection style template conversations into the same sequence to generate the collection template sub-conversation sequence corresponding to the a-th collection style template. Suppose Y = 5, when a = 7, the server will load the 3rd (7 - 5 + 1 = 3) collection style template conversation to the 7th collection style template conversation into the same sequence. If this is the collection process for debtors with an increasing overdue amount and a longer overdue time, then this sequence covers the collection conversations from an earlier stage to the current one. The earlier collection conversations may be gentle reminders, and as the overdue situation worsens, the subsequent collection conversations may gradually increase the urging intensity, mentioning possible overdue fees, credit impacts, etc. The server integrates these collection style template conversations at different stages into one sequence, which completely reflects the development and changes of the collection process for such debtors, thereby generating the collection template sub-conversation sequence corresponding to the a-th collection style template. In this way, the server can accurately construct the collection template sub-conversation sequence corresponding to each collection style template conversation for different collection situations, providing a basis for subsequent collection style analysis and conversion operations.

[0112] In a possible implementation manner, the method includes:

[0113] Step S101: Obtain the sample digital human style portrait vector and the sample collection conversation flow of the first sample virtual collection digital human. The sample digital human style portrait vector of the first sample virtual collection digital human is generated by extracting the collection style portrait feature from the sample historical digital human collection conversations. The sample historical digital human collection conversations include the sample collection style templates of the first virtual collection digital human. The digital human style portrait vector is a sample collection style semantic vector used to identify the sample collection style portrait of the virtual collection digital human. The sample collection conversation flow includes X sample collection style template conversations, and each sample collection style template conversation includes the sample collection style template of the first sample virtual collection digital human.

[0114] Step S102: Extract the style flow transition vectors for the sample collection template sub-conversation sequences corresponding to each sample collection style template conversation, and generate the sample style flow transition vectors corresponding to each sample collection style template conversation. The sample collection template sub-conversation sequence corresponding to any sample collection style template conversation includes Y sample collection template sub-conversations, and each sample collection template sub-conversation includes the sample collection style template of the second sample virtual collection digital human, where Y is less than X.

[0115] Step S103: Perform the first collection style conversion on the sample style flow transition vector corresponding to each sample collection style template conversation, the sample collection template sub-conversation sequence corresponding to this sample collection style template conversation, and the sample digital human style portrait vector of the first virtual collection digital human according to the initialized collection style conversion network, and generate the sample collection style conversion results of each sample collection style template conversation.

[0116] Step S104: Train the initialized collection style conversion network according to the sample collection style conversion results of each sample collection style template conversation to generate the trained collection style conversion network.

[0117] In a possible implementation manner, step S104 includes:

[0118] Step S1041: Perform identification processing on the sample collection style conversion results of each sample collection style template conversation according to the identification network, and generate the identification results of the X sample collection style conversion results. The identification result of any sample collection style conversion result indicates whether this collection style conversion result is a converted conversation.

[0119] Step S1042: Train the GAN network according to the identification results of the X sample collection style conversion results and the X sample collection style conversion results to generate the trained collection style conversion network. The GAN network includes the identification network and the initialized collection style conversion network.

[0120] In this embodiment, first, the server needs to obtain the sample digital human style portrait vector and the sample collection conversation flow of the first sample virtual collection digital human. Taking the handling of overdue collection of small loans in a financial institution as an example, the first sample virtual collection digital human has accumulated rich collection experience in past sample historical digital human collection conversations. These sample historical digital human collection conversations contain the sample collection style templates of the first virtual collection digital human. For example, this sample collection style template may be to politely address the debtor at the beginning of the conversation, then briefly explain the debt situation, and maintain a professional and gentle attitude throughout the conversation. The server will extract the collection style portrait features from these sample historical digital human collection conversations to generate the sample digital human style portrait vector of the first sample virtual collection digital human. This vector is a sample collection style semantic vector that can identify the sample collection style portrait of this virtual digital human. At the same time, the sample collection conversation flow obtained by the server contains X sample collection style template conversations, and each sample collection style template conversation contains the sample collection style template of the first sample virtual collection digital human. These sample collection style template conversations may be collection conversations for debtors with different overdue durations or different debt amounts. For example, one conversation template for debtors with an overdue duration of 1 - 7 days and a debt amount of 1000 - 5000 yuan, and another conversation template for debtors with an overdue duration of 8 - 15 days and a debt amount of 5000 - 10000 yuan, etc.

[0121] Next, the server extracts the style flow transition vectors for the sample collection template sub-conversation sequences corresponding to each sample collection style template conversation to generate the sample style flow transition vectors corresponding to each sample collection style template conversation. Any sample collection template sub-conversation sequence corresponding to a sample collection style template conversation includes Y sample collection template sub-conversations (Y is less than X). For example, for the sample collection style template conversation for debtors with an overdue duration of 1 - 7 days and a debt amount of 1000 - 5000 yuan, its sample collection template sub-conversation sequence may contain 3 sample collection template sub-conversations (assuming Y = 3). The first sub-conversation may be a repayment reminder sent by the digital human, containing information such as the debt amount and the number of overdue days; the second sub-conversation may be the debtor's reply, expressing their repayment difficulties; the third sub-conversation may be some repayment suggestions given by the digital human in response to the debtor's reply. The server will analyze each sub-conversation in this sequence, extract the context semantic nodes, determine the semantic connection features, construct a semantic relationship graph, perform node clustering analysis, construct an intra-group semantic vector, analyze the inter-group relationship, arrange the elements and encode, etc. A series of operations are finally used to generate the sample style flow transition vector corresponding to this sample collection style template conversation. For each sample collection style template conversation, the server will repeat such operations.

[0122] Then, based on the initialized collection style conversion network, the server performs a first collection style conversion on the example style flow direction vector corresponding to each example collection style template dialogue, the example collection template sub-dialogue sequence corresponding to the example collection style template dialogue, and the example digital human style portrait vector of the first virtual collection digital human, and generates the example collection style conversion result of each example collection style template dialogue. Taking one of the example collection style template dialogues as an example, the initialized collection style conversion network will comprehensively consider the style flow reflected by the example style flow direction vector of this dialogue, the specific content in the example collection template sub-dialogue sequence, and the overall style characteristics represented by the example digital human style portrait vector. For example, if the example style flow direction vector shows that the original style of this dialogue is relatively rigid, while the example digital human style portrait vector represents a gentle style, the initialized collection style conversion network may adjust the dialogue content to make the tone more gentle and guiding, so as to generate the example collection style conversion result. This result may be a converted collection dialogue content that is more in line with the ideal collection style.

[0123] After that, the server trains the initialized collection style conversion network based on the example collection style conversion results of each example collection style template dialogue, and generates the trained collection style conversion network. First, the discrimination network performs discrimination processing on the example collection style conversion results of each example collection style template dialogue, and generates the discrimination results of X example collection style conversion results. The role of this discrimination network is to judge whether each example collection style conversion result is a valid conversion dialogue. For example, for a certain example collection style conversion result, if it is semantically incoherent or the style is far from the target style, the discrimination network may determine that it is not a valid conversion dialogue. Then, based on the discrimination results of these X example collection style conversion results and the X example collection style conversion results, the GAN network (including the discrimination network and the initialized collection style conversion network) is trained. Taking a specific example, for an example collection style conversion result determined by the discrimination network as an invalid conversion, the initialized collection style conversion network will be adjusted according to this discrimination result, such as adjusting the weights or algorithm logic when processing the example style flow direction vector, the example collection template sub-dialogue sequence, and the example digital human style portrait vector. At the same time, the discrimination network will also continuously optimize its discrimination criteria based on more example data. Through continuous adjustment and optimization, the finally generated trained collection style conversion network can perform collection style conversion more accurately, improving the quality and effect of collection dialogues.

[0124] In a possible implementation manner, step S1042 includes:

[0125] Step S1042-1: Determine the discrimination cost parameter based on the error between the discrimination result of the b-th collection style conversion result and the discrimination result of the b-th sample collection style template dialogue, where b is a positive integer not greater than X.

[0126] Step S1042-2: Determine the collection style conversion cost parameter based on the error between the b-th sample collection style conversion result and the b-th sample collection style template dialogue.

[0127] Step S1042-3: Train the GAN network based on the discrimination cost parameter and the collection style conversion cost parameter.

[0128] In a possible implementation, Step S1042-2 includes:

[0129] Step S1042-21: Respectively extract the semantic understanding vectors of the b-th sample collection style conversion result and the b-th sample collection style template dialogue at multiple semantic understanding depths according to the semantic understanding model. The semantic understanding vectors at different semantic understanding depths are generated by different semantic understanding units in the semantic understanding model.

[0130] Step S1042-22: Perform a fusion calculation on the semantic understanding vector errors of the b-th sample collection style conversion result and the b-th sample collection style template dialogue at each semantic understanding depth to generate the semantic understanding cost parameter between the b-th sample collection style conversion result and the b-th sample collection style template dialogue.

[0131] Step S1042-23: Determine the style semantic cost parameter between the b-th sample collection style conversion result and the b-th sample collection style template dialogue based on the style vector error between the b-th sample collection style conversion result and the b-th sample collection style template dialogue.

[0132] Step S1042-24: Calculate the collection style portrait cost parameter of the b-th collection style conversion result based on the feature distance between the sample digital human style portrait vector of the b-th sample collection style conversion result and the b-th sample collection style template dialogue.

[0133] Step S1042-25: Perform a fusion calculation on the semantic understanding cost parameter, the style semantic cost parameter, the collection style portrait cost parameter and the game cost parameter to generate the collection style conversion cost parameter. The game cost parameter is determined based on the discrimination result of the b-th collection style conversion result.

[0134] First, the server determines the discrimination cost parameter based on the error between the discrimination result of the b-th collection style conversion result and the discrimination result of the b-th sample collection style template dialogue (where b is a positive integer not greater than X). Taking the previous example of overdue collection for microloans, assume that the b-th sample collection style template dialogue is a collection dialogue for debtors with overdue days of 1 - 7 and an outstanding amount of 1000 - 5000 yuan. Its original collection style template dialogue has a specific discrimination result. For example, it is determined to be highly effective because it is good in aspects such as semantic expression and style compliance. However, the discrimination result of the b-th collection style conversion result after conversion may be different from the original in the discrimination network. If the original is determined to be very compliant (e.g., with a score of 0.9, where the score is a hypothetical value representing the degree of compliance), and the converted result is determined to be 0.7 (indicating relatively worse), then the error (0.2) between them will be used to determine the discrimination cost parameter. This discrimination cost parameter reflects the degree of difference in the discrimination results between the conversion result and the original template dialogue, and is a quantitative manifestation of the conversion effect at the discrimination level.

[0135] Next, the server determines the collection style conversion cost parameter based on the error between the b-th sample collection style conversion result and the b-th sample collection style template dialogue. This process is relatively complex and involves multiple aspects.

[0136] First, semantic understanding vectors at multiple semantic understanding depths of the b-th sample collection style conversion result and the b-th sample collection style template dialogue are respectively extracted according to the semantic understanding model. The semantic understanding vectors at different semantic understanding depths are generated by different semantic understanding units in the semantic understanding model. For example, at a relatively shallow semantic understanding depth, it may mainly focus on the understanding at the lexical level. For some key words in the collection dialogue, such as "overdue", "repayment", "outstanding amount", etc., the semantic understanding model will generate corresponding semantic understanding vectors. At this level, there may be differences in the semantic understanding vectors of these key words in the b-th sample collection style conversion result and the b-th sample collection style template dialogue. For example, the semantic understanding vector of the word "overdue" in the original template dialogue represents a clear calculation method of overdue days in a certain dimension, while the semantic understanding vector of this word in the conversion result may deviate in this dimension. At a deeper semantic understanding depth, it may involve the understanding of sentence structure and logical relationships. For example, there is a causal sentence structure in the original template dialogue (because of overdue, additional fees need to be paid), and its semantic understanding vector will reflect this relationship. If the expression of this relationship in the conversion result is unclear or changed, the semantic understanding vector will also reflect this difference.

[0137] Then, the semantic understanding vector errors of the collection style conversion result of the b-th example and the collection style template dialogue of the b-th example at each semantic understanding depth are fused and calculated to generate the semantic understanding cost parameter between the collection style conversion result of the b-th example and the collection style template dialogue of the b-th example. This fusion calculation process may comprehensively consider the importance weights of different semantic understanding depths. For example, a deeper semantic understanding depth may be given a higher weight because it is more crucial for grasping the overall semantics. Suppose the error at the lexical level is 0.1 and the error at the sentence structure level is 0.3 (the values here are illustrative of the error magnitudes), and through weighted fusion (assuming the weight at the lexical level is 0.3 and the weight at the sentence structure level is 0.7), the semantic understanding cost parameter is obtained.

[0138] Next, based on the style vector error between the collection style conversion result of the b-th example and the collection style template dialogue of the b-th example, the style semantic cost parameter between the collection style conversion result of the b-th example and the collection style template dialogue of the b-th example is determined. For example, the original collection style template dialogue of the example has a gentle and professional style vector representation, where certain dimensions of the style vector represent characteristics such as the politeness of the words used and the mildness of the tone. If the collection style conversion result of the b-th example after conversion has changed in style, for example, if there is a large deviation in the dimension of the politeness of the words used, such as originally using very polite words and becoming more blunt after conversion, then the error in this dimension will be calculated into the style vector error, and then the style semantic cost parameter is determined.

[0139] Then, based on the feature distance between the example digital human style portrait vector of the collection style conversion result of the b-th example and the collection style template dialogue of the b-th example, the collection style portrait cost parameter of the collection style conversion result of the b-th example is calculated. The example digital human style portrait vector contains the comprehensive characteristics of the entire collection style. For example, the overall attitude tendency of the digital human in the collection process, the interaction style with the debtor, and other characteristics. If the distance between the example digital human style portrait vector of the conversion result and the original collection style template dialogue of the b-th example in these characteristics is large, such as the attitude tendency changing from gentle to relatively tough, this feature distance will be quantified as the collection style portrait cost parameter.

[0140] Finally, the semantic understanding cost parameter, the style semantic cost parameter, the collection style portrait cost parameter, and the game cost parameter are fused and calculated to generate the collection style conversion cost parameter. The game cost parameter here is determined based on the discrimination result of the b-th collection style conversion result. For example, if the discrimination result of the b-th collection style conversion result is poor (indicating a bad conversion effect), then the game cost parameter may be relatively high, which will have a greater impact on the final collection style conversion cost parameter during the fusion calculation. This fusion calculation may assign different weights according to the importance of each parameter, and comprehensively obtain a parameter that comprehensively reflects the conversion cost between the b-th sample collection style conversion result and the b-th sample collection style template dialogue.

[0141] Based on the discrimination cost parameter and the collection style conversion cost parameter, the server trains the GAN network. During the training process, these two cost parameters are like compasses guiding the network to adjust. If the discrimination cost parameter is relatively high, it indicates that there is a large deviation in the discrimination network's judgment of the conversion result, then the discrimination network part in the GAN network will adjust its discrimination criteria and algorithm logic. If the collection style conversion cost parameter is relatively high, it indicates that there is a large error in the conversion network's conversion of the sample collection style template dialogue into the sample collection style conversion result, then the initialized collection style conversion network part in the GAN network will adjust its conversion method, such as adjusting the processing methods of the style flow steering vector, the sample collection template sub-dialogue sequence, and the sample digital human style portrait vector, etc. By continuously adjusting the GAN network according to these cost parameters, a trained collection style conversion network is finally generated, which can perform collection style conversion more accurately and improve the quality and effect of collection conversations.

[0142] Figure 2 The hardware structure diagram of the collection conversation generation system 100 based on AI semantic understanding provided by the embodiment of the present invention for implementing the above-mentioned collection conversation generation method based on AI semantic understanding is shown. As Figure 2 shown, the collection conversation generation system 100 based on AI semantic understanding may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0143] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from external terminals. In some embodiments, the machine-readable storage medium 120 can store the data and / or instructions used by the collection conversation generation system 100 based on AI semantic understanding to execute or use to complete the exemplary methods described in the present invention.

[0144] In a specific implementation process, one or more processors 110 execute computer-executable instructions stored in a machine-readable storage medium 120, enabling the processors 110 to execute the collection conversation generation method based on AI semantic understanding in the above method embodiments. The processors 110, the machine-readable storage medium 120, and the communication unit 140 are connected through a bus 130. The processors 110 can be used to control the sending and receiving actions of the communication unit 140.

[0145] For the specific implementation process of the processors 110, reference can be made to the respective method embodiments executed by the above collection conversation generation system 100 based on AI semantic understanding. Their implementation principles and technical effects are similar, and will not be elaborated herein.

[0146] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the collection conversation generation method based on AI semantic understanding as described above is implemented.

[0147] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the previous description of the embodiments of the present invention, multiple features are sometimes merged into one embodiment, drawing, or description thereof.

Claims

1. A debt collection dialogue generation method based on AI semantic understanding, characterized in that: The method comprises: Obtaining a digital human style portrait vector and an initial collection dialogue flow for a first virtual collection digital human. The digital human style portrait vector for the first virtual collection digital human is generated by extracting collection style portrait features from historical digital human collection dialogues, wherein the historical digital human collection dialogues contain a collection style template for the first virtual collection digital human. The digital human style portrait vector is a collection style semantic vector used to identify the collection style portrait of the virtual collection digital human. The initial collection dialogue flow includes X collection style template dialogues, each of which contains a collection style template for the second virtual collection digital human. Perform style flow vector extraction on the collection template sub-dialogue sequence corresponding to each collection style template dialogue to generate a style flow vector corresponding to each collection style template dialogue. The collection template sub-dialogue sequence corresponding to any collection style template dialogue includes Y collection template sub-dialogues, where Y is less than X. Performing a first debt collection style conversion on the style flow vector corresponding to each debt collection style template dialogue, the debt collection template sub-dialogue sequence corresponding to the debt collection style template dialogue, and the digital human style portrait vector of the first virtual debt collection digital human through a debt collection style conversion network to generate a debt collection style conversion result for each debt collection style template dialogue; Generating a target collection dialogue flow corresponding to the initial collection dialogue flow based on the collection style conversion results of the X collection style template dialogues; The step of extracting style flow vectors from the collection template sub-dialogue sequence corresponding to each collection style template dialogue to generate style flow vectors corresponding to each collection style template dialogue includes: Get the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue, where a is not greater than X; Extracting contextual semantic nodes from each collection template sub-dialogue in the collection template sub-dialogue sequence corresponding to the ath collection style template dialogue, and generating semantic connection features of the contextual semantic nodes of each collection template sub-dialogue; For each collection template sub-dialogue corresponding to the ath collection style template dialogue, its contextual semantic nodes are treated as nodes in a semantic relationship graph. Based on the semantic connection characteristics of the contextual semantic nodes, edges are established between the nodes. The nodes are arranged according to the semantic jump order of the collection template sub-dialogue to construct the semantic relationship graph. The semantic relationship is used to reflect the structural relationship of the collection template sub-dialogue corresponding to the ath collection style template dialogue at the semantic level. Perform node clustering analysis on the constructed semantic relationship graph. Specifically, starting from a target node in the semantic relationship graph, search for associated nodes whose edge weight with the target node is greater than the set weight and adjacent in the semantic jump order. Mark the target node and the corresponding associated nodes as a potential cluster, and generate a clustered semantic node group. For each clustered semantic node group, construct its corresponding intra-group semantic vector, perform inter-group relationship analysis on the intra-group semantic vectors corresponding to all semantic node groups, and form an inter-group relationship matrix based on the relationship measurement values between the intra-group semantic vectors of every two semantic node groups. The inter-group relationship matrix is used to reflect the semantic similarity or difference between different semantic node groups. Arranging the elements in the inter-group relationship matrix according to a set order, wherein the set order is based on the original order or a predefined order of the nodes in the semantic relationship graph; Performing an encoding operation on the elements in the arranged inter-group relationship matrix to generate an initial style flow vector. The initial style flow vector is used to reflect the overall style flow characteristics of the collection template sub-dialogue corresponding to the a-th collection style template dialogue under the semantic jump order and semantic connection characteristics, and includes the relationship information between each semantic node group; Perform feature space transformation on the initial style flow vector to generate a style flow vector corresponding to the ath debt collection style template dialogue.

2. The debt collection dialogue generation method based on AI semantic understanding according to claim 1 is characterized in that: The step of performing a first debt collection style conversion on the style flow vector corresponding to each debt collection style template dialogue, the debt collection template sub-dialogue sequence corresponding to the debt collection style template dialogue, and the digital human style portrait vector of the first virtual debt collection digital human through the debt collection style conversion network to generate a debt collection style conversion result for each debt collection style template dialogue includes: Encode the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue to generate a latent style vector corresponding to the a-th collection style template dialogue, where a is a positive integer not greater than X; Integrating the style flow vector corresponding to the ath debt collection style template dialogue with the digital human style portrait vector of the first virtual debt collection digital human to generate an integrated vector of the ath debt collection style template dialogue; Aggregating the integrated vector of the ath debt collection style template dialogue and the latent style vector corresponding to the ath debt collection style template dialogue to generate an aggregated vector of the ath debt collection style template dialogue; The aggregate vector of the a-th debt collection style template dialogue is decoded and represented to generate a debt collection style conversion result of the a-th debt collection style template dialogue.

3. The debt collection dialogue generation method based on AI semantic understanding according to claim 1 is characterized in that: The X collection style template dialogues are consecutive collection dialogues in the initial collection dialogue flow; obtaining the collection template sub-dialogue sequence corresponding to the ath collection style template dialogue includes: If a is less than Y, then a collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue is generated based on the first a collection style template dialogues, and the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue contains multiple identical collection style template dialogues; If a is not less than Y, then the a-Y+1th collection style template dialogue to the ath collection style template dialogue among the X collection style template dialogues are loaded into the same sequence to generate a collection template sub-dialogue sequence corresponding to the ath collection style template dialogue.

4. The method for generating debt collection dialogues based on AI semantic understanding according to any one of claims 1 to 3, characterized in that: The method comprises: Obtain a sample digital human style portrait vector and a sample collection dialogue flow of a first sample virtual collection digital human. The sample digital human style portrait vector of the first sample virtual collection digital human is generated by extracting collection style portrait features from sample historical digital human collection dialogues. The sample historical digital human collection dialogues contain a sample collection style template of the first virtual collection digital human. The digital human style portrait vector is a sample collection style semantic vector used to identify the sample collection style portrait of the virtual collection digital human. The sample collection dialogue flow includes X sample collection style template dialogues, each of which contains the sample collection style template of the first sample virtual collection digital human. Perform style flow vector extraction on the sample collection template sub-dialogue sequence corresponding to each sample collection style template dialogue to generate a sample style flow vector corresponding to each sample collection style template dialogue, wherein the sample collection template sub-dialogue sequence corresponding to any sample collection style template dialogue includes Y sample collection template sub-dialogues, each sample collection template sub-dialogue includes the sample collection style template of the second sample virtual collection digital person, and Y is less than X; performing a first collection style conversion on the sample style flow vector corresponding to each sample collection style template dialogue, the sample collection template sub-dialogue sequence corresponding to the sample collection style template dialogue, and the sample digital human style portrait vector of the first virtual collection digital human based on the initialized collection style conversion network, thereby generating a sample collection style conversion result for each sample collection style template dialogue; The initialized debt collection style conversion network is trained according to the sample debt collection style conversion results of each sample debt collection style template dialogue to generate a trained debt collection style conversion network.

5. The method for generating debt collection dialogue based on AI semantic understanding according to claim 4 is characterized in that: The initializing collection style conversion network is trained based on the sample collection style conversion results of each sample collection style template dialogue to generate a trained collection style conversion network, including: The sample collection style conversion results of each sample collection style template dialogue are identified based on the identification network to generate identification results of X sample collection style conversion results; the identification result of any sample collection style conversion result indicates whether the collection style conversion result is a conversion dialogue; The GAN network is trained based on the identification results of the X sample debt collection style conversion results and the X sample debt collection style conversion results to generate a trained debt collection style conversion network, wherein the GAN network includes the identification network and the initialized debt collection style conversion network.

6. The method for generating debt collection dialogue based on AI semantic understanding according to claim 5 is characterized in that: The step of training a GAN network based on the identification results of the X sample debt collection style conversion results and the X sample debt collection style conversion results to generate a trained debt collection style conversion network includes: Determine the identification cost parameter based on the error between the identification result of the bth debt collection style conversion result and the identification result of the bth sample debt collection style template dialogue, where b is a positive integer not greater than X; Determining a collection style conversion cost parameter based on an error between the bth sample collection style conversion result and the bth sample collection style template dialogue; The GAN network is trained based on the identification cost parameter and the collection style conversion cost parameter.

7. The method for generating debt collection dialogue based on AI semantic understanding according to claim 6 is characterized in that: Determining a cost parameter of the debt collection style conversion based on an error between the bth sample debt collection style conversion result and the bth sample debt collection style template dialogue includes: Extracting semantic understanding vectors of the bth sample debt collection style conversion result and the bth sample debt collection style template dialogue at multiple semantic understanding depths based on the semantic understanding model, wherein the semantic understanding vectors at different semantic understanding depths are generated by different semantic understanding units in the semantic understanding model; Performing a fusion calculation on the semantic understanding vector error of the bth sample collection style conversion result and the bth sample collection style template dialogue at each semantic understanding depth to generate a semantic understanding cost parameter between the bth sample collection style conversion result and the bth sample collection style template dialogue; Determining a style semantic cost parameter between the bth sample collection style conversion result and the bth sample collection style template dialogue based on a style vector error between the bth sample collection style conversion result and the bth sample collection style template dialogue; Calculating a collection style portrait cost parameter of the bth collection style conversion result based on a feature distance between the sample digital human style portrait vector of the bth sample collection style conversion result and the bth sample collection style template dialogue; The semantic understanding cost parameter, the style semantic cost parameter, the collection style portrait cost parameter and the game cost parameter are fused and calculated to generate a collection style conversion cost parameter; the game cost parameter is determined based on the identification result of the bth collection style conversion result.

8. A debt collection dialogue generation system based on AI semantic understanding, characterized by: The debt collection dialogue generation system based on AI semantic understanding includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the debt collection dialogue generation method based on AI semantic understanding as described in any one of claims 1 to 7 above.

Images