AI semantic understanding-based collection dialogue generation method and system

Abstract

According to the AI semantic understanding-based collection dialogue generation method and system provided by the invention, the digital person style portrait vector of the first virtual collection digital person and the initial collection dialogue stream are acquired and utilized, so that the collection dialogue style is accurately controlled and converted. According to the method, style flow vector extraction is carried out on a collection template sub-dialogue sequence of each collection style template dialogue, a digital person style portrait vector of a first virtual collection digital person is combined, first collection style conversion is carried out through a collection style conversion network, and finally a target collection dialogue flow matched with a target collection style is generated. The generation efficiency of the collection dialogue is improved, and the accuracy and consistency of the collection dialogue style are ensured, so that the collection effect and the user satisfaction are improved.

Description

Technical Field

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

[0002] In the financial services industry, debt collection is an important part of ensuring the return of funds and reducing the risk of bad debts. Traditional debt collection methods mostly rely on manual labor, which is not only costly, but also difficult to achieve efficient and consistent debt collection results due to multiple factors such as personnel quality and emotional management. 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 debt collection field.

[0003] In the existing technology, there have been attempts to use AI technology to assist in debt collection, such as through automated voice response systems, text message collection, etc., but most of these methods are based on fixed scripts or templates, lack flexibility and personalization, and are difficult to adapt to the actual situation and psychological state of different debtors. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for generating a debt collection dialogue based on AI semantic understanding, the method comprising: Obtaining a digital human style portrait vector of a first virtual debt collection digital human and an initial debt collection dialogue flow, wherein the digital human style portrait vector of the first virtual debt collection digital human is generated by extracting the characteristics of the debt collection style portrait of a historical digital human debt collection dialogue, wherein the historical digital human debt collection dialogue includes a debt collection style template of the first virtual debt collection digital human, and the digital human style portrait vector is a debt collection style semantic vector used to identify the debt collection style portrait of the virtual debt collection digital human; the initial debt collection dialogue flow includes X debt collection style template dialogues, each of which includes the debt collection style template of the second virtual debt collection digital human; 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, wherein 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 collection style conversion on the style flow 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 through a collection style conversion network to generate a collection style conversion result for each collection style template dialogue; A target collection dialogue flow corresponding to the initial collection dialogue flow is generated according to the collection style conversion results of the X collection style template dialogues.

[0005] In a possible implementation of the first aspect, the step of performing a first collection style conversion on the style flow 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 through a collection style conversion network to generate a collection style conversion result for each collection style template dialogue includes: Encoding the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue to generate a potential style vector corresponding to the a-th collection style template dialogue, where a is a positive integer not greater than X; Integrate the style flow vector corresponding to the a-th 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 a-th debt collection style template dialogue; Aggregating the integrated vector of the a-th debt collection style template dialogue and the potential style vector corresponding to the a-th debt collection style template dialogue to generate an aggregated vector of the a-th 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.

[0006] In a possible implementation of the first aspect, 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; 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 vector extraction is performed on each context semantic node to generate the style flow vector corresponding to the a-th collection style template dialogue.

[0007] In a possible implementation of the first aspect, extracting style flow vectors from each context semantic node based on the semantic jump order of the collection template sub-dialog corresponding to the a-th collection style template dialogue and the semantic connection features of the context semantic nodes of each collection template sub-dialog to generate the style flow vector corresponding to the a-th collection style template dialogue includes: According to the semantic jump order of the collection template sub-dialog corresponding to the ath collection style template dialogue and the semantic connection features of the context semantic nodes of each collection template sub-dialog, style flow vector extraction is performed on each context semantic node to generate an initial style flow vector; Performing feature space transformation on the initial style flow vector to generate a style flow vector corresponding to the ath debt collection style template dialogue; The step of extracting style flow vectors from each context semantic node based on the semantic jump order of the collection template sub-dialog corresponding to the ath collection style template dialogue and the semantic connection features of the context semantic nodes of each collection template sub-dialog to generate an initial style flow vector comprises: For each collection template sub-dialogue corresponding to the a-th collection style template dialogue, its context semantic nodes are regarded as nodes in the semantic relationship graph, edges are established between the nodes according to the semantic connection characteristics of the context semantic nodes, and the nodes are arranged according to the semantic jump order of the collection template sub-dialogue to construct the semantic relationship graph, and the semantic relationship is used to reflect the structural relationship of the collection template sub-dialogue corresponding to the a-th collection style template dialogue at the semantic level; 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 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 with the relationship measurement values ​​between the semantic vectors of every two semantic node groups. The inter-group relationship matrix is ​​used to reflect the semantic similarity or difference relationship 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; The elements in the arranged inter-group relationship matrix are encoded 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 order and semantic connection characteristics, and includes the relationship information between each semantic node group.

[0008] In a possible implementation of the first aspect, the X collection style template dialogues are continuous collection dialogues in the initial collection dialogue flow; and obtaining the collection template sub-dialogue sequence corresponding to the a-th 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 there are multiple identical collection style template dialogues in the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue; 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.

[0009] In a possible implementation of the first aspect, the method includes: Obtaining a sample digital human style portrait vector and a sample collection dialogue flow of a first sample virtual collection digital human, wherein the sample digital human style portrait vector of the first sample virtual collection digital human is generated by extracting collection style portrait features of a sample historical digital human collection dialogue, wherein the sample historical digital human collection dialogue contains a sample collection style template of the first virtual collection digital human, and the digital human style portrait vector is a sample collection style semantic vector used to identify a sample collection style portrait of a 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; Performing 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 according to the initialized collection style conversion network, and generating a sample collection style conversion result for each sample collection style template dialogue; The initialized collection style conversion network is trained according to the sample collection style conversion results of each sample collection style template dialogue to generate a trained collection style conversion network.

[0010] In a possible implementation of the first aspect, the training of the initialized collection style conversion network according to the sample collection style conversion results of each sample collection style template dialogue to generate a trained collection style conversion network includes: The sample collection style conversion results of each sample collection style template dialogue are subjected to identification processing according to the identification network, and identification results of X sample collection style conversion results are generated; the identification result of any sample collection style conversion result indicates whether the collection style conversion result is a conversion dialogue; According to the identification results of the X sample collection style conversion results and the X sample collection style conversion results, the GAN network is trained to generate a trained collection style conversion network, wherein the GAN network includes the identification network and the initialized collection style conversion network.

[0011] In a possible implementation of the first aspect, the step of training a GAN network based on the identification results of the X sample collection style conversion results and the X sample collection style conversion results to generate a trained collection style conversion network includes: Determine the identification cost parameter according to the error between the identification result of the b-th collection style conversion result and the identification result of the b-th sample collection style template dialogue, where b is a positive integer not greater than X; Determining a collection style conversion cost parameter according to an error between a b-th sample collection style conversion result and the b-th sample collection style template dialogue; Based on the identification cost parameter and the collection style conversion cost parameter, the GAN network is trained.

[0012] In a possible implementation of the first aspect, determining the collection style conversion cost parameter according to an error between the b-th sample collection style conversion result and the b-th sample collection style template dialogue includes: Extracting 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 respectively according to 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 b-th sample collection style conversion result and the b-th sample collection style template dialogue at each semantic understanding depth, and generating a semantic understanding cost parameter between the b-th sample collection style conversion result and the b-th 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 according to 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 b-th collection style conversion result according to a feature distance between a sample digital human style portrait vector of the b-th sample collection style conversion result and the b-th 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.

[0013] On the other hand, an embodiment of the present invention also provides a collection dialogue generation system based on AI semantic understanding, including a processor and a machine-readable storage medium, wherein 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.

[0014] Based on the above aspects, the embodiment of the present application achieves accurate control and conversion of the collection dialogue style by acquiring and utilizing the digital human style portrait vector and the initial collection dialogue flow of the first virtual collection digital human. By performing style flow vector extraction on the collection template sub-dialogue sequence of each collection style template dialogue, and combining 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, which not only improves the generation efficiency of the collection dialogue, but also ensures the accuracy and consistency of the collection dialogue style, thereby improving the collection effect and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the execution flow of the collection dialogue generation method based on AI semantic understanding provided by an embodiment of the present invention.

[0016] Figure 2 It is a schematic diagram of the hardware architecture of a debt collection dialogue generation system based on AI semantic understanding provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a method for generating a collection dialogue based on AI semantic understanding provided by an embodiment of the present invention. The method for generating a collection dialogue based on AI semantic understanding is introduced in detail below.

[0018] Step S110, obtaining a digital human style portrait vector of a first virtual debt collection digital human and an initial debt collection dialogue flow, wherein the digital human style portrait vector of the first virtual debt collection digital human is generated by extracting the characteristics of the debt collection style portrait of a historical digital human debt collection dialogue, wherein the historical digital human debt collection dialogue contains the debt collection style template of the first virtual debt collection digital human, and the digital human style portrait vector is a debt collection style semantic vector used to identify the debt collection style portrait of the virtual debt collection digital human. The initial debt collection dialogue flow includes X debt collection style template dialogues, each of which contains the debt collection style template of the second virtual debt collection digital human.

[0019] In this embodiment, there is a first virtual collection digital person, which has accumulated a large number of historical digital person collection dialogues in the past collection work. The server first needs to obtain the digital person style portrait vector of the first virtual collection digital person. For example, in the historical digital person collection dialogue, this first virtual collection digital person has a unique collection style template, such as always politely introducing himself and the purpose of collection at the beginning of the dialogue, and maintaining a gentle but firm attitude during the communication process, and always patiently asking and expressing understanding for the reasons for the overdue debtor. The server will perform in-depth collection style portrait feature extraction on these historical digital person collection dialogues. It may analyze many elements such as the words, tone, and questioning methods in the dialogue through natural language processing technology. For example, if this digital person often uses polite and inquiring sentences such as "Hello, we noticed that your repayment is overdue, can you tell me the reason?" 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 person style portrait vector. This vector is like the "style fingerprint" of the first virtual debt collection digital person, and is a collection style semantic vector that can identify its collection style portrait.

[0020] At the same time, 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 collection style template dialogue here contains the collection style template of the second virtual collection digital person. For example, these X collection style template dialogues may be divided according to different collection stages or different types of debts. For example, for small-amount short-term overdue debtors, there may be a set of collection style template dialogues, and the content of the dialogue may focus more on reminding the convenience of repayment and the slight impact of overdue; while for large-amount long-term overdue debtors, there is another set of collection style template dialogues, which will emphasize legal risks and credit impact. These different collection style template dialogues together constitute the initial collection dialogue flow, and the server obtains it in full to prepare for subsequent processing.

[0021] Step S120, extracting style flow vectors from the collection template sub-dialogue sequence corresponding to each collection style template dialogue, generating style flow vectors corresponding to each collection style template dialogue, wherein 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.

[0022] In this embodiment, the server starts to process each collection style template dialogue in the acquired 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 contains Y collection template sub-dialogues (Y is less than X). Assume that this collection style template dialogue is a collection dialogue for customers who are recently overdue and have a medium amount of debt.

[0023] The server first obtains the collection template sub-dialogue sequence corresponding to the collection style template dialogue. For example, the first collection template sub-dialogue in this sequence may be a message sent by the digital person to the debtor to remind him to repay the debt, which reads "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 his temporary financial constraints.

[0024] The server then extracts contextual semantic nodes for each collection template sub-dialogue in this sequence. For the repayment reminder message sent by the digital human, the server will analyze the semantic nodes therein, such as "the repayment date has passed", "repay as soon as possible", "unnecessary expenses", etc. There are certain semantic connection features between these nodes. For example, there is a causal relationship between "the 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 expenses", and repaying as soon as possible can avoid unnecessary expenses. Similar semantic node extraction will also be performed for the debtor's reply, such as the semantic node of "tight funds".

[0025] Next, the server extracts style flow vectors from each contextual semantic node based on the semantic jump order of the collection template sub-dialogue corresponding to the collection style template dialogue and the semantic connection features of the contextual semantic nodes of each collection template sub-dialogue. The server first constructs a semantic relationship graph, regards the contextual semantic nodes of each collection template sub-dialogue as nodes in the graph, establishes edges between nodes based on the semantic connection features, and arranges each node according to the semantic jump order of the collection template sub-dialogue. For example, from the reminder message of the digital person 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 the node "repayment date has passed", search for associated nodes whose edge weights are greater than the set weight and are adjacent in the semantic jump order. Assuming that the node "repay as soon as possible" meets the conditions, then "repayment date has passed" and "repay as soon as possible" are marked as a potential cluster, thus generating a clustered semantic node group.

[0026] For each clustered semantic node group, the server constructs its corresponding intra-group semantic vector. For example, for the semantic node group containing "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, the inter-group relationship analysis is performed on the intra-group semantic vectors corresponding to all semantic node groups, and the relationship measurement values ​​between the semantic vectors of each two semantic node groups are combined into an inter-group relationship matrix. This matrix reflects the semantic similarity or difference relationship between different semantic node groups.

[0027] Afterwards, 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, the elements in the arranged inter-group relationship matrix are encoded to generate an initial style flow vector. This initial style flow vector reflects the overall style flow characteristics of the collection template sub-dialogue corresponding to this collection style template dialogue under the semantic jump order and semantic connection characteristics, and contains the relationship information between each semantic node group. However, this initial style flow vector may need further processing. The server will perform feature space transformation on this initial style flow vector and finally generate the style flow vector corresponding to this collection style template dialogue. For each collection style template dialogue in the initial collection dialogue flow, the server will repeat this process to obtain the style flow vector corresponding to each collection style template dialogue.

[0028] Step S130, performing a first collection style conversion on the style flow 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 through the collection style conversion network to generate a collection style conversion result for each collection style template dialogue.

[0029] After obtaining the style flow vector corresponding to each collection style template dialogue, the server starts to perform the first collection style conversion through the collection style conversion network. Taking one of the collection style template dialogues as an example, it is assumed that this collection style template dialogue is a dialogue for debtors who are more resistant to collection.

[0030] First, the server encodes the collection template sub-dialogue sequence corresponding to the collection style template dialogue. For example, this collection template sub-dialogue sequence contains dialogue content in which the digital person repeatedly tries to communicate with the debtor about repayment, but the debtor always shirks the responsibility for various reasons. The server converts each sub-dialogue content in this sequence into a latent style vector through a specific encoding method. This encoding process may involve converting text content into digital representation, and taking into account factors such as the position and semantics of each sub-dialogue in the entire sequence. This latent style vector can reflect the intrinsic style characteristics of the collection template sub-dialogue sequence corresponding to this collection style template dialogue to a certain extent.

[0031] Then, the server integrates the style flow vector corresponding to the 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 the 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 vector of the collection style template dialogue reflects a relatively tough but lacking guiding style. Through the integration operation, the characteristics of the two can complement and adjust each other.

[0032] Next, the server aggregates the integrated vector of the collection style template dialogue and the previously generated latent style vector. This aggregation process combines the multi-faceted feature information of the collection style template dialogue to form a more comprehensive aggregate vector.

[0033] Finally, the server decodes the aggregate vector. The decoding process is the inverse of the encoding process, converting the aggregate vector into a collection style conversion result that can be understood by humans. This result may be an adjusted collection dialogue content. For example, in the original collection dialogue, the digital person may simply ask the debtor to repay the debt, but after the style conversion, the collection style conversion result may be that the digital person not only asks for repayment, but also provides some feasible repayment plan suggestions, and the tone is gentler and more guiding, so as to better deal with those debtors who resist collection. For each collection style template dialogue in the initial collection dialogue flow, the server will repeat this process to obtain the collection style conversion result of each collection style template dialogue.

[0034] Step S140: generating a target collection dialogue flow corresponding to the initial collection dialogue flow according to the collection style conversion results of the X collection style template dialogues.

[0035] After obtaining the collection style conversion results of X collection style template dialogues, the server starts to generate the target collection dialogue flow. Assume that these X collection style template dialogues cover various collection dialogue types from initial collection reminders to multiple collection follow-ups and for different debt situations.

[0036] The server will combine these X collection style conversion results in a certain logical order. For example, according to the time sequence of collection, the collection style conversion results for the first overdue debtors are placed first, followed by the collection style conversion results for the debtors who have been overdue for many times and the amount of debts is gradually increasing. This combination method can make the entire collection dialogue flow more logically coherent and reasonable.

[0037] At the same time, the server will also consider the connectivity between the various collection style conversion results. For example, if the previous collection style conversion result mentioned that the debtor would be contacted again at a certain time, then the opening of the next collection style conversion result should echo this promise. In this way, the server organically combines the various collection style conversion results to form a complete target collection dialogue flow. Compared with the initial collection dialogue flow, this target collection dialogue flow is more unified and effective in collection style, and can better adapt to the situations of different debtors, thereby improving the success rate of collection.

[0038] During this process, the server will also conduct a final check and optimization of the target collection dialogue flow. For example, check whether there are any semantic incoherence or unclear expressions. If it is found that the collection style conversion result of a certain place is not in harmony with the overall style, the server may fine-tune it to ensure the consistency and effectiveness of the style of the entire target collection dialogue flow. The target collection dialogue flow finally generated can be used in actual collection work, and the virtual collection digital person will interact with the debtor according to this dialogue flow to improve the efficiency and success rate of the collection work.

[0039] Based on the above steps, the embodiment of the present application achieves accurate control and conversion of the collection dialogue style by acquiring and utilizing the digital human style portrait vector and the initial collection dialogue flow of the first virtual collection digital human. By performing style flow vector extraction on the collection template sub-dialogue sequence of each collection style template dialogue, and combining 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, which not only improves the generation efficiency of the collection dialogue, but also ensures the accuracy and consistency of the collection dialogue style, thereby improving the collection effect and user satisfaction.

[0040] In a possible implementation, step S130 includes: Step S131, encoding and representing the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue, and generating a potential style vector corresponding to the a-th collection style template dialogue, where a is a positive integer not greater than X.

[0041] Step S132: Integrate the style flow vector corresponding to the a-th 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 a-th debt collection style template dialogue.

[0042] Step S133 , aggregating the integrated vector of the a th debt collection style template dialogue and the latent style vector corresponding to the a th debt collection style template dialogue to generate an aggregated vector of the a th debt collection style template dialogue.

[0043] Step S134, decoding the aggregate vector of the a-th debt collection style template dialogue to generate a debt collection style conversion result of the a-th debt collection style template dialogue.

[0044] Taking the processing of the ath debt collection style template dialogue as an example, a here is a positive integer not greater than X.

[0045] First, the server encodes the collection template sub-dialogue sequence corresponding to the ath collection style template dialogue to generate the corresponding latent style vector. Assuming that the ath collection style template dialogue is a collection dialogue for debtors with a long overdue period and a large amount of debt, its corresponding collection template sub-dialogue sequence contains multiple collection interaction contents. The server will conduct a detailed analysis and digital processing of each sub-dialogue content in this sequence. This process is like converting each dialogue into a special string of digital codes according to certain rules. For example, the amount of debt mentioned in the sub-dialogue, the description of the overdue period, the tone and words of the repayment request, etc. are all given specific digital representations. In addition, the order of the sub-dialogue in the entire sequence will also be taken into account, because the influence of sub-dialogues in different orders on the overall style is different. Through such a complex processing process, the server finally generates the latent style vector corresponding to the ath collection style template dialogue, which reflects the style characteristics of this collection template sub-dialogue sequence from a digital level.

[0046] Next, the server integrates the style flow vector corresponding to the ath collection style template dialogue with the digital human style portrait vector of the first virtual collection digital human. Taking the collection style template dialogue for large overdue debtors as an example, the style flow vector of the ath collection style template dialogue may contain the unique communication style trend of this dialogue, such as the style flow that gradually changes from gentle reminders to more serious warnings during multiple collection processes. The digital human style portrait vector of the first virtual collection digital human represents its overall collection style characteristics, such as the style of this digital human who is usually more rational and good at using data to persuade debtors. The server integrates these two vectors, just 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 based on the relationship between the elements in the vector, thereby generating an integrated vector of the ath collection style template dialogue. This integrated vector combines the style flow of a specific dialogue and the overall style characteristics of the digital human.

[0047] Then, the server aggregates the integrated vector of the ath collection style template dialogue and the latent style vector generated previously. Continuing with the collection dialogue of a large overdue debtor as an example, the integrated vector contains the information after style fusion, and the latent style vector reflects the style characteristics of the collection template sub-dialogue sequence itself. The server aggregates these two vectors, just like comprehensively summarizing the style information obtained from analyzing this dialogue from different angles. During the aggregation process, the server will calculate and combine the corresponding elements in the two vectors according to a pre-set algorithm, fully considering the meaning and importance of each element, and finally generate the aggregate vector of the ath collection style template dialogue. This aggregate vector contains the most comprehensive style information obtained from analyzing this collection style template dialogue from all aspects.

[0048] Finally, the server decodes the aggregate vector of the ath collection style template dialogue to generate the collection style conversion result. Still taking this collection scenario as an example, the server will perform the reverse operation according to the rules followed by the previous encoding process to convert the digital information in the aggregate vector into the collection dialogue content that humans can understand. This conversion process needs to consider how the various elements in the vector correspond to the actual collection dialogue words, tone, logical structure, etc. For example, if an element in the aggregate vector represents a more serious collection attitude, then in the decoded collection style conversion result, it will be reflected as the use of tougher words to require the debtor to repay. After such a decoding process, the server finally generates the collection style conversion result of the ath collection style template dialogue. This result is a collection dialogue content after style conversion, which can better adapt to the situation of a specific debtor and improve the effectiveness of collection. For each collection style template dialogue, the server will repeat this operation to obtain the collection style conversion result of each dialogue.

[0049] In a possible implementation, step S120 includes: Step S121, obtaining the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue, where a is not greater than X.

[0050] 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.

[0051] 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 vector extraction is performed on each context semantic node to generate the style flow vector corresponding to the a-th collection style template dialogue.

[0052] In a possible implementation, step S123 includes: Step S1231, based on the semantic jump order of the collection template sub-dialog corresponding to the ath collection style template dialogue and the semantic connection features of the context semantic nodes of each collection template sub-dialog, style flow vectors are extracted for each context semantic node to generate an initial style flow vector.

[0053] Step S1232, performing feature space transformation on the initial style flow turning vector to generate a style flow turning vector corresponding to the ath debt collection style template dialogue.

[0054] Wherein, step S1231 includes: Step S1231-1, for each collection template sub-dialogue corresponding to the a-th collection style template dialogue, regard its context semantic node as a node 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 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 at the semantic level of the collection template sub-dialogue corresponding to the a-th collection style template dialogue.

[0055] Step S1231-2, perform 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 the 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.

[0056] Step S1231-3, 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 with the relationship measurement values ​​between the semantic vectors of every two semantic node groups. The inter-group relationship matrix is ​​used to reflect the semantic similarity or difference relationship between different semantic node groups.

[0057] Step S1231 - 4 , 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.

[0058] Step S1231-5, 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 order and semantic connection characteristics, and includes the relationship information between each semantic node group.

[0059] In this embodiment, the processing of the ath collection style template dialogue is taken 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 ath collection style template dialogue. Assume that the ath collection style template dialogue is a collection dialogue for customers with medium overdue time and medium amount of debt. The collection template sub-dialogue sequence corresponding to this collection style template dialogue includes a series of interactive processes, such as the initial repayment reminder, to the customer's feedback on financial difficulties, and then to the collection party's proposal of some repayment suggestions and other sub-dialogues. These sub-dialogues constitute the entire collection template sub-dialogue sequence in a certain order. The server accurately obtains this sequence to prepare for subsequent analysis.

[0060] Next, the server extracts 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 generates semantic connection features of contextual semantic nodes of each collection template sub-dialogue. Taking the above collection dialogue for customers with medium overdue and arrears as an example, for the initial repayment reminder collection template sub-dialogue, the server will identify the key semantic elements therein as semantic nodes, such as "repayment date", "overdue situation", "amount to be repaid", etc. There are semantic connection features between these semantic nodes, such as there is a temporal correlation between "repayment date" and "overdue situation", because the overdue situation occurs only after the repayment date has passed; there is a causal relationship between "overdue situation" and "amount to be repaid", because the amount to be repaid due to the overdue payment includes additional parts such as overdue fees. For the sub-dialogue in which the customer reports financial difficulties, the server will also identify semantic nodes such as "financial difficulties" and "temporarily unable to repay", and analyze the semantic connection features between them and between them and the semantic nodes of the previous sub-dialogue. For example, "financial difficulties" is an explanation for the inability to pay the "amount due" on time.

[0061] Then, based on the semantic jump order of the collection template sub-dialogue corresponding to the ath collection style template dialogue and the semantic connection features of the context semantic nodes of each collection template sub-dialogue, the server extracts the style flow vector of each context semantic node to generate the style flow vector corresponding to the ath collection style template dialogue. This process is relatively complicated, as follows: First, according to the semantic jump order of the collection template sub-dialogue corresponding to the ath 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 for each context semantic node to generate an initial style flow vector. For each collection template sub-dialogue corresponding to the ath collection style template dialogue, its context semantic nodes are regarded as nodes in the semantic relationship graph, and according to the semantic connection features of the context semantic nodes, edges are established between the nodes, and the nodes are arranged according to the semantic jump order of the collection template sub-dialogue to construct a semantic relationship graph. Taking the above collection dialogue as an example, in the repayment reminder sub-dialogue, the semantic nodes "repayment date", "overdue situation" and "amount to be repaid" are constructed into a semantic relationship graph, and the edge between "repayment date" and "overdue situation" represents their association relationship, and they are arranged according to the order of this sub-dialogue. For the sub-dialogue of customer feedback on financial difficulties, semantic nodes such as "financial difficulties" and "temporarily unable to repay" are also constructed into corresponding semantic relationship graphs, and the order of this sub-dialogue in the entire collection template sub-dialogue sequence determines its order relationship with the semantic relationship graph of the previous repayment reminder sub-dialogue. This semantic relationship graph reflects the structural relationship of the collection template sub-dialogue corresponding to the a-th collection style template dialogue at the semantic level.

[0062] Next, the constructed semantic relationship graph is subjected to node clustering analysis. Specifically, starting from a target node in the semantic relationship graph, the associated nodes whose edge weight with the target node is greater than the set weight and adjacent in the semantic jump order are searched, and the target node and the corresponding associated nodes are marked as a potential cluster to generate a 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 to be repaid" and "overdue situation" is greater than the set weight (this weight may be set according to the closeness of the semantic association, for example, if the semantic association between the two nodes is very direct, the weight is high), and they are adjacent in the semantic jump order (in this sub-dialogue, the overdue situation is mentioned first, and the amount to be repaid is immediately involved), then "overdue situation" and "amount to be repaid" are marked as a potential cluster, thus generating a clustered semantic node group. The same operation is performed on the semantic relationship graph of the sub-dialogue of customer feedback on financial difficulties.

[0063] Then, for each clustered semantic node group, the corresponding intra-group semantic vector is constructed. For the clustered semantic node group containing "overdue situation" and "amount to be repaid", the server will construct an intra-group semantic vector that can reflect the semantic relationship between the two nodes. This vector may contain some numerical values, which represent certain quantitative characteristics of the two nodes in the semantic relationship, such as the degree of semantic closeness, the importance weight of semantics, etc. Inter-group relationship analysis is performed on the intra-group semantic vectors corresponding to all semantic node groups, and the relationship measurement values ​​between the intra-group semantic vectors of each two semantic node groups are combined into an inter-group relationship matrix. For example, the relationship measurement between the intra-group semantic vectors of the semantic node group in the repayment reminder sub-dialogue and the intra-group semantic vectors of the semantic node group in the customer feedback financial difficulties sub-dialogue may be calculated by calculating the semantic similarity or difference between them, and then these measurement values ​​are combined 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.

[0064] After that, the elements in the inter-group relationship matrix are arranged in a set order, which is based on the original order of the nodes in the semantic relationship graph or a predefined order. For example, if it is based on the original order of the nodes in the semantic relationship graph, the elements in the inter-group relationship matrix are arranged in the order of the semantic nodes in the repayment reminder sub-dialogue first and the semantic nodes in the customer feedback financial difficulties sub-dialogue later.

[0065] Finally, the elements in the arranged inter-group relationship matrix are encoded to generate an initial style flow vector. This initial style flow vector is used to reflect the overall style flow characteristics of the collection template sub-dialogue corresponding to the ath 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 the initial emphasis on repayment facts (repayment date, amount to be repaid, etc.) to the response style after the customer feedback difficulties (how to deal with the customer's financial difficulties).

[0066] After generating the initial style flow vector, the server also needs to perform feature space transformation on the initial style flow vector to generate a style flow vector corresponding to the ath collection style template dialogue. This feature space transformation process may involve readjusting, mapping or transforming the elements in the initial style flow vector to make it more consistent with the requirements of the entire collection style analysis system. For example, based on the results of the previous analysis of a large number of collection dialogues, some features that are more representative or more distinguishing in a specific style may be determined, and then the elements in the initial style flow vector are transformed into this new feature space, and finally the style flow vector corresponding to the ath collection style template dialogue is obtained. This style flow vector can accurately reflect the semantic and stylistic characteristics of the ath collection style template dialogue, and provide an important basis for subsequent operations such as collection style conversion. For each collection style template dialogue, the server will repeat the entire operation process above to obtain the style flow vector corresponding to each collection style template dialogue.

[0067] In a possible implementation, the X collection style template dialogues are continuous collection dialogues in the initial collection dialogue flow. Step S121 includes: Step S1211, 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 there are multiple identical collection style template dialogues in the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue.

[0068] Step S1212: 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.

[0069] Taking the ath collection style template dialogue as an example, when a is less than Y, the server generates the collection template sub-dialogue sequence corresponding to the ath collection style template dialogue based on the previous a collection style template dialogues. Assuming that this is a collection process for debtors with a small overdue amount and a short overdue time, the continuous collection style template dialogues in the initial collection dialogue flow have a certain logical order and association. For example, the first three (assuming a = 3 and Y is greater than 3) collection style template dialogues may all be relatively mild reminder-type dialogues. The server will integrate these three collection style template dialogues to construct the collection template sub-dialogue sequence corresponding to the third collection style template dialogue. Since it is constructed based on the previous a dialogues, there may be multiple identical collection style template dialogues in this sequence. In this sequence, there will be repeated mild reminders, 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 and gently reminding such debtors in the early stage of debt collection. In this way, the server generates a collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue.

[0070] When a is not less than Y, the server loads the a-Y+1th collection style template dialogue to the ath collection style template dialogue among the X collection style template dialogues into the same sequence, thereby generating a collection template sub-dialogue sequence corresponding to the ath collection style template dialogue. Assuming Y = 5, when a = 7, the server will load the 3rd (7-5+1=3)th collection style template dialogue to the 7th collection style template dialogue into the same sequence. If this is a collection process for debtors with gradually increasing overdue amounts and longer overdue periods, then this sequence covers the collection dialogues from earlier ones to the current one. Early collection dialogues may be gentle reminders. As the overdue situation worsens, subsequent collection dialogues may gradually increase the intensity of the urging, mentioning possible overdue fees, credit impacts, and other content. The server integrates these collection style template dialogues at different stages into a sequence that fully reflects the development and changes of the collection process for this type of debtor, thereby generating a collection template sub-dialogue sequence corresponding to the ath collection style template dialogue. In this way, the server can accurately construct the collection template sub-dialogue sequence corresponding to each collection style template dialogue for different collection situations, providing a basis for subsequent collection style analysis and conversion operations.

[0071] In one possible implementation, the method includes: Step S101, obtaining a sample digital human style portrait vector and a sample collection dialogue flow of a first sample virtual collection digital human, wherein the sample digital human style portrait vector of the first sample virtual collection digital human is generated by extracting the collection style portrait features of a sample historical digital human collection dialogue, wherein the sample historical digital human collection dialogue contains a sample collection style template of the first virtual collection digital human, and 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.

[0072] Step S102, extracting style flow vectors from the sample collection template sub-dialogue sequence corresponding to each sample collection style template dialogue, generating sample style flow vectors 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 a sample collection style template of the second sample virtual collection digital person, and Y is less than X.

[0073] Step S103, based on the initialized collection style conversion network, a first collection style conversion is performed 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, to generate a sample collection style conversion result for each sample collection style template dialogue.

[0074] Step S104, training the initialized collection style conversion network according to the sample collection style conversion results of each sample collection style template dialogue to generate a trained collection style conversion network.

[0075] In a possible implementation, step S104 includes: Step S1041, the sample collection style conversion results of each sample collection style template dialogue are subjected to identification processing according to the identification network, and identification results of X sample collection style conversion results are generated. The identification result of any sample collection style conversion result indicates whether the collection style conversion result is a conversion dialogue.

[0076] Step S1042: training a GAN network based on the identification results of the X sample collection style conversion results and the X sample collection style conversion results to generate a trained collection style conversion network, wherein the GAN network includes the identification network and the initialized collection style conversion network.

[0077] In this embodiment, first, the server needs to obtain the sample digital person style portrait vector and sample collection dialogue flow of the first example virtual collection digital person. Taking the overdue collection of small loans in financial institutions as an example, the first example virtual collection digital person has accumulated rich collection experience in the past sample historical digital person collection dialogues, and these sample historical digital person collection dialogues contain the sample collection style template of the first virtual collection digital person. For example, this sample collection style template may be to politely address the debtor at the beginning of the conversation, and 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 of these sample historical digital person collection dialogues, thereby generating a sample digital person style portrait vector of the first example virtual collection digital person, which is a sample collection style semantic vector that can identify the sample collection style portrait of the virtual collection digital person. At the same time, the sample collection dialogue flow obtained by the server includes X sample collection style template dialogues, each of which includes the sample collection style template of the first sample virtual collection digital person. These sample collection style template dialogues may be collection dialogues for debtors with different overdue periods or different amounts of debt, such as one dialogue template for debtors with an overdue period of 1 to 7 days and an amount of 1,000 to 5,000 yuan, and another dialogue template for debtors with an overdue period of 8 to 15 days and an amount of 5,000 to 10,000 yuan.

[0078] Next, the server extracts style flow vectors from the sample collection template sub-dialogue sequence corresponding to each sample collection style template dialogue to generate sample style flow vectors corresponding to each sample collection style template dialogue. The sample collection template sub-dialogue sequence corresponding to any sample collection style template dialogue includes Y sample collection template sub-dialogues (Y is less than X). For example, for a sample collection style template dialogue for a debtor who is overdue for 1-7 days and owes a debt of 1,000-5,000 yuan, its sample collection template sub-dialogue sequence may include 3 sample collection template sub-dialogues (assuming Y = 3). The first sub-dialogue may be a repayment reminder sent by a digital person, including information such as the amount of debt owed and the number of days overdue; the second sub-dialogue may be the debtor's reply, expressing his repayment difficulties; the third sub-dialogue may be some repayment suggestions given by the digital person in response to the debtor's reply. The server will analyze each sub-dialogue in this sequence, extract contextual semantic nodes, determine semantic connection features, construct semantic relationship graphs, perform node clustering analysis, construct intra-group semantic vectors, analyze inter-group relationships, arrange elements and encode, and finally generate a sample style flow vector corresponding to this sample collection style template dialogue. The server will repeat this operation for each sample collection style template dialogue.

[0079] Then, the server performs 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 person style portrait vector of the first virtual collection digital person according to the initialized collection style conversion network, and generates a sample collection style conversion result for each sample collection style template dialogue. Taking one of the sample collection style template dialogues as an example, the initialized collection style conversion network will comprehensively consider the style flow reflected by the sample style flow vector of the dialogue, the specific content in the sample collection template sub-dialogue sequence, and the overall style characteristics represented by the sample digital person style portrait vector. For example, if the sample style flow vector shows that the original style of the dialogue is relatively stiff, and the sample digital person style portrait vector represents a gentle style, the initialized collection style conversion network may adjust the dialogue content to make the tone gentler and more guiding, thereby generating a sample collection style conversion result. This result may be a converted collection dialogue content that is more in line with the ideal collection style.

[0080] Afterwards, the server trains the initialized collection style transfer network based on the sample collection style transfer results of each sample collection style template dialogue to generate a trained collection style transfer network. First, the sample collection style transfer results of each sample collection style template dialogue are identified based on the identification network to generate identification results of X sample collection style transfer results. The function of this identification network is to determine whether each sample collection style transfer result is a valid conversion dialogue. For example, for a sample collection style transfer result, if it is semantically incoherent or the style is far from the target style, the identification network may determine that it is not a valid conversion dialogue. Then, based on the identification results of these X sample collection style transfer results and the X sample collection style transfer results, the GAN network (including the identification network and the initialized collection style transfer network) is trained. For a specific example, for a sample collection style conversion result that is determined by the identification network to be an invalid conversion, the initial collection style conversion network will be adjusted according to this identification result, such as adjusting its weights or algorithm logic when processing the sample style flow vector, the sample collection template sub-dialogue sequence, and the sample digital human style portrait vector. At the same time, the identification network will also continuously optimize its identification criteria based on more sample data. Through continuous adjustment and optimization, a trained collection style conversion network is finally generated. This network can perform collection style conversion more accurately and improve the quality and effectiveness of collection dialogues.

[0081] In a possible implementation, step S1042 includes: Step S1042-1, determining an identification cost parameter according to an error between an identification result of the b-th collection style conversion result and an identification result of the b-th sample collection style template dialogue, where b is a positive integer not greater than X.

[0082] Step S1042-2, determining a collection style conversion cost parameter according to an error between the b-th sample collection style conversion result and the b-th sample collection style template dialogue.

[0083] Step S1042-3, training the GAN network based on the identification cost parameter and the collection style conversion cost parameter.

[0084] In a possible implementation, step S1042-2 includes: Step S1042-21, extracting 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 respectively according to the semantic understanding model, and the semantic understanding vectors of different semantic understanding depths are generated by different semantic understanding units in the semantic understanding model.

[0085] Step S1042-22, performing a fusion calculation on the semantic understanding vector error of the b-th sample collection style conversion result and the b-th sample collection style template dialogue at each semantic understanding depth, and generating a semantic understanding cost parameter between the b-th sample collection style conversion result and the b-th sample collection style template dialogue.

[0086] Step S1042-23, determining the style semantic cost parameter between the bth sample collection style conversion result and the bth sample collection style template dialogue based on the style vector error between the bth sample collection style conversion result and the bth sample collection style template dialogue.

[0087] Step S1042-24, calculating the collection style portrait cost parameter of the bth collection style conversion result based on the 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.

[0088] Step S1042-25, the semantic understanding cost parameter, the style semantic cost parameter, the collection style portrait cost parameter and the game cost parameter are integrated 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.

[0089] The server first determines the identification cost parameter (here b is a positive integer not greater than X) based on the error between the identification result of the bth collection style conversion result and the identification result of the bth sample collection style template dialogue. Taking the above-mentioned overdue collection of small loans as an example, assuming that the bth sample collection style template dialogue is a collection dialogue for debtors who are overdue for 1-7 days and the amount of debt is 1000-5000 yuan, the original collection style template dialogue has a specific identification result, for example, it is judged to have high validity because it is good in semantic expression, style conformity, etc. After conversion, the identification result of the bth collection style conversion result in the identification network may be different from the original one. If the original one is judged to be very consistent with the requirements (for example, the score is 0.9, where the score is a hypothetical value indicating the degree of conformity), and the converted result is judged to be 0.7 (indicating that it is relatively poor), then the error between them (0.2) will be used to determine the identification cost parameter. This identification cost parameter reflects the degree of difference between the conversion result and the original template dialogue in the identification result, and is a quantitative reflection of the conversion effect at the identification level.

[0090] Next, the server determines the collection style conversion cost parameter based on the error between the bth sample collection style conversion result and the bth sample collection style template dialogue. This process is relatively complex and involves multiple aspects.

[0091] First, 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 are extracted based on the semantic understanding model. Semantic understanding vectors of different semantic understanding depths are generated by different semantic understanding units in the semantic understanding model. For example, at a shallower semantic understanding depth, the main focus may be on vocabulary-level understanding. For some key words in the collection dialogue, such as "overdue", "repayment", "amount of debt", 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 method of calculating the number of overdue days in a certain dimension, while the semantic understanding vector of this word in the conversion result may deviate from this dimension. At a deeper level of semantic understanding, it may involve the understanding of sentence structure and logical relationships. For example, if there is a sentence structure with a causal relationship in the original template dialogue (additional fees need to be paid because of overdue payment), its semantic understanding vector will reflect this relationship. If the expression of this relationship is unclear or changed in the conversion result, the semantic understanding vector will also reflect this difference.

[0092] Then, 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 are fused and calculated 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. This fusion calculation process may comprehensively consider the importance weights of different semantic understanding depths. For example, a deeper level of semantic understanding depth may be given a higher weight because it is more critical to the overall semantics. Assuming that the error at the lexical level is 0.1 and the error at the sentence structure level is 0.3 (the numerical value here is a schematic representation of the error size), after weighted fusion (assuming that 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.

[0093] Next, the style semantic cost parameter between the b-th sample collection style conversion result and the b-th sample collection style template dialogue is determined based on the style vector error between the b-th sample collection style conversion result and the b-th sample collection style template dialogue. For example, the original sample collection style template dialogue has a gentle and professional style vector representation, in which certain dimensions of the style vector represent features such as the politeness of the words and the calmness of the tone. The converted b-th sample collection style conversion result may change in style. If there is a large deviation in the politeness dimension of the words, for example, the original words used are very polite, but the words used after conversion are relatively stiff, then the error in this dimension will be calculated into the style vector error, and then the style semantic cost parameter will be determined.

[0094] Then, 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, the collection style portrait cost parameter of the b-th collection style conversion result is calculated. The sample digital human style portrait vector contains the comprehensive features of the entire collection style, such as the overall attitude tendency of the digital human in the collection process, the interaction style with the debtor, and other features. If the distance between the sample digital human style portrait vector of the conversion result and the original b-th sample collection style template dialogue in these features is large, for example, when the attitude tendency changes from mild to tough, this feature distance will be quantified as the collection style portrait cost parameter.

[0095] Finally, the semantic understanding cost parameter, style semantic cost parameter, collection style portrait cost parameter and 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 identification result of the b-th collection style conversion result. For example, if the identification result of the b-th collection style conversion result is poor (indicating that the conversion effect is not good), then the game cost parameter may be 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 fully reflects the conversion cost between the b-th sample collection style conversion result and the b-th sample collection style template dialogue.

[0096] 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 a compass to guide the network adjustment. If the discrimination cost parameter is high, it means that the discrimination network has a large deviation in judging the conversion result, then the discrimination network part in the GAN network will adjust its judgment criteria and algorithm logic. If the collection style conversion cost parameter is high, it means that the conversion network has a large error when converting the sample collection style template dialogue into the sample collection style conversion result, then the initialization collection style conversion network part in the GAN network will adjust its conversion method, such as adjusting the style flow vector, the sample collection template sub-dialogue sequence, and the sample digital human style portrait vector. By continuously adjusting the GAN network according to these cost parameters, a trained collection style conversion network is finally generated. This network can perform collection style conversion more accurately and improve the quality and effect of collection dialogues.

[0097] Figure 2 The hardware structure of the system 100 for generating debt collection dialogue based on AI semantic understanding for realizing the above-mentioned method for generating debt collection dialogue based on AI semantic understanding provided by an embodiment of the present invention is shown as follows: Figure 2 As shown, the collection dialogue 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.

[0098] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the collection dialogue generation system 100 based on AI semantic understanding uses to execute or use to complete the exemplary method described in the present invention.

[0099] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the collection dialogue generation method based on AI semantic understanding in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0100] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned collection dialogue generation system 100 based on AI semantic understanding. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0101] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the above-mentioned method for generating collection dialogues based on AI semantic understanding is implemented.

[0102] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A collection dialogue generation method based on AI semantic understanding, characterized in that: The method comprises: Obtaining a digital human style portrait vector of a first virtual debt collection digital human and an initial debt collection dialogue flow, wherein the digital human style portrait vector of the first virtual debt collection digital human is generated by extracting the characteristics of the debt collection style portrait of a historical digital human debt collection dialogue, wherein the historical digital human debt collection dialogue includes a debt collection style template of the first virtual debt collection digital human, and the digital human style portrait vector is a debt collection style semantic vector used to identify the debt collection style portrait of the virtual debt collection digital human; the initial debt collection dialogue flow includes X debt collection style template dialogues, each of which includes the debt collection style template of the second virtual debt collection digital human; 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, wherein 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 collection style conversion on the style flow 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 through a collection style conversion network to generate a collection style conversion result for each collection style template dialogue; A target collection dialogue flow corresponding to the initial collection dialogue flow is generated according to the collection style conversion results of the X collection style template dialogues.

2. The method for generating debt collection dialogue based on AI semantic understanding according to claim 1 is characterized in that: The step of performing a first collection style conversion on the style flow 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 through the collection style conversion network to generate a collection style conversion result for each collection style template dialogue includes: Encoding the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue to generate a potential style vector corresponding to the a-th collection style template dialogue, where a is a positive integer not greater than X; Integrate the style flow vector corresponding to the a-th 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 a-th debt collection style template dialogue; Aggregating the integrated vector of the a-th debt collection style template dialogue and the potential style vector corresponding to the a-th debt collection style template dialogue to generate an aggregated vector of the a-th 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 method for generating debt collection dialogue based on AI semantic understanding according to claim 1 is characterized in that: 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; 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 vector extraction is performed on each context semantic node to generate the style flow vector corresponding to the a-th collection style template dialogue.

4. The method for generating debt collection dialogue based on AI semantic understanding according to claim 3 is characterized in that: The extracting of style flow vectors from each context semantic node based on the semantic jump order of the collection template sub-dialog corresponding to the ath collection style template dialogue and the semantic connection features of the context semantic nodes of each collection template sub-dialog to generate the style flow vector corresponding to the ath collection style template dialogue includes: According to the semantic jump order of the collection template sub-dialog corresponding to the ath collection style template dialogue and the semantic connection features of the context semantic nodes of each collection template sub-dialog, style flow vector extraction is performed on each context semantic node to generate an initial style flow vector; Performing feature space transformation on the initial style flow vector to generate a style flow vector corresponding to the ath debt collection style template dialogue; The step of extracting style flow vectors from each context semantic node based on the semantic jump order of the collection template sub-dialog corresponding to the ath collection style template dialogue and the semantic connection features of the context semantic nodes of each collection template sub-dialog to generate an initial style flow vector comprises: For each collection template sub-dialogue corresponding to the a-th collection style template dialogue, its context semantic nodes are regarded as nodes in the semantic relationship graph, edges are established between the nodes according to the semantic connection characteristics of the context semantic nodes, and the nodes are arranged according to the semantic jump order of the collection template sub-dialogue to construct the semantic relationship graph, and the semantic relationship is used to reflect the structural relationship of the collection template sub-dialogue corresponding to the a-th collection style template dialogue at the semantic level; 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 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 with the relationship measurement values ​​between the semantic vectors of every two semantic node groups. The inter-group relationship matrix is ​​used to reflect the semantic similarity or difference relationship 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; The elements in the arranged inter-group relationship matrix are encoded 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 order and semantic connection characteristics, and includes the relationship information between each semantic node group.

5. The method for generating debt collection dialogue based on AI semantic understanding according to claim 3 is characterized in that: The X collection style template dialogues are continuous collection dialogues in the initial collection dialogue flow; the acquisition of 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 there are multiple identical collection style template dialogues in the collection template sub-dialogue sequence corresponding to the a-th collection style template dialogue; 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.

6. The method for generating debt collection dialogue based on AI semantic understanding according to any one of claims 1 to 5, characterized in that: The method comprises: Obtaining a sample digital human style portrait vector and a sample collection dialogue flow of a first sample virtual collection digital human, wherein the sample digital human style portrait vector of the first sample virtual collection digital human is generated by extracting collection style portrait features of a sample historical digital human collection dialogue, wherein the sample historical digital human collection dialogue contains a sample collection style template of the first virtual collection digital human, and the digital human style portrait vector is a sample collection style semantic vector used to identify a sample collection style portrait of a 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; Performing 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 according to the initialized collection style conversion network, and generating a sample collection style conversion result for each sample collection style template dialogue; The initialized collection style conversion network is trained according to the sample collection style conversion results of each sample collection style template dialogue to generate a trained collection style conversion network.

7. The method for generating debt collection dialogue based on AI semantic understanding according to claim 6 is characterized in that: The initializing collection style conversion network is trained according to 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 subjected to identification processing according to the identification network, and identification results of X sample collection style conversion results are generated; the identification result of any sample collection style conversion result indicates whether the collection style conversion result is a conversion dialogue; According to the identification results of the X sample collection style conversion results and the X sample collection style conversion results, the GAN network is trained to generate a trained collection style conversion network, wherein the GAN network includes the identification network and the initialized collection style conversion network.

8. The method for generating debt collection dialogue based on AI semantic understanding according to claim 7 is characterized in that: The step of training the GAN network based on the identification results of the X sample collection style conversion results and the X sample collection style conversion results to generate a trained collection style conversion network includes: Determine the identification cost parameter according to the error between the identification result of the b-th collection style conversion result and the identification result of the b-th sample collection style template dialogue, where b is a positive integer not greater than X; Determining a collection style conversion cost parameter according to an error between a b-th sample collection style conversion result and the b-th sample collection style template dialogue; Based on the identification cost parameter and the collection style conversion cost parameter, the GAN network is trained.

9. The method for generating debt collection dialogue based on AI semantic understanding according to claim 8 is characterized in that: The step of determining a collection style conversion cost parameter based on an error between the b-th sample collection style conversion result and the b-th sample collection style template dialogue includes: Extracting 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 respectively according to 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 b-th sample collection style conversion result and the b-th sample collection style template dialogue at each semantic understanding depth, and generating a semantic understanding cost parameter between the b-th sample collection style conversion result and the b-th 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 according to 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 b-th collection style conversion result according to a feature distance between a sample digital human style portrait vector of the b-th sample collection style conversion result and the b-th 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.

10. A debt collection dialogue generation system based on AI semantic understanding, characterized in that: The 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 collection dialogue generation method based on AI semantic understanding as described in any one of claims 1 to 9 above.

Images