Information decision method and system based on collection session big data analysis

Abstract

The invention provides an information decision-making method and system based on collection session big data analysis, and the method comprises the steps: firstly obtaining collection session big data, and then generating sample collection session training data for a session data streams in the collection session big data through enhancing key semantic nodes of collection target users in the session data streams. In the process, if a risk fraud event exists in the semantic representation vector associated with the key semantic node of the collection target user, the risk fraud event is included into sample collection session training data. And finally, parameter learning is performed on the collection session decision network by using the sample collection session training data, so that the collection session decision network for completing parameter learning, which can accurately decide whether risk fraud events exist in the collection session data, is generated, and key information in the collection session big data is deeply mined, so that the risk fraud events in the collection session data can be accurately determined. And efficient identification of risk fraud events is realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an information decision-making method and system based on collection conversation big data analysis. Background Art

[0002] In the financial and credit industries, debt collection is an important part of ensuring the return of funds and maintaining creditor rights. However, traditional debt collection methods often rely on manual judgment and experience, which is not only inefficient, but also difficult to accurately identify and handle risky fraud incidents that may occur during the debt collection process. Especially when faced with a large amount of debt collection conversation data, manual analysis seems to be unable to cope with it, and it is easy to make omissions and misjudgments.

[0003] With the rapid development of big data technology, it has become possible to use big data analysis methods to conduct in-depth mining and analysis of collection conversation data. By collecting and organizing collection conversation data, a large collection conversation data set containing rich information can be formed. These data sets contain information on the behavior patterns, repayment willingness, fraud risks, and other aspects of the collection target users, providing strong data support for collection decisions.

[0004] However, how to effectively extract key information from collection conversation big data and make accurate decisions based on it is still an urgent problem to be solved. Most existing collection conversation analysis methods focus on simple data statistics and rule matching, which makes it difficult to capture the deep semantic information and risk fraud characteristics in the conversation. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an information decision method based on collection conversation big data analysis, the method comprising: Acquire collection conversation big data, the collection conversation big data including a plurality of conversation data streams associated with collection target users sorted according to the order of collection conversation nodes; For a session data streams in the collection session big data, the key semantic nodes of the collection target user in the session data streams are enhanced to generate sample collection session training data, and the semantic representation vectors associated with the key semantic nodes of the collection target user have risk fraud events in the sample collection session training data, where a is a positive integer; Parameter learning is performed on the collection session decision network based on the sample collection session training data to generate a collection session decision network that has completed parameter learning. The collection session decision network that has completed parameter learning is used to decide whether there is a risk of fraud in the collection session data.

[0006] In a possible implementation manner of the first aspect, the enhancing the key semantic nodes of the collection target user in the conversation data stream to generate sample collection conversation training data includes: enhancing each session semantic unit in the session data stream to generate an enhanced session data stream, wherein the session semantic units in the enhanced session data stream are inconsistent with the session semantic units in the session data stream; Aggregating the enhanced session data stream and the session data stream to generate a fictitious session data stream, wherein key semantic nodes of the target user for collection in the fictitious session data stream show inconsistency with key semantic nodes of the target user for collection in the session data stream, and non-key semantic nodes of the target user for collection in the fictitious session data stream show consistency with non-key semantic nodes of the target user for collection in the session data stream; The a conversation data streams are overwritten with various fictitious conversation data streams corresponding to the a conversation data streams to generate the sample collection conversation training data.

[0007] In a possible implementation manner of the first aspect, the aggregating the enhanced session data stream and the session data stream to generate a fictitious session data stream includes: Acquire multiple key semantic node sets corresponding to the target user for collection in the session data stream, each of the key semantic node sets including multiple key semantic nodes; For any one of the multiple key semantic node sets, generating a semantic feature domain corresponding to the session data stream according to the key semantic node set, wherein the semantic feature domain represents a penetration segment of the key semantic node set in the session data stream; Under the restriction of various corresponding semantic feature domains of the plurality of key semantic node sets, the enhanced session data stream and the session data stream are aggregated to generate the fictitious session data stream.

[0008] In a possible implementation manner of the first aspect, generating a semantic feature domain corresponding to the session data flow according to the key semantic node set includes: For any session semantic unit in the session data stream, obtaining a first correlation between the session semantic unit and the penetration paragraph where the key semantic node set is located; Based on the first correlation, determining a semantic feature value of the conversation semantic unit, wherein the semantic feature value of the conversation semantic unit is in a positive correlation relationship with the first correlation; The semantic feature domain is generated according to the semantic feature value of each conversation semantic unit in the conversation data stream.

[0009] In a possible implementation manner of the first aspect, aggregating the enhanced session data stream and the session data stream to generate the fictitious session data stream under the restriction of various corresponding semantic feature domains of the multiple key semantic node sets includes: For any one of the semantic feature domains corresponding to the multiple key semantic node sets, optimizing the enhanced session data stream according to the semantic feature domain to generate a first composite session data stream, and optimizing the session data stream according to an inverse semantic feature domain corresponding to the semantic feature domain to generate a second composite session data stream, wherein the sum of the semantic feature values ​​of the session semantic units of the same penetration paragraph in the semantic feature domain and the inverse semantic feature domain is 1; generating a composite session data stream corresponding to the semantic feature domain according to the first composite session data stream and the second composite session data stream; Aggregate the composite conversation data streams corresponding to the plurality of the semantic feature domains to generate the fictitious conversation data stream.

[0010] In a possible implementation manner of the first aspect, optimizing the enhanced session data stream according to the semantic feature domain to generate a first composite session data stream, and optimizing the session data stream according to an inverse semantic feature domain corresponding to the semantic feature domain to generate a second composite session data stream include: For any session semantic unit in the enhanced session data stream, multiply the session semantic unit by a semantic feature value corresponding to the session semantic unit in the semantic feature domain to generate the first composite session data stream; For any session semantic unit in the session data stream, the session semantic unit is merged with the semantic feature value corresponding to the session semantic unit in the inverse semantic feature domain to generate the second composite session data stream.

[0011] In a possible implementation manner of the first aspect, the enhancing each session semantic unit in the session data stream to generate the enhanced session data stream includes: Preliminarily segmenting the session data stream according to semantic integrity to generate multiple independent session semantic units, identifying the vocabulary type in each session semantic unit, and marking the key entities and semantic roles in each session semantic unit based on the vocabulary type to generate marking information for each session semantic unit, wherein the key entities include debt information related to debt collection, financial products involved, identity-related information of target users for debt collection, and domain vocabulary related to risk fraud, and the semantic roles include agent and recipient; Based on the tag information of each conversation semantic unit, a corresponding semantic association knowledge base is constructed. The construction of the semantic association knowledge base is obtained by analyzing the corpus related to debt collection and the general corpus. Specifically, for the nouns in each conversation semantic unit, an association relationship with other related nouns is established in the semantic association knowledge base. For the verbs in each conversation semantic unit, its semantic association with different subject and object combinations is established. For adjectives and adverbs, the semantic association between them and the modified objects and the semantic change association in different contexts is established. For each conversation semantic unit, extract the core semantics of the conversation semantic unit according to the semantic association knowledge base to obtain a conversation semantic unit set containing the core semantics; For each conversation semantic unit containing core semantics, semantic expansion is performed according to the semantic association knowledge base, wherein, for the noun part in the core semantics, other nouns, adjectives, and verbs related thereto are searched from the semantic association knowledge base and then expanded into the conversation semantic unit, and, for the verb part in the core semantics, its related synonyms, near synonyms, and different tenses and voice forms are searched and then expanded into the conversation semantic unit, thereby obtaining a set of conversation semantic units after semantic expansion; Performing a semantic conversion operation on the basis of the semantically expanded conversation semantic unit set, converting part of the semantics in the conversation semantic unit according to the semantic association knowledge base, and obtaining a semantically converted conversation semantic unit set; Based on the prior context information of the collection conversation, the conversation semantic unit set after the semantic conversion is adjusted for context adaptability to obtain the conversation semantic unit set after context adaption; Analyze the logical relationship between each conversation semantic unit in the conversation semantic unit set after context adaptation, and determine whether there is a semantic jump or semantic contradiction between adjacent conversation semantic units. If so, adjust the relevant conversation semantic units to obtain a logically coherent conversation semantic unit set. The logically coherent conversation semantic units are reassembled according to the order in the original implementation conversation data stream to form an enhanced conversation data stream.

[0012] In a possible implementation of the first aspect, performing parameter learning on the collection session decision network according to the sample collection session training data to generate the collection session decision network that has completed parameter learning includes: Using the collection conversation decision network to obtain a first mapping knowledge vector of the sample collection conversation training data, the first mapping knowledge vector representing semantic transition information in the time domain of a semantic representation vector corresponding to the collection target user in the sample collection conversation training data; Using the collection conversation decision network to generate a second mapping knowledge vector based on the first mapping knowledge vector, the second mapping knowledge vector represents the semantic transformation information of the semantic representation vector corresponding to the collection target user in the sample collection conversation training data in the spatial domain; Using the collection conversation decision network to generate a decision result of the sample collection conversation training data according to the second mapping knowledge vector, the decision result of the sample collection conversation training data characterizes whether there is a risk fraud event in the semantic change of the semantic representation vector associated with the key semantic node of the collection target user in the sample collection conversation training data; Based on the decision result, parameter learning is performed on the debt collection conversation decision network to generate the debt collection conversation decision network that has completed parameter learning.

[0013] In a possible implementation of the first aspect, the generating a second mapping knowledge vector by using the collection session decision network according to the first mapping knowledge vector includes: Performing attention weight allocation processing on the first mapping knowledge vector to generate a first mapping knowledge vector after the attention weight allocation processing; Generate the second mapping knowledge vector by using the first mapping knowledge vector processed by the collection conversation decision network according to the attention weight distribution; The knowledge fragments in the first mapping knowledge vector are arranged in the time domain, and the knowledge fragments in the first mapping knowledge vector after the attention weight allocation process are arranged in the space domain; In a possible implementation of the first aspect, the collection session decision network includes a time domain knowledge encoder and a space domain knowledge encoder, the time domain knowledge encoder is generated according to a first initialization encoder and a first recurrent neural network, the space domain knowledge encoder is generated according to a second initialization encoder and a second recurrent neural network, the first recurrent neural network is used to obtain the first mapping knowledge vector according to the first initialization encoder, and the second recurrent neural network is used to obtain the second mapping knowledge vector according to the second initialization encoder; The performing parameter learning on the collection session decision network based on the decision result to generate the collection session decision network that has completed parameter learning includes: Based on the decision result, determining a training error parameter of the collection session decision network; The weight parameter information of the first initialization encoder and the weight parameter information of the second initialization encoder are locked, and the weight parameter information of the first recurrent neural network and the weight parameter information of the second recurrent neural network are optimized based on the training error parameters to generate the collection session decision network that completes parameter learning.

[0014] On the other hand, an embodiment of the present invention also provides an information decision-making system based on collection conversation big data analysis, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0015] Based on the above aspects, the embodiment of the present application obtains and organizes the collection conversation big data, enhances the key semantic nodes to generate sample collection conversation training data, and then performs parameter learning on the collection conversation decision network, and finally generates a collection conversation decision network that has completed parameter learning and can accurately decide whether there is a risk of fraud in the collection conversation data. This effectively improves the accuracy and efficiency of the collection conversation data analysis, and helps to timely discover and deal with potential risk fraud events. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the execution flow of the information decision-making method based on collection conversation big data analysis provided by an embodiment of the present invention.

[0017] Figure 2 It is a schematic diagram of the hardware architecture of an information decision-making system based on collection conversation big data analysis provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of an information decision-making method based on collection conversation big data analysis provided by an embodiment of the present invention. The information decision-making method based on collection conversation big data analysis is introduced in detail below.

[0019] Step S110, obtaining collection conversation big data, wherein the collection conversation big data includes a plurality of conversation data streams associated with collection target users sorted according to the order of collection conversation nodes.

[0020] In this embodiment, the collection session big data is a collection of a series of session information related to the target users who need to be collected. For example, a financial institution provides a variety of financial products, such as personal loans, credit card consumption and other services. When the user fails to repay on time, the collection process will be entered.

[0021] The server collects these conversation data from multiple systems. For example, the telephone communication record system between the customer service department and the target user of the collection. When the customer service staff makes a collection call, the telephone system will automatically record the entire call process, including what the customer service said and the user's response. These call records are sorted according to the order in which the calls occurred, that is, the order of the collection conversation nodes. The call record of each call forms a conversation data stream. For another example, the chat record system between the online customer service and the target user of the collection. When the user has a conversation with the online customer service about repayment collection through the official website of the financial institution or the mobile APP, these chat records will also be sorted according to the order of the conversation to become a conversation data stream.

[0022] There are also some possible channels, such as offline debt collectors communicating with users face to face and entering the communication content into a special system. The server will also obtain this data and organize it into a conversation data stream according to the order of the debt collection conversation nodes. These numerous conversation data streams associated with the debt collection target users are gathered together to form the debt collection conversation big data obtained by the server.

[0023] Take a specific credit card collection scenario as an example. User A's credit card is overdue. The customer service staff first called on the first day after the overdue date. During the call, the customer service staff introduced themselves, informed user A that the credit card was overdue, and asked about the repayment plan. User A said that he was short of money recently and it might take a few days to repay. This call record is part of a session data stream. On the third day, the customer service staff called again to remind user A that the overdue situation might affect his credit record. User A said that he was trying to raise money. These two call records, in sequence, completely constitute a session data stream associated with user A, the collection target user. The server will obtain many such session data streams of different users, thus forming collection session big data.

[0024] Step S120, for a session data streams in the collection session big data, enhance the key semantic nodes of the collection target user in the session data stream, generate sample collection session training data, and the semantic representation vector associated with the key semantic nodes of the collection target user in the sample collection session training data has risk fraud events, and a is a positive integer.

[0025] In this embodiment, after the collection session big data is obtained, the a session data streams are processed. Assume that the server selects 100 (ie, a = 100) session data streams from the numerous session data streams as processing objects.

[0026] Take one of the conversation data streams related to user B as an example. User B is a customer who has overdue personal loan repayments. The original conversation data stream contains multiple conversation records between the customer service and user B. Some of the key semantic nodes may include user B's mention of repayment ability, source of funds, repayment time, and other related statements.

[0027] First, the server needs to enhance each session semantic unit in the session data stream to generate an enhanced session data stream. For the processing of session semantic units, the server will perform a series of complex operations. For example, the session data stream is initially segmented according to semantic integrity to generate multiple independent session semantic units. Assume that in a conversation, user B says "I recently lost my job and have no source of income, so I can't pay back the loan." The server will identify the vocabulary types in this session semantic unit, analyze the words "unemployed", "source of income", "can't pay back the loan", etc., and mark the key entities (such as "loan", which is debt information related to collection) and semantic roles (such as user B as the agent, indicating that he is unemployed and has no income, so he cannot repay the loan) based on these vocabulary types, and generate the marking information of this session semantic unit.

[0028] Then, the server builds the corresponding semantic association knowledge base based on the tag information of each conversation semantic unit. For the noun "unemployment" in this conversation semantic unit, the server will establish associations with other related nouns in the semantic association knowledge base. For example, "unemployment" may be associated with nouns such as "economic situation" and "employment market"; for the verb "cannot pay it back", establish its semantic associations with different subject and object combinations, such as the semantic association of causal relationship "because there is no income, so I can't pay back the loan", and for adjectives and adverbs (if they exist), establish their semantic associations with the modified objects and semantic change associations in different contexts.

[0029] Next, for each conversation semantic unit, the server extracts the core semantics of the conversation semantic unit based on the semantic association knowledge base. For the conversation semantic unit of user B, the core semantics may be "unemployment leads to inability to repay". Then, semantic expansion is performed based on the semantic association knowledge base. For the noun part "unemployment" in the core semantics, other nouns, adjectives, and verbs related to it are searched from the semantic association knowledge base and expanded to the conversation semantic unit. It may become "because of the industry downturn leading to unemployment, there is no source of income, and there is really no ability to repay the loan"; for the verb part "repay" in the core semantics, its related synonyms, near synonyms, and different tenses and voice forms are searched and expanded to the conversation semantic unit, such as "unable to fulfill the repayment obligation". After obtaining the conversation semantic unit set after semantic expansion, a semantic conversion operation is performed on this basis. According to the semantic association knowledge base, some semantics in the conversation semantic unit are converted, for example, "inability to repay the loan" is converted to "difficult to bear the repayment responsibility of the loan". Then, based on the prior context information of the collection conversation, the conversation semantic unit set after semantic conversion is adjusted to adapt to the context, and the conversation semantic unit set after context adaptation is obtained. For example, if the previous conversation mentioned some preferential repayment policies of the bank, then in this conversation semantic unit, it can be adjusted to "Although I know that the bank has preferential repayment policies, I am unemployed now and still find it difficult to bear the repayment responsibility of the loan."

[0030] After that, the logical relationship between each conversation semantic unit in the context-adapted conversation semantic unit set is analyzed to determine whether there is a semantic jump or semantic contradiction between adjacent conversation semantic units. If so, the relevant conversation semantic units are adjusted to obtain a logically coherent conversation semantic unit set. Finally, the logically coherent conversation semantic units are reassembled according to the order in the original conversation data stream to form an enhanced conversation data stream.

[0031] After the enhanced session data stream is generated, the server aggregates the enhanced session data stream and the original session data stream to generate a fictitious session data stream. First, the server obtains multiple key semantic node sets corresponding to the target users in the original session data stream. For example, for the session data stream of user B, a key semantic node set may be all expressions about repayment ability, including key semantic nodes such as "unemployed and no income" and "no savings". Then, for each key semantic node set, the semantic feature domain corresponding to the session data stream is generated based on this key semantic node set. Assume that for the key semantic node set of repayment ability, the server obtains the first correlation between this session semantic unit and the penetration paragraph where the key semantic node set is located for any session semantic unit in the session data stream. For example, user B said "I recently spent a sum of money on medical treatment", and the first correlation between this session semantic unit and the penetration paragraph where the key semantic node set of repayment ability is located (the paragraph about the statement of financial status) may be high, because medical expenses also affect the financial status and thus affect the repayment ability. Based on the first correlation, the semantic feature value of the conversation semantic unit is determined. The semantic feature value of the conversation semantic unit is positively correlated with the first correlation. If the first correlation is 0.8, then the semantic feature value may be 0.8. The semantic feature domain is generated based on the semantic feature values ​​of each conversation semantic unit in the conversation data stream.

[0032] After obtaining the semantic feature domains corresponding to multiple sets of key semantic nodes, for any one of the semantic feature domains, the enhanced conversation data stream is optimized according to the semantic feature domain to generate a first composite conversation data stream, and the original conversation data stream is optimized according to the inverse semantic feature domain corresponding to the semantic feature domain to generate a second composite conversation data stream. For example, for any conversation semantic unit in the enhanced conversation data stream corresponding to the semantic feature domain of repayment ability, the conversation semantic unit is multiplied by the semantic feature value corresponding to the semantic feature domain to generate a first composite conversation data stream; for any conversation semantic unit in the original conversation data stream, the conversation semantic unit is fused with the semantic feature value corresponding to the inverse semantic feature domain to generate a second composite conversation data stream. In the semantic feature domain and the inverse semantic feature domain, the sum of the semantic feature values ​​of the conversation semantic units of the same penetration paragraph is 1. Then, based on the first composite conversation data stream and the second composite conversation data stream, a composite conversation data stream corresponding to the semantic feature domain is generated. Finally, the composite conversation data streams corresponding to multiple semantic feature domains are aggregated to generate a fictitious conversation data stream. In this fictitious conversation data stream, the key semantic nodes of the target user for collection show inconsistency with the key semantic nodes in the original conversation data stream, while the non-key semantic nodes show consistency with the non-key semantic nodes in the original conversation data stream.

[0033] The server overwrites these 100 conversation data streams with corresponding fictitious conversation data streams in the above manner to generate sample collection conversation training data. In this sample collection conversation training data, the semantic representation vector associated with the key semantic node of the collection target user contains risk fraud events. For example, in the fictitious conversation data stream, user B may suddenly make unreasonable statements related to repayment ability, such as saying that he has a large source of funds but does not repay the loan in a short period of time, instead of saying that he has no income. This semantic representation vector may indicate the existence of risk fraud events.

[0034] Step S130, performing parameter learning on the collection session decision network based on the sample collection session training data to generate a collection session decision network that has completed parameter learning. The collection session decision network that has completed parameter learning is used to decide whether there is a risk fraud event in the collection session data.

[0035] In this embodiment, the above-mentioned embodiment has generated sample collection conversation training data, and now the collection conversation decision network is to be parameterized using these data. The collection conversation decision network includes a time domain knowledge encoder and a space domain knowledge encoder, wherein the time domain knowledge encoder is generated based on the first initialization encoder and the first recurrent neural network, and the space domain knowledge encoder is generated based on the second initialization encoder and the second recurrent neural network.

[0036] First, the server uses the collection conversation decision network to obtain the first mapping knowledge vector of the sample collection conversation training data. This first mapping knowledge vector represents the semantic transformation information of the semantic representation vector corresponding to the collection target user in the sample collection conversation training data in the time domain. Taking the previously mentioned fictitious conversation data stream of user B as an example, in this conversation data stream, user B's expression of repayment ability has a semantic transformation in the time domain as the conversation progresses, from the initial lack of repayment ability to the sudden appearance of unreasonable fund source expression later. These semantic transformation information is encoded in the first mapping knowledge vector.

[0037] Then, the server uses the collection session decision network to generate a second mapping knowledge vector based on the first mapping knowledge vector. In this process, the server will perform attention weight allocation processing on the first mapping knowledge vector to generate a first mapping knowledge vector after the attention weight allocation processing. Assume that in the first mapping knowledge vector, the knowledge fragments about the early part and the later sudden change part of the description of user B's repayment ability are arranged in the time domain. After the attention weight allocation processing, these knowledge fragments are rearranged in the spatial domain according to their importance to form the first mapping knowledge vector after the attention weight allocation processing. Then, the collection session decision network is used to generate a second mapping knowledge vector based on the first mapping knowledge vector after the attention weight allocation processing. This second mapping knowledge vector represents the semantic transformation information in the spatial domain of the semantic representation vector corresponding to the collection target user in the sample collection session training data.

[0038] Next, the server uses the collection session decision network to generate a decision result for the sample collection session training data based on the second mapping knowledge vector. This decision result characterizes whether there is a risk fraud event in the semantic change of the semantic representation vector associated with the key semantic node of the collection target user in the sample collection session training data. For the fictitious session data stream of user B, if there are obvious unreasonable aspects in the semantic changes reflected by the second mapping knowledge vector, such as the expression of repayment ability is illogical and matches the pattern related to risk fraud, then the decision result will determine that there is a risk fraud event.

[0039] Finally, the collection session decision network is trained with parameters based on this decision result. The server will determine the training error parameters of the collection session decision network based on the decision result. If it is determined that there is a risk fraud event but it is actually a misjudgment, or vice versa, a training error will be generated. The weight parameter information of the first initialization encoder and the weight parameter information of the second initialization encoder are locked, and the weight parameter information of the first recurrent neural network and the weight parameter information of the second recurrent neural network are optimized based on this training error parameter, thereby generating a collection session decision network that has completed parameter learning. This collection session decision network that has completed parameter learning can be used to make decisions on new collection session data to determine whether there is a risk fraud event. For example, when new collection session data of user C enters, this collection session decision network can analyze it according to the same mechanism to determine whether there is a risk fraud event in the session data of user C.

[0040] Based on the above steps, the embodiment of the present application obtains and organizes the collection conversation big data, enhances the key semantic nodes to generate sample collection conversation training data, and then performs parameter learning on the collection conversation decision network, and finally generates a collection conversation decision network that has completed parameter learning and can accurately decide whether there is a risk of fraud in the collection conversation data. This effectively improves the accuracy and efficiency of the collection conversation data analysis, and helps to promptly discover and deal with potential risk of fraud events.

[0041] In a possible implementation, step S120 includes: Step S121 , enhancing each session semantic unit in the session data stream to generate an enhanced session data stream, wherein the session semantic units in the enhanced session data stream are inconsistent with the session semantic units in the session data stream.

[0042] Step S122, aggregating the enhanced session data stream and the session data stream to generate a fictional session data stream, wherein the key semantic nodes of the target user for collection in the fictional session data stream show inconsistency with the key semantic nodes of the target user for collection in the session data stream, and the non-key semantic nodes of the target user for collection in the fictional session data stream show consistency with the non-key semantic nodes of the target user for collection in the session data stream.

[0043] Step S123: Overwrite the a conversation data streams with various corresponding fictitious conversation data streams of the a conversation data streams to generate the sample collection conversation training data.

[0044] In this embodiment, taking the conversation data stream of user A whose credit card is overdue as mentioned above as an example, the first step is to enhance each conversation semantic unit in the conversation data stream to generate an enhanced conversation data stream. In the original conversation data stream, when the customer service communicates with user A, user A says, "I have been spending too much recently and cannot pay back the credit card debt for the time being." The server will first divide this conversation data stream into independent conversation semantic units according to semantic integrity. For this conversation semantic unit, the server identifies the vocabulary type, marks the key entity such as the debt information "credit card debt", and the semantic role. Here, user A is the agent, indicating that he cannot repay the debt due to high expenses, and generates marking information. Then, a semantic association knowledge base is constructed. For the noun "expenses", it is associated with "living expenses", "consumption items", etc., and for the verb "cannot pay back", a semantic association with different subject and object combinations is established. Then, the core semantics extracted may be "high expenses lead to inability to repay", and semantic expansion is performed to expand "high expenses" to "due to unexpected large medical expenses and other living expenses, funds are tight", and the verb part is also similarly expanded. Then, semantic transformation is performed, such as changing it to "Funds are insufficient due to various expenses, making it difficult to repay the credit card debt", and then adjusting it according to the prior context, such as the content of the bank's repayment reminder SMS mentioned earlier, which is adjusted to "Although I received the bank's repayment reminder, I am insufficient due to various expenses and it is difficult to repay the credit card debt". Finally, the logical relationship is analyzed to ensure coherence and reassembled to form an enhanced conversation data stream. The conversation semantic units in this enhanced conversation data stream show inconsistencies with the conversation semantic units in the original conversation data stream, such as more detailed and specific expressions.

[0045] After the enhanced session data stream is generated, it is necessary to aggregate it with the original session data stream to generate a fictitious session data stream. Still taking user A as an example, the server first obtains the key semantic node set corresponding to user A in the original session data stream, such as the statement about repayment ability and repayment willingness. For the key semantic node set of repayment ability, the server determines the relevance of each session semantic unit in the session data stream to the penetration paragraph where the key semantic node set is located. For example, when user A mentions "I have other debts to repay", this session semantic unit has a high relevance to repayment ability, determines its semantic feature value, and generates a semantic feature domain based on the semantic feature values ​​of all session semantic units. Then, under the restriction of this semantic feature domain, the enhanced session data stream is optimized to generate a first composite session data stream, and the original session data stream is optimized to generate a second composite session data stream. For example, a session semantic unit in the enhanced session data stream is multiplied by the semantic feature value to obtain the corresponding part in the first composite session data stream, and the session semantic unit in the original session data stream is fused with the inverse semantic feature value to obtain the part in the second composite session data stream. Then, a composite session data stream corresponding to the semantic feature domain is generated based on the first composite session data stream and the second composite session data stream, and a plurality of such composite session data streams are aggregated to obtain a fictitious session data stream. In this fictitious session data stream, the key semantic nodes of user A, such as expressions related to repayment ability, are inconsistent with the key semantic nodes in the original session data stream, while some insignificant non-key semantic nodes, such as irrelevant life trivia mentioned by user A, remain consistent.

[0046] Finally, the server converts the previously selected a conversation data streams (such as the 100 mentioned above) into their corresponding fictitious conversation data streams in the above manner, and these fictitious conversation data streams constitute the sample collection conversation training data. This is like creating a special training material library for the collection conversation decision network, in which the sample data contains carefully constructed fictitious conversations that have a specific relationship with the original conversations, so that the network can learn the characteristic patterns of risky fraud events.

[0047] In a possible implementation, step S122 includes: Step S1221, obtaining a plurality of key semantic node sets corresponding to the target collection user in the session data stream, each of the key semantic node sets including a plurality of key semantic nodes.

[0048] Step S1221, for any one of the multiple key semantic node sets, generate a semantic feature domain corresponding to the session data stream based on the key semantic node set, and the semantic feature domain represents the penetration paragraph of the key semantic node set in the session data stream.

[0049] Step S1222 , under the restriction of various corresponding semantic feature domains of the plurality of key semantic node sets, the enhanced session data stream and the session data stream are aggregated to generate the fictitious session data stream.

[0050] In a possible implementation, step S1221 includes: Step S1221 - 1 , for any session semantic unit in the session data stream, obtain a first correlation between the session semantic unit and the penetration paragraph where the key semantic node set is located.

[0051] Step S1221 - 2 : determining the semantic feature value of the conversation semantic unit based on the first correlation, wherein the semantic feature value of the conversation semantic unit is positively correlated with the first correlation.

[0052] Step S1221 - 3 , generating the semantic feature domain according to the semantic feature value of each conversation semantic unit in the conversation data stream.

[0053] In this embodiment, taking the above-mentioned credit card overdue user A as an example, the server first needs to obtain multiple key semantic node sets corresponding to the session data stream of user A. For user A, the target user for collection, one key semantic node set may be about repayment ability, which includes multiple key semantic nodes such as "monthly income amount", "existing deposits", and "other debts"; another key semantic node set may be about repayment willingness, including key semantic nodes such as "whether to promise to repay" and "attitude towards overdue".

[0054] Then, for this key semantic node set about repayment ability, the server will generate the semantic feature domain corresponding to the session data stream based on it. The server will operate on any session semantic unit in the session data stream. For example, in a conversation between customer service and user A, user A said, "I used all my salary last month to pay the rent, and there is not much left." This session semantic unit is related to the penetration paragraph where the key semantic node set of repayment ability is located (that is, the paragraph about the impact of capital income and expenditure on repayment ability). The server will obtain the first correlation between this session semantic unit and the penetration paragraph. The correlation here may be relatively high, assuming it is 0.8. Based on this first correlation, since the semantic feature value of the session semantic unit is positively correlated with the first correlation, the semantic feature value of this session semantic unit is determined to be 0.8. For another example, user A said, "I ate bread this morning." The correlation between this session semantic unit and the penetration paragraph where repayment ability is located is very low. Assuming the correlation is 0.1, the semantic feature value is 0.1. The server generates this semantic feature domain based on the semantic feature values ​​of each session semantic unit in the session data stream. This semantic feature domain is like a quantitative description of the semantics related to repayment ability in the conversation data stream. It characterizes the penetration of the key semantic node set of repayment ability in the entire conversation data stream, that is, which parts are closely related to the conversation semantic unit in terms of repayment ability, and which parts are loosely related.

[0055] The same operation process is used for the key semantic node set of repayment willingness. For example, if user A says "I know that overdue is not good, I will repay as soon as possible", the correlation between this conversation semantic unit and the penetration paragraph (paragraph about attitude towards overdue repayment) where the key semantic node set of repayment willingness is located is very high. Assuming the correlation is 0.9, the semantic feature value is 0.9; if user A says "the weather is bad recently", the correlation between this conversation semantic unit and the penetration paragraph where the repayment willingness is located is extremely low. Assuming it is 0.05, the semantic feature value is 0.05, thus generating a semantic feature domain about repayment willingness.

[0056] After obtaining the semantic feature domains corresponding to multiple key semantic node sets (such as repayment ability, repayment willingness, etc.), the server will aggregate the enhanced session data stream and the session data stream under the constraints of these semantic feature domains to generate a fictitious session data stream. Taking user A as an example, for the semantic feature domain of repayment ability, the server will optimize the enhanced session data stream to generate the first composite session data stream, and optimize the session data stream to generate the second composite session data stream. In the enhanced session data stream, for example, if the previously determined semantic feature value of a certain session semantic unit is 0.7, then during the optimization process, this session semantic unit will be adjusted according to this semantic feature value, for example, its semantic weight will be adjusted according to 0.7, so as to generate the corresponding part in the first composite session data stream. For a certain session semantic unit in the session data stream, if its semantic feature value in the inverse semantic feature domain corresponding to the semantic feature domain of repayment ability is 0.3 (because the sum of the semantic feature values ​​of the session semantic units in the same penetration paragraph in the semantic feature domain and the inverse semantic feature domain is 1), then the session semantic unit is fused according to the semantic feature value of 0.3 to generate the part in the second composite session data stream. Then, based on the first composite session data stream and the second composite session data stream, a composite session data stream corresponding to the semantic feature domain of repayment ability is generated.

[0057] The same operation is also applied to the semantic feature domains corresponding to other key semantic node sets such as willingness to repay. Finally, the server aggregates the composite session data streams corresponding to all these key semantic node sets to generate a fictitious session data stream. In this fictitious session data stream, the key semantic nodes of user A (such as expressions related to repayment ability and willingness to repay) show inconsistency with the key semantic nodes in the original session data stream, while some non-key semantic nodes that are not related to repayment ability and willingness to repay, such as some daily trivia mentioned by user A, remain consistent. This completes the generation process from the original session data stream and the enhanced session data stream to the fictitious session data stream, providing a special data structure for subsequent operations, which helps to effectively analyze and judge risk fraud events in collection scenarios.

[0058] In a possible implementation, step S1222 includes: Step S1222-1, for any one of the various corresponding semantic feature domains of the multiple key semantic node sets, optimize the enhanced session data stream according to the semantic feature domain to generate a first composite session data stream, and optimize the session data stream according to the inverse semantic feature domain corresponding to the semantic feature domain to generate a second composite session data stream, wherein the sum of the semantic feature values ​​of the session semantic units of the same penetration paragraph in the semantic feature domain and the inverse semantic feature domain is 1.

[0059] Step S1222-2: Generate a composite session data stream corresponding to the semantic feature domain according to the first composite session data stream and the second composite session data stream.

[0060] Step S1222-3, aggregating the composite conversation data streams corresponding to the plurality of the semantic feature domains to generate the fictitious conversation data stream.

[0061] In a possible implementation, step S1222-1 includes: Step S1222-11: for any session semantic unit in the enhanced session data stream, multiply the session semantic unit by the semantic feature value corresponding to the session semantic unit in the semantic feature domain to generate the first composite session data stream.

[0062] Step S1222-12: for any session semantic unit in the session data stream, fuse the session semantic unit with the semantic feature value corresponding to the session semantic unit in the inverse semantic feature domain to generate the second composite session data stream.

[0063] In this embodiment, taking the case of user A with overdue credit card as an example, after obtaining the semantic feature domains corresponding to multiple key semantic node sets (such as repayment ability, repayment willingness, etc.) about user A, the server starts the aggregation operation.

[0064] For the semantic feature domain of repayment ability, the server will optimize the enhanced session data stream based on it to generate the first composite session data stream, and optimize the session data stream based on its corresponding inverse semantic feature domain to generate the second composite session data stream. In the enhanced session data stream, for example, there is a session semantic unit that is "In addition to the credit card arrears, I also have a car loan to pay back, which is very stressful." It was previously determined that the semantic feature value of this session semantic unit in the semantic feature domain of repayment ability is 0.8. Then the server will multiply this session semantic unit with its corresponding 0.8 in this semantic feature domain to obtain an adjusted session semantic unit, and this adjusted session semantic unit will become part of the first composite session data stream. In this way, the server operates on any session semantic unit in the enhanced session data stream and finally generates the first composite session data stream.

[0065] For the conversation data stream, for the conversation semantic units therein, for example, there is a conversation semantic unit that is "I originally planned to use the bonus to pay off my credit card, but the bonus was not issued", assuming that the semantic feature value of this conversation semantic unit in the inverse semantic feature domain corresponding to the repayment ability semantic feature domain is 0.2 (because in the semantic feature domain and the inverse semantic feature domain, the sum of the semantic feature values ​​of the conversation semantic units of the same penetration paragraph is 1). The server will fuse this conversation semantic unit with the corresponding 0.2 in the inverse semantic feature domain. This fusion operation may be based on a certain weighted calculation method, adjusting the semantics of this conversation semantic unit according to the weight of 0.2, thereby generating a part in the second composite conversation data stream. In this way, the server operates on any conversation semantic unit in the conversation data stream, and finally generates the second composite conversation data stream.

[0066] Next, a composite session data stream corresponding to the semantic feature domain of repayment ability is generated based on the generated first composite session data stream and the second composite session data stream. The composite session data stream integrates information of the optimized enhanced session data stream and the session data stream in the semantic feature domain of repayment ability.

[0067] The same operation process is applied to the semantic feature domain corresponding to the key semantic node set of willingness to repay. For example, in the enhanced conversation data stream, the semantic feature value of the conversation semantic unit "I know that overdue payment will affect my credit, and I want to repay as soon as possible" in the semantic feature domain of willingness to repay is 0.9, and it is multiplied by this semantic feature value to obtain the adjusted conversation semantic unit as part of the first composite conversation data stream; in the conversation data stream, the semantic feature value of the conversation semantic unit "I have always repaid on time before" in the inverse semantic feature domain corresponding to the semantic feature domain of willingness to repay is 0.1, and it is fused with 0.1 to obtain a part of the second composite conversation data stream, and then the composite conversation data stream corresponding to the semantic feature domain of willingness to repay is generated.

[0068] Finally, the server aggregates the composite session data streams corresponding to all these semantic feature domains (such as repayment ability, repayment willingness, etc.). By combining these composite session data streams according to certain rules, a fictitious session data stream is generated. In this fictitious session data stream, the key semantic nodes of user A (such as expressions related to repayment ability and repayment willingness) show inconsistency with the key semantic nodes in the original session data stream, while some non-key semantic nodes remain consistent. In this way, the generation process from enhanced session data streams and session data streams to fictitious session data streams is completed under the constraints of various corresponding semantic feature domains of multiple key semantic node sets. This fictitious session data stream is of great significance for the subsequent identification of risk fraud events in collection scenarios.

[0069] In a possible implementation, step S121 includes: Step S1211, preliminarily segmenting the conversation data stream according to semantic integrity, generating multiple independent conversation semantic units, identifying the vocabulary type in each conversation semantic unit, and marking the key entities and semantic roles in each conversation semantic unit based on the vocabulary type to generate marking information for each conversation semantic unit, wherein the key entities include debt information related to collection, financial products involved, identity-related information of the collection target user, and domain vocabulary related to risk fraud, and the semantic roles include agent and patient.

[0070] Step S1212, based on the tag information of each conversation semantic unit, construct a corresponding semantic association knowledge base, the construction of the semantic association knowledge base is obtained by analyzing the corpus related to debt collection and the general corpus, specifically for the nouns in each conversation semantic unit, establish associations with other related nouns in the semantic association knowledge base, and for the verbs in each conversation semantic unit, establish semantic associations with different subject and object combinations, and for adjectives and adverbs, establish semantic associations between them and the modified objects and semantic change associations in different contexts.

[0071] Step S1213: for each conversation semantic unit, extract the core semantics of the conversation semantic unit according to the semantic association knowledge base to obtain a conversation semantic unit set containing the core semantics.

[0072] Step S1214, for each conversation semantic unit containing core semantics, semantic expansion is performed according to the semantic association knowledge base, wherein, for the noun part in the core semantics, other nouns, adjectives, and verbs related thereto are searched from the semantic association knowledge base and then expanded into the conversation semantic unit, and, for the verb part in the core semantics, its related synonyms, near-synonyms, and different tenses and voice forms are searched and then expanded into the conversation semantic unit, thereby obtaining a set of conversation semantic units after semantic expansion.

[0073] Step S1215, performing a semantic conversion operation on the basis of the semantically expanded conversation semantic unit set, converting part of the semantics in the conversation semantic unit according to the semantic association knowledge base, and obtaining a semantically converted conversation semantic unit set.

[0074] Step S1216, based on the prior context information of the collection conversation, context adaptability adjustment is performed on the conversation semantic unit set after semantic conversion to obtain a context adapted conversation semantic unit set.

[0075] Step S1217, analyze the logical relationship between each conversation semantic unit in the context-adapted conversation semantic unit set, and determine whether there are semantic jumps and semantic contradictions between adjacent conversation semantic units. If so, adjust the relevant conversation semantic units to obtain a logically coherent conversation semantic unit set.

[0076] Step S1218: reassemble the logically coherent conversation semantic units according to the order in the original implementation conversation data stream to form an enhanced conversation data stream.

[0077] In this embodiment, taking the conversation data stream of user A whose credit card is overdue as mentioned above as an example, the server first preliminarily segments the conversation data stream according to semantic integrity, thereby generating multiple independent conversation semantic units. For example, in a conversation between customer service and user A, "I spent a lot of money in hospital due to illness, so my credit card is overdue" will be segmented into a conversation semantic unit. After that, the server begins to identify the vocabulary types in this conversation semantic unit, such as "sick hospital" belongs to a noun phrase, "spent" is a verb, "a lot of money" is a noun phrase, "credit card" is a noun, and "overdue" is a verb. Based on these vocabulary types, the server marks key entities and semantic roles. Among them, the key entity includes "credit card", which is debt information related to collection, and "I (user A)" is the agent, indicating that the action of "spending money in hospital due to illness" was performed, which led to the patient situation of "credit card overdue", thus generating the marking information of this conversation semantic unit.

[0078] For the tag information of each conversation semantic unit, the server will build a corresponding semantic association knowledge base. This knowledge base is obtained by analyzing the corpus related to debt collection and general corpus. For the noun "sick hospitalization" in this conversation semantic unit, establish an association relationship with other related nouns in the semantic association knowledge base, such as "medical expenses" and "medical insurance", etc.; for the verb "spent", establish its semantic association with different subject and object combinations, such as "he spent his savings on travel"; for the adjective "a lot", establish its semantic association with the modified object "money", as well as the semantic change association in different contexts, such as "a lot of money" when describing poor families and rich families, the quantity range represented by the semantics of "a lot" may be different.

[0079] Next, based on this semantic association knowledge base, the server extracts the core semantics for each conversation semantic unit. For the above conversation semantic unit, the core semantics may be "the cost of hospitalization due to illness caused the credit card to be overdue." In this way, a set of conversation semantic units containing core semantics is obtained. Then semantic expansion is performed for each conversation semantic unit containing core semantics. For the noun part "hospitalization due to illness" in the core semantics, other nouns, adjectives, and verbs related to it are searched from the semantic association knowledge base and then expanded to the conversation semantic unit, for example, it is expanded to "because of a sudden serious illness that requires hospitalization, a lot of medical expenses were spent"; for the verb part "spend" in the core semantics, its related synonyms, near synonyms, and different tenses and voice forms are searched and expanded to the conversation semantic unit, such as "consume", etc., so as to obtain a set of conversation semantic units after semantic expansion.

[0080] On the basis of the set of conversation semantic units obtained after semantic expansion, the server performs semantic conversion operations according to the semantic association knowledge base. For example, "a lot of medical expenses were spent" is converted into "high medical expenses were generated" to obtain a set of conversation semantic units after semantic conversion. Afterwards, the set of conversation semantic units after semantic conversion is context-adapted based on the prior context information of the collection conversation. If the previous conversation mentioned that the bank has a special policy for overdue payments due to illness, then the conversation semantic unit can be adjusted to "Although I know that the bank has a special policy for overdue payments due to illness, because of a sudden serious illness that requires hospitalization, high medical expenses were incurred, so the credit card is overdue", and a set of conversation semantic units after context adaptation is obtained.

[0081] Then analyze the set of conversation semantic units after context adaptation to determine the logical relationship between each conversation semantic unit. For example, in this set, if there is a conversation semantic unit saying "I don't have money to pay my credit card" and the next conversation semantic unit says "I just bought a very expensive computer", there will be a semantic jump and semantic contradiction. The server will adjust the relevant conversation semantic units, such as adjusting the latter conversation semantic unit to "During my illness, my family borrowed money to buy a very expensive computer to facilitate my recovery, so now I don't have money to pay my credit card", so as to obtain a logically coherent conversation semantic unit set.

[0082] Finally, these logically coherent conversation semantic units are reassembled in the order of the original conversation data stream. For example, if the original conversation data stream first mentions being hospitalized due to illness, and then mentions overdue credit card payments, then this order is followed when reassembling, thus forming an enhanced conversation data stream. Compared with the conversation semantic units in the original conversation data stream, the conversation semantic units in this enhanced conversation data stream have been processed and optimized in many aspects, and are richer, more accurate, and more logically coherent in semantic expression, providing a better data foundation for subsequent further operations.

[0083] In a possible implementation, step S130 includes: Step S131, using the collection conversation decision network to obtain a first mapping knowledge vector of the sample collection conversation training data, wherein the first mapping knowledge vector represents semantic transition information in the time domain of a semantic representation vector corresponding to the collection target user in the sample collection conversation training data.

[0084] Step S132, using the collection session decision network to generate a second mapping knowledge vector based on the first mapping knowledge vector, wherein the second mapping knowledge vector represents the semantic transformation information in the spatial domain of the semantic representation vector corresponding to the collection target user in the sample collection session training data.

[0085] Step S133, using the collection session decision network to generate a decision result of the sample collection session training data based on the second mapping knowledge vector, the decision result of the sample collection session training data characterizes whether there is a risk of fraud event in the semantic changes of the semantic representation vector associated with the key semantic node of the collection target user in the sample collection session training data.

[0086] Step S134, performing parameter learning on the collection session decision network based on the decision result, and generating the collection session decision network that has completed parameter learning.

[0087] In this embodiment, taking the sample collection conversation training data related to user A whose credit card is overdue as an example, the server first uses the collection conversation decision network to obtain the first mapping knowledge vector of the sample collection conversation training data. In the conversation data of user A, as the conversation progresses, there is semantic transition information in the time domain for its semantic representation vector. For example, at the beginning of the conversation, user A mentioned that his credit card was overdue due to huge expenses for hospitalization due to illness. At this time, the semantic representation vector reflects a situation where repayment is difficult due to unexpected expenses. As the conversation progresses, user A also said that he is working hard to raise funds to repay the loan, which is a transition of the semantic representation vector in the time domain. The collection conversation decision network will encode these semantic transition information in the time domain to form a first mapping knowledge vector. This first mapping knowledge vector can accurately characterize the semantic transition of the semantic representation vector corresponding to user A in the entire sample collection conversation training data in the time domain.

[0088] Next, the server uses the collection conversation decision network to generate a second mapping knowledge vector based on the first mapping knowledge vector. In this process, the network will process the first mapping knowledge vector to obtain a second mapping knowledge vector that reflects the semantic transformation information of the semantic representation vector in the spatial domain. For example, for the conversation data of user A, the information in the first mapping knowledge vector may be arranged in the order in which the conversation occurred, and after being processed by the network, this information will be rearranged in the spatial domain according to factors such as semantic importance. For example, the semantic information related to the statement about repayment ability, such as source of income, debt status, etc., is reorganized in the spatial domain according to its importance in judging risk fraud events, thereby generating a second mapping knowledge vector. This second mapping knowledge vector can characterize the semantic transformation information in the spatial domain of the semantic representation vector corresponding to user A in the sample collection conversation training data.

[0089] Then, the server generates the decision result of the sample collection session training data based on the second mapping knowledge vector using the collection session decision network. For the sample collection session training data of user A, based on the semantic transition information contained in the second mapping knowledge vector, it is determined whether the semantic change of the semantic representation vector associated with the key semantic node of user A in the sample collection session training data is a risk fraud event. If in the user A's session, from expressing efforts to repay to suddenly appearing some unreasonable explanations of the source of funds, or a sudden change in repayment attitude, these semantic changes match the known risk fraud patterns, then the decision result will determine that there is a risk fraud event. For example, if user A initially said that he had a meager income and was trying to repay the loan, but later claimed that he had a large amount of funds of unknown origin and was unwilling to explain the source in detail to repay the loan, but did not actually repay the loan, this semantic change may be determined as a risk fraud event.

[0090] Finally, the collection session decision network is parameter-learned based on this decision result. If the decision result determines that there is a risk fraud event, but it is actually a misjudgment, or vice versa, a training error will be generated. The server will determine the training error parameters of the collection session decision network based on this decision result. For example, if the judgment of user A is that there is a risk fraud event, but in fact user A just can't express it clearly, this is a misjudgment, which will produce a certain training error. Then, the server locks the weight parameter information of the first initialization encoder and the weight parameter information of the second initialization encoder in the collection session decision network, and optimizes the weight parameter information of the first recurrent neural network and the weight parameter information of the second recurrent neural network based on this training error parameter. Through such an optimization process, the network can better adapt to the data, continuously adjust its own parameters to improve the accuracy of the judgment of risk fraud events, and finally generate a collection session decision network that has completed parameter learning. This collection session decision network that has completed parameter learning can more accurately analyze the new collection session data and determine whether there is a risk fraud event.

[0091] In a possible implementation, step S132 includes: Step S1321, perform attention weight allocation processing on the first mapping knowledge vector to generate a first mapping knowledge vector after the attention weight allocation processing.

[0092] Step S1322: Generate the second mapping knowledge vector by using the first mapping knowledge vector processed by the collection conversation decision network according to the attention weight allocation.

[0093] Among them, the knowledge fragments in the first mapping knowledge vector are arranged in the time domain, and the knowledge fragments in the first mapping knowledge vector after the attention weight allocation processing are arranged in the spatial domain.

[0094] In one possible implementation, the collection session decision network includes a time domain knowledge encoder and a space domain knowledge encoder, the time domain knowledge encoder is generated based on a first initialization encoder and a first recurrent neural network, the space domain knowledge encoder is generated based on a second initialization encoder and a second recurrent neural network, the first recurrent neural network is used to obtain the first mapping knowledge vector based on the first initialization encoder, and the second recurrent neural network is used to obtain the second mapping knowledge vector based on the second initialization encoder.

[0095] Step S134 includes: Step S1341: determining a training error parameter of the collection session decision network based on the decision result.

[0096] Step S1342, lock the weight parameter information of the first initialized encoder and the weight parameter information of the second initialized encoder, optimize the weight parameter information of the first recurrent neural network and the weight parameter information of the second recurrent neural network based on the training error parameters, and generate the collection session decision network that completes parameter learning.

[0097] In this embodiment, taking the collection session data of the previously mentioned overdue credit card user A as an example, when the server generates the second mapping knowledge vector based on the first mapping knowledge vector using the collection session decision network, the first mapping knowledge vector must first be processed for attention weight allocation. In the collection session of user A, the first mapping knowledge vector contains knowledge fragments arranged in the time domain, which reflect the semantic transformation information of user A throughout the session. For example, in the early stages of the session, user A mentioned that he had spent a lot of money in the hospital due to illness, which caused his credit card to be overdue. This information exists as a knowledge fragment in the first mapping knowledge vector; as the session progresses, user A indicates that he is trying to repay the credit card debt by borrowing from relatives, which is another knowledge fragment. These knowledge fragments are arranged in the order in which the conversation occurs, that is, in the time domain.

[0098] When the server performs the attention weight allocation process, it will allocate weights according to the importance of each knowledge fragment to the judgment of risk fraud events. For example, for knowledge fragments directly related to repayment ability, such as income source, debt situation, etc., a higher attention weight will be assigned. For relatively unimportant information such as the hospital environment during the illness and hospitalization mentioned by user A, a lower weight is assigned. Through such weight allocation, the first mapping knowledge vector after the attention weight allocation process is generated, and the knowledge fragments therein are arranged in the spatial domain. This transition from time domain arrangement to spatial domain arrangement is the result of reorganizing the knowledge fragments based on attention weights. For example, all knowledge fragments related to repayment ability and with higher weights will be put together, while knowledge fragments that are not closely related to repayment ability will be placed in other positions.

[0099] Then, the collection conversation decision network is used to generate a second mapping knowledge vector based on the first mapping knowledge vector processed by the attention weight allocation. The time domain knowledge encoder in the collection conversation decision network is generated based on the first initialization encoder and the first recurrent neural network, and the first recurrent neural network obtains the first mapping knowledge vector based on the first initialization encoder. The spatial domain knowledge encoder is generated based on the second initialization encoder and the second recurrent neural network. In this process, the second recurrent neural network is based on the second initialization encoder, takes the first mapping knowledge vector processed by the attention weight allocation as input, and generates a second mapping knowledge vector after a series of complex calculations and processing. This second mapping knowledge vector represents the semantic transformation information of the corresponding semantic representation vector of user A in the sample collection conversation training data in the spatial domain. For example, in the spatial domain, the relationship between different semantically related knowledge fragments is reconstructed, which can more clearly reflect the changes in semantics in space, which helps to more comprehensively analyze whether there are risky fraud events in user A's conversation data.

[0100] Next, when the collection session decision network is parameter-learned based on the decision results to generate a collection session decision network that has completed parameter learning, the server must first determine the training error parameters of the collection session decision network based on the decision results. For the collection session decision results of user A, if it is determined that there is a risk of fraud, but after further investigation it is found that user A is just unclear due to some misunderstandings in communication, and there is actually no fraud, this will result in a misjudgment, resulting in a training error. Conversely, if it is determined that there is no risk of fraud, but in fact user A has committed fraud, this is also a misjudgment.

[0101] After determining the training error parameters, the server will lock the weight parameter information of the first initialization encoder and the weight parameter information of the second initialization encoder. This is because the weight parameters of the first initialization encoder and the second initialization encoder have a certain stability in the network structure or have been determined through preliminary optimization. Then, the weight parameter information of the first recurrent neural network and the weight parameter information of the second recurrent neural network are optimized based on this training error parameter. When optimizing the weight parameter information of the first recurrent neural network, the weights related to the processing of the first mapping knowledge vector will be adjusted according to the training error parameters. For example, if a certain weight causes a deviation in the processing of the knowledge fragment related to the repayment ability of user A, thereby affecting the final decision result, then this weight will be adjusted according to the training error. The optimization of the weight parameter information of the second recurrent neural network is also a similar process. The weights related to the generation of the second mapping knowledge vector are adjusted according to the training error parameters to improve the accuracy of the processing of semantic transformation information in the spatial domain.

[0102] Through this process of continuously adjusting the weight parameter information of the first recurrent neural network and the second recurrent neural network based on the decision results, the server gradually optimizes the collection session decision network and finally generates a collection session decision network that has completed parameter learning. This collection session decision network that has completed parameter learning can more accurately determine whether there are risk fraud events when processing new collection session data. For example, when processing the collection session data of a new credit card overdue user B, this network can more accurately analyze the conversation semantics of user B, determine whether he has risk fraud behavior, reduce the possibility of misjudgment, and thus improve the efficiency and accuracy of the entire collection process. In actual applications, as more and more sample collection session training data are used for training, the collection session decision network will continue to be optimized and can adapt to the risk fraud judgment needs in various complex collection scenarios.

[0103] To further elaborate on the case of user A, suppose that in a collection session, user A mentioned that he was sick and hospitalized and spent a lot of money, that he had other debts in addition to his credit card debt, and that he was currently unemployed. This information constitutes part of the knowledge fragments in the first mapping knowledge vector in the order of the time domain. When performing the attention weight allocation process, knowledge fragments such as "no job" and "other debts" that are directly related to repayment ability are assigned higher weights, because these factors are crucial to judging whether user A has risky fraud. Some details in "sick and hospitalized and spent a lot of money", such as the name of the hospital, the level of the ward, and other relatively less important information, are assigned lower weights. After the attention weight allocation process, these knowledge fragments are rearranged in the airspace. Then, the collection session decision network generates the second mapping knowledge vector based on this processed first mapping knowledge vector, and the semantic transition information in the airspace is clearer at this time. For example, the relationship changes between knowledge fragments related to repayment ability and their association with other information can be clearly seen.

[0104] When learning parameters based on decision results, if user A is initially judged to have a risk of fraud, because user A's expression of his attitude and ability to repay in the conversation is somewhat vague, but later it is found that it is because user A has just recovered from an illness and his thinking is not clear enough, so it is not actually a fraud. This misjudgment will produce a training error. Based on this training error, the server optimizes the weight parameter information of the first recurrent neural network and the second recurrent neural network on the basis of locking the weight parameter information of the first initialization encoder and the second initialization encoder. For the first recurrent neural network, the weights of the knowledge fragments related to processing user A's repayment ability may be adjusted so that it can more accurately reflect the actual situation when processing similar situations in the future. For the second recurrent neural network, the weights related to constructing spatial semantic relationships may be adjusted to better process the semantic transformation information in the spatial domain and avoid similar misjudgments from happening again. Through such continuous optimization, the collection session decision network will be more accurate and reliable when processing similar collection sessions.

[0105] In the collection business of the entire financial institution, there is a large amount of collection session data that needs to be processed. Different users have different situations. For example, some users may have credit card overdue due to sudden economic difficulties, such as unemployment, investment failure, etc., while some users may have malicious fraud and deliberately default on their debts. The collection session decision network needs to accurately distinguish these situations, and through the above parameter learning process, the network can continuously improve its accuracy. Take another user C as an example. User C's credit card is overdue due to investment failure. In the collection session, user C mentioned information such as the project he invested in, the amount of loss, and the current source of income. The server processes the collection session data of user C in the same process, from obtaining the first mapping knowledge vector, performing attention weight allocation processing to generate the second mapping knowledge vector, to parameter learning based on the decision results. In this process, each link is crucial to accurately determine whether user C has risky fraud. For example, when determining the training error parameter, if the misjudgment is caused by insufficient understanding of user C's investment project, then it is necessary to accurately identify this problem and adjust it when optimizing the network weight parameters to ensure that the network can make more accurate judgments when processing collection sessions of similar users.

[0106] As more and different types of collection conversation data are used for training, the collection conversation decision network will continue to learn new patterns and features, and can better adapt to various complex collection scenarios. Whether facing simple users who are overdue due to unexpected expenses, or facing complex users who may have fraudulent behavior, the collection conversation decision network that has completed parameter learning can accurately analyze the conversation semantics and determine whether there is a risk of fraud, thereby providing strong support for the collection work of financial institutions, improving collection efficiency, and reducing risk losses.

[0107] Figure 2 The hardware structure of the information decision system 100 based on the collection conversation big data analysis for implementing the above-mentioned information decision method based on the collection conversation big data analysis provided by the embodiment of the present invention is shown as follows: Figure 2 As shown, the information decision system 100 based on the collection conversation big data analysis may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0108] 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 information decision system 100 based on the collection conversation big data analysis uses to execute or use to complete the exemplary method described in the present invention.

[0109] 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 information decision-making method based on the collection conversation big data analysis of 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.

[0110] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned information decision system 100 based on collection conversation big data analysis. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0111] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned information decision-making method based on the collection conversation big data analysis is implemented.

[0112] 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. An information decision-making method based on big data analysis of collection conversations, characterized in that: The method comprises: Acquire collection conversation big data, the collection conversation big data including a plurality of conversation data streams associated with collection target users sorted according to the order of collection conversation nodes; For a session data streams in the collection session big data, the key semantic nodes of the collection target user in the session data streams are enhanced to generate sample collection session training data, and the semantic representation vectors associated with the key semantic nodes of the collection target user have risk fraud events in the sample collection session training data, where a is a positive integer; Parameter learning is performed on the collection session decision network based on the sample collection session training data to generate a collection session decision network that has completed parameter learning. The collection session decision network that has completed parameter learning is used to decide whether there is a risk of fraud in the collection session data.

2. The information decision-making method based on collection conversation big data analysis according to claim 1 is characterized in that: The step of enhancing the key semantic nodes of the target user for collection in the conversation data stream to generate sample collection conversation training data includes: enhancing each session semantic unit in the session data stream to generate an enhanced session data stream, wherein the session semantic units in the enhanced session data stream are inconsistent with the session semantic units in the session data stream; Aggregating the enhanced session data stream and the session data stream to generate a fictitious session data stream, wherein key semantic nodes of the target user for collection in the fictitious session data stream show inconsistency with key semantic nodes of the target user for collection in the session data stream, and non-key semantic nodes of the target user for collection in the fictitious session data stream show consistency with non-key semantic nodes of the target user for collection in the session data stream; The a conversation data streams are overwritten with various fictitious conversation data streams corresponding to the a conversation data streams to generate the sample collection conversation training data.

3. The information decision-making method based on collection conversation big data analysis according to claim 2 is characterized in that: The step of aggregating the enhanced session data stream and the session data stream to generate a fictitious session data stream includes: Acquire multiple key semantic node sets corresponding to the target user for collection in the session data stream, each of the key semantic node sets including multiple key semantic nodes; For any one of the multiple key semantic node sets, generating a semantic feature domain corresponding to the session data stream according to the key semantic node set, wherein the semantic feature domain represents a penetration segment of the key semantic node set in the session data stream; Under the restriction of various corresponding semantic feature domains of the plurality of key semantic node sets, the enhanced session data stream and the session data stream are aggregated to generate the fictitious session data stream.

4. The information decision-making method based on collection conversation big data analysis according to claim 3 is characterized in that: The generating, based on the key semantic node set, a semantic feature domain corresponding to the session data flow includes: For any session semantic unit in the session data stream, obtaining a first correlation between the session semantic unit and the penetration paragraph where the key semantic node set is located; Based on the first correlation, determining a semantic feature value of the conversation semantic unit, wherein the semantic feature value of the conversation semantic unit is in a positive correlation relationship with the first correlation; The semantic feature domain is generated according to the semantic feature value of each conversation semantic unit in the conversation data stream.

5. The information decision-making method based on collection conversation big data analysis according to claim 4 is characterized in that: The step of aggregating the enhanced session data stream and the session data stream under the constraints of the semantic feature domains corresponding to the plurality of key semantic node sets to generate the fictitious session data stream includes: For any one of the semantic feature domains corresponding to the multiple key semantic node sets, optimizing the enhanced session data stream according to the semantic feature domain to generate a first composite session data stream, and optimizing the session data stream according to an inverse semantic feature domain corresponding to the semantic feature domain to generate a second composite session data stream, wherein the sum of the semantic feature values ​​of the session semantic units of the same penetration paragraph in the semantic feature domain and the inverse semantic feature domain is 1; generating a composite session data stream corresponding to the semantic feature domain according to the first composite session data stream and the second composite session data stream; Aggregate the composite conversation data streams corresponding to the plurality of the semantic feature domains to generate the fictitious conversation data stream.

6. The information decision-making method based on collection conversation big data analysis according to claim 5 is characterized in that: The step of optimizing the enhanced session data stream according to the semantic feature domain to generate a first composite session data stream, and optimizing the session data stream according to an inverse semantic feature domain corresponding to the semantic feature domain to generate a second composite session data stream includes: For any session semantic unit in the enhanced session data stream, multiply the session semantic unit by a semantic feature value corresponding to the session semantic unit in the semantic feature domain to generate the first composite session data stream; For any session semantic unit in the session data stream, the session semantic unit is merged with the semantic feature value corresponding to the session semantic unit in the inverse semantic feature domain to generate the second composite session data stream.

7. The information decision-making method based on collection conversation big data analysis according to any one of claims 2 to 6, characterized in that: The step of enhancing each session semantic unit in the session data stream to generate an enhanced session data stream includes: Preliminarily segmenting the session data stream according to semantic integrity to generate multiple independent session semantic units, identifying the vocabulary type in each session semantic unit, and marking the key entities and semantic roles in each session semantic unit based on the vocabulary type to generate marking information for each session semantic unit, wherein the key entities include debt information related to debt collection, financial products involved, identity-related information of target users for debt collection, and domain vocabulary related to risk fraud, and the semantic roles include agent and recipient; Based on the tag information of each conversation semantic unit, a corresponding semantic association knowledge base is constructed. The construction of the semantic association knowledge base is obtained by analyzing the corpus related to debt collection and the general corpus. Specifically, for the nouns in each conversation semantic unit, an association relationship with other related nouns is established in the semantic association knowledge base. For the verbs in each conversation semantic unit, its semantic association with different subject and object combinations is established. For adjectives and adverbs, the semantic association between them and the modified objects and the semantic change association in different contexts is established. For each conversation semantic unit, extract the core semantics of the conversation semantic unit according to the semantic association knowledge base to obtain a conversation semantic unit set containing the core semantics; For each conversation semantic unit containing core semantics, semantic expansion is performed according to the semantic association knowledge base, wherein, for the noun part in the core semantics, other nouns, adjectives, and verbs related thereto are searched from the semantic association knowledge base and then expanded into the conversation semantic unit, and, for the verb part in the core semantics, its related synonyms, near synonyms, and different tenses and voice forms are searched and then expanded into the conversation semantic unit, thereby obtaining a set of conversation semantic units after semantic expansion; Performing a semantic conversion operation on the basis of the semantically expanded conversation semantic unit set, converting part of the semantics in the conversation semantic unit according to the semantic association knowledge base, and obtaining a semantically converted conversation semantic unit set; Based on the prior context information of the collection conversation, the conversation semantic unit set after the semantic conversion is adjusted for context adaptability to obtain the conversation semantic unit set after context adaption; Analyze the logical relationship between each conversation semantic unit in the conversation semantic unit set after context adaptation, and determine whether there is a semantic jump or semantic contradiction between adjacent conversation semantic units. If so, adjust the relevant conversation semantic units to obtain a logically coherent conversation semantic unit set. The logically coherent conversation semantic units are reassembled according to the order in the original implementation conversation data stream to form an enhanced conversation data stream.

8. The information decision-making method based on collection conversation big data analysis according to claim 1 is characterized in that: The step of performing parameter learning on the collection session decision network based on the sample collection session training data to generate a collection session decision network that has completed parameter learning includes: Using the collection conversation decision network to obtain a first mapping knowledge vector of the sample collection conversation training data, the first mapping knowledge vector representing semantic transition information in the time domain of a semantic representation vector corresponding to the collection target user in the sample collection conversation training data; Using the collection conversation decision network to generate a second mapping knowledge vector based on the first mapping knowledge vector, the second mapping knowledge vector represents the semantic transformation information of the semantic representation vector corresponding to the collection target user in the sample collection conversation training data in the spatial domain; Using the collection conversation decision network to generate a decision result of the sample collection conversation training data according to the second mapping knowledge vector, the decision result of the sample collection conversation training data characterizes whether there is a risk fraud event in the semantic change of the semantic representation vector associated with the key semantic node of the collection target user in the sample collection conversation training data; Based on the decision result, parameter learning is performed on the debt collection conversation decision network to generate the debt collection conversation decision network that has completed parameter learning.

9. The information decision-making method based on collection conversation big data analysis according to claim 8 is characterized in that: The step of using the collection session decision network to generate a second mapping knowledge vector according to the first mapping knowledge vector includes: Performing attention weight allocation processing on the first mapping knowledge vector to generate a first mapping knowledge vector after the attention weight allocation processing; Generate the second mapping knowledge vector by using the first mapping knowledge vector processed by the collection conversation decision network according to the attention weight distribution; The knowledge fragments in the first mapping knowledge vector are arranged in the time domain, and the knowledge fragments in the first mapping knowledge vector after the attention weight allocation process are arranged in the space domain; The collection conversation decision network includes a time domain knowledge encoder and a space domain knowledge encoder, the time domain knowledge encoder is generated according to a first initialization encoder and a first recurrent neural network, the space domain knowledge encoder is generated according to a second initialization encoder and a second recurrent neural network, the first recurrent neural network is used to obtain the first mapping knowledge vector according to the first initialization encoder, and the second recurrent neural network is used to obtain the second mapping knowledge vector according to the second initialization encoder; The performing parameter learning on the collection session decision network based on the decision result to generate the collection session decision network that has completed parameter learning includes: Based on the decision result, determining a training error parameter of the collection session decision network; The weight parameter information of the first initialization encoder and the weight parameter information of the second initialization encoder are locked, and the weight parameter information of the first recurrent neural network and the weight parameter information of the second recurrent neural network are optimized based on the training error parameters to generate the collection session decision network that completes parameter learning.

10. An information decision-making system based on big data analysis of collection conversations, characterized in that: The information decision-making system based on collection conversation big data analysis 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 information decision-making method based on collection conversation big data analysis as described in any one of claims 1 to 9.

Images