Data processing method oriented to customer relationship management

Abstract

According to the customer relationship management-oriented data processing method provided by the invention, in a high-value scene such as mixed asset management, a customer instruction is expressed as a dynamic operation semantic tree by means of setting an annular buffer area, dynamically setting an alignment time window, adopting a reinjection mechanism and the like; therefore, the domain enhanced pre-training language model can perform reasoning based on the most accurate and timely domain facts, so that ambiguity in a client instruction is eliminated. By dynamically adjusting the conditional probability table, the reasoning process of the credibility probabilistic reasoning model can be adaptively adjusted along with the environment context, the expression ability of the model is greatly improved, and the risk assessment result is more accurate and fine. The credibility probability and the multi-dimensional feature vector of the path to be executed are used as input through the decision function, so that the system realizes conversion from condition judgment based on a threshold value to quantitative decision based on random optimization, and the customer instruction execution risk can be effectively controlled.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a data processing method for customer relationship management. BACKGROUND

[0002] With the diversification of high net worth client assets and the complication of interaction modes, a new and rare customer relationship management scenario has emerged: the mixed asset management scenario of simultaneously hosting physical assets (such as precious metals, art) and digital assets (such as digital keys, digital copyrights) for customers. In this scenario, customers initiate complex, cross-asset class operation instructions (for example, "authorize the exhibition of a certain painting and transfer the proceeds to a specific account or purchase a certain digital copyright") through multiple modalities such as voice, video, and text. The demand of the management institution in this scenario has changed from the traditional "maintaining customer satisfaction" to "safely, accurately, and correctly understanding and executing customer instructions in a highly adversarial environment, while avoiding irreversible and large loss risks".

[0003] In this scenario, the existing CRM system's technical architecture cannot meet the institution's needs, mainly in the following aspects: 1. The existing CRM system is good at handling preset processes and structured queries, but cannot technically understand the referential ambiguity (such as "it" refers to which asset) and semantic ambiguity (such as the specific definition of "market price") in natural language instructions and cross-asset dependency relationships. Its underlying lacks a technical means that can convert multi-modal interaction information (such as voice, text) into a formal operation semantic tree that can be unambiguously understood and executed by machines, resulting in a fundamental misunderstanding risk in the subsequent execution phase.

[0004] 2. Existing identity authentication technologies (such as multi-modal biometric recognition) can only solve the problem of "whether it is someone", but cannot answer the higher-level security question of "whether the instruction is issued under the free will of the customer". The existing CRM system lacks a technical framework that can fuse multi-dimensional weak signals (such as stress features in voiceprints, anomalies in interaction timing, deviations between instruction content and historical behavior) in real time and perform uncertainty reasoning, so it cannot calculate a dynamic, quantitative credibility probability value for each instruction as a basis for subsequent decision-making.

[0005] 3、Most importantly, the result of the existing CRM system management is determined and static, which cannot deal with the fundamental contradiction in the high-value scenario: both maximizing the efficiency of executing customer instructions and minimizing the loss caused by potential risks. The current customer management technology lacks a technical solution that can jointly model "instruction credibility" and "resource cost and risk coefficient of the back-end execution path" and solve the stochastic optimization control problem within milliseconds, so it cannot autonomously and intelligently select the optimal risk-reward ratio strategy among multiple execution paths (such as direct execution, requiring secondary confirmation, or starting a sandbox simulation). SUMMARY

[0006] The present application aims to accurately understand the customer's instruction intention in the mixed asset management scenario, dynamically evaluate the credibility of the instruction source in the adversarial environment, increase transaction security, maximize the efficiency of executing customer instructions while minimizing the loss caused by potential risks, and improve the reliability of resource scheduling. A data processing method for customer relationship management is provided.

[0007] To achieve this purpose, the following technical solutions are adopted: A data processing method for customer relationship management is provided, applied in a mixed asset management scenario, comprising the steps of: S1, dynamically generating an operation semantic tree based on multi-modal time alignment and domain knowledge injection; S2, dynamically calculating the credibility probability of customer instructions based on multi-source uncertainty reasoning; S3, dynamically generating the optimal security strategy for executing the customer instructions based on the pre-constructed credibility cost game model.

[0008] Preferably, step S1 specifically includes the steps of: A1, time alignment and fusion of multi-modal input; A2, using a domain-enhanced pre-training language model to understand the intention of the aligned data set obtained in step A1, while querying the identified preliminary entities and the background asset knowledge graph in real time, and feeding the query results back to the domain-enhanced pre-training language model; A3, dynamically generating an operation semantic tree based on the structured understanding results predicted by the model.

[0009] Preferably, step A1 specifically includes the steps of: A11, time stamping and packaging multi-modal data; A12, creating a dynamically scalable ring buffer and dynamically setting an alignment time window Then write the data packet after time stamping and packaging into the ring buffer, and automatically trigger alignment checking to screen out data packets of all modalities within a time range, representing the current moment in time; A13, binding each modal data packet screened in step A12 into an "alignment data group", and then releasing it to the domain-enhanced pre-training language model.

[0010] Preferably, step A2 specifically comprises the following steps: A21, processing the alignment data group obtained in step A1 in parallel and asynchronously through a bypass and a main path, the bypass processing method being: performing preliminary entity recognition on the text data transcribed from the alignment data group through a pre-constructed dictionary, forming a preliminary entity list from the recognized preliminary entities, and then initiating a batch query to the background asset knowledge graph, and structuring the query result into a knowledge prompt; the main path processing method being: determining whether the knowledge prompt is generated for the text data, if yes, inserting the knowledge prompt into the specified position in the text data, and then proceeding to step A22; if no, proceeding to step A22; A22, inputting or annotating the same text data currently being processed by the bypass or the text data processed by the bypass to the domain-enhanced pre-training language model, and the model predicting and outputting the structured understanding result.

[0011] Preferably, step A3 specifically comprises the following steps: A31, according to the defined domain syntax rules, matching the structured understanding result output by the model with the corresponding production rule template, and filling each data in the understanding result into the corresponding slot in the template; A32, taking the <operation> slot matched in step A31 as the root node of the semantic tree, and the remaining slots as the leaf nodes, and generating the production rule template with slot data filling as the operation semantic tree corresponding to the alignment data group.

[0012] Preferably, in step S2, the method of dynamically calculating the credibility probability of the customer instruction comprises the following steps: B1, calculating the multi-modal behavior deviation score and compliance conflict score using the operation semantic tree; B2, inputting the scores calculated in step B1 into a credibility probability inference model, and the credibility probability inference model dynamically adjusts the score weight according to the environmental context, and finally predicts and outputs the credibility probability of the customer instruction expressed in the operation semantic tree.

[0013] Preferably, the method for calculating the multi-modal behavioral deviation score in step B1 comprises the steps of: B11, the system continuously extracts behavioral features from each interaction with the same customer and maintains a probability distribution model for each of the behavioral features within a sliding time window as a dynamic baseline for the calculation of the deviation score; B12, extract the current behavior features of the customer from the operation semantic tree, and calculate the standard score relative to the dynamic baseline; B13, calculate the square root of the weighted sum of the standard scores associated with different types of current behavior features as the multi-modal behavior deviation score.

[0014] Preferably, the method for calculating the compliance conflict score comprises the steps of: C1, match the operation semantic tree with the dynamically updated compliance knowledge graph in terms of nodes and edges to search for conflicting paths in the compliance knowledge graph that exist in the operation semantic tree; C2, query the conflict severity values corresponding to each of the conflicting paths found to conflict with the operation semantic tree, calculate the square root of the weighted sum of each of the conflict severity values, and output the compliance conflict score of the customer at the current time.

[0015] Preferably, the method for model prediction outputting the credibility probability comprises the steps of: B21, construct the network topology structure of the credibility probability inference model depending on the pre-defined risk nodes and node states; B22, map the score calculated in step B1 to the node state of the corresponding node in the network topology structure; B23, dynamically adjust the conditional probability table, which comprises the steps of: B231, define the environmental context; B232, predefine a corresponding conditional probability table for each of the environmental contexts; B233, according to the environmental context in which the customer instruction is currently located, match the conditional probability table to be loaded from the pre-defined conditional probability tables; B24, perform probability propagation on the conditional probability table to be loaded and the node state mapped after step B22, and finally calculate and output the credibility probability of the overall credibility node in the network topology structure constructed in step B21.

[0016] Preferably, step S3 comprises the steps of: S31, multi-dimensional feature vector modeling of the path to be executed; S32, input the multi-dimensional feature vector of each proposed execution path and the associated confidence probability to the decision function, the function solves the expected loss value of each proposed execution path, and takes the proposed execution path with the minimum expected loss value as the optimal security strategy dynamically generated for the client instruction; The decision function Expressed as: ,in, Indicates the first The probability of credibility of the proposed execution path. Indicates the first The risk and cost of the proposed implementation path Indicates the first The execution cost of the proposed execution path.

[0017] This application has the following beneficial effects: 1. By setting up a circular buffer, dynamically setting alignment time windows, and employing a back-injection mechanism, client instructions are represented as dynamic operation semantic trees. This enables the domain-enhanced pre-trained language model to reason based on the most accurate and timely domain facts, thereby eliminating ambiguity in client instructions. By dynamically adjusting the conditional probability table, the reasoning process of the credibility probability inference model can adaptively adjust with the environmental context, greatly improving the model's expressive power and making risk assessment results more accurate and refined. By using the credibility probability and the multi-dimensional feature vector of the proposed execution path as input to the decision function, the system achieves a shift from threshold-based conditional judgment to quantitative decision-making based on stochastic optimization, effectively controlling the risk of client instruction execution.

[0018] 2. The provided circular buffer dynamically adjusts according to the actual workload and dynamically adjusts the alignment time window by automatically triggering alignment checks. It has self-management and self-optimization capabilities and can find the best balance between achieving the highest alignment success rate and the lowest additional latency.

[0019] 3. Place the domain-enhanced pre-trained language model in a closed loop that interacts with the knowledge graph in real time. Through parallel processing and information feedback mechanisms, the model can reason based on the most accurate and timely domain facts, thereby eliminating ambiguity in customer instructions.

[0020] 4. By dynamically adjusting the conditional probability table, the reasoning process of the credibility probability reasoning model can be adaptively adjusted according to the environmental context, which greatly improves the expressive power of the model and makes the risk assessment results more accurate and refined.

[0021] 5. The abstract execution strategy is accurately and quantitatively mathematically modeled in the form of a multi-dimensional feature vector, and by introducing the expected loss index, the credibility of the customer instruction and the path risk cost and execution cost are combined through probability theory to form a globally optimal decision function, so that the system realizes the transformation from the threshold-based conditional judgment to the quantitative decision based on random optimization, which can effectively control the customer instruction execution risk while improving the accuracy of customer instruction intention understanding. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments of the present application will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0023] Figure 1 is the implementation step diagram of the data processing method for customer relationship management provided by the embodiments of the present application. DETAILED DESCRIPTION

[0024] The technical solutions of the present application will be further described below by combining the drawings and through specific embodiments.

[0025] Among them, the drawings are only used for illustrative description, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation on the present application; in order to better illustrate the embodiments of the present application, some components of the drawings will be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it can be understood that some well-known structures and their descriptions in the drawings can be omitted.

[0026] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only used for illustrative description, and cannot be understood as a limitation on the present application, and for those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0027] In the description of the present application, unless otherwise explicitly specified and limited, if the term "connection" and the like appear to indicate the connection relationship between components, the term should be interpreted broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two components or the interaction relationship between two components. For those skilled in the art, the specific meaning of the above-mentioned terms in the present application can be understood according to the specific circumstances.

[0028] The data processing method for customer relationship management provided by the embodiments of the present application comprises the steps of: Figure 1 as shown, comprising the steps of: S1, generating an operation semantic tree based on multi-modal time alignment and domain knowledge injection; In the present embodiment, the purpose of generating the operation semantic tree is to achieve accurate intent understanding and unambiguous formalization representation of complex customer instructions. In order to convert the ambiguous natural language instructions into clear and unambiguous "operation semantic tree", the present application introduces a time alignment mechanism and a real-time domain knowledge injection mechanism, and the specific implementation method comprises the steps of: A1, time alignment and fusion of multi-modal input; The system receives the instruction stream input by the customer through different modalities such as voice, video, and text. Since the transmission and processing delays of different modalities are different, the present application creates a shared time alignment buffer, timestamps all input data with high precision, and dynamically aligns the features of other modalities (such as key frames in video, submission time in text) based on the text stream recognized by voice, to ensure that all context information is in the same time dimension when analyzing semantics.

[0029] The following examples will explain how the present application creates a shared time alignment buffer for multi-modal input: Suppose a customer issues an instruction through a video call: "(points to something outside the picture) I want to evaluate this item together with that box", where "this item" points to an object outside the camera's field of view, and "that box" points to a box on the screen.

[0030] In the above example, the modalities of the customer's input data include video data and voice data. The method of time alignment and fusion of the input data of the two modalities comprises the steps of: A11, time stamping and packaging multi-modal data, the specific method is: Deploy a high-precision, monotonically increasing hardware clock source on the system server side, and synchronize the time of all client acquisition devices (such as microphones and cameras) through the network time protocol.

[0031] Each acquisition device, upon capturing a raw data frame, immediately requests an absolute timestamp from the clock source. For example, upon capturing the client's input audio, several audio frame packets are generated from the input audio stream at a frequency of one packet every 20ms, and each audio frame packet is assigned a corresponding timestamp, denoted as . , Indicates the first The timestamp is added to each audio frame packet; at the instant the client's input video is captured, several video frame packets are generated from the input video stream at a frequency of one video frame packet every 40ms, and a corresponding timestamp is added to each video frame packet, denoted as . , Indicates the first Each video frame packet is timestamped. Additionally, to improve the reliability of the optimal security strategy for generating and executing client instructions, it is preferable to also timestamp other interactive events, such as when a client submits the text "Start Evaluation" on a tablet.

[0032] Then, all data packets (audio frames, video frames, events) are encapsulated together with their corresponding timestamps and sent to the server's timing alignment buffer.

[0033] A12 creates a dynamically expanding and shrinking circular buffer and dynamically sets the alignment time window. Then, it writes the time-stamped and encapsulated multimodal data packets into the circular buffer and automatically triggers an alignment check to filter out data within the circular buffer. Data packets of all modalities within the time range, Indicates the current time; After time-stamping and encapsulating the multimodal data in step A11, a dynamically growing circular buffer will be created in the server memory for this session. This buffer is not directly indexed by time, but managed by sequence number. Each slot in the buffer stores a data packet with a timestamp.

[0034] The method of creating a dynamically growing circular buffer for each session can be summarized as follows: create a circular buffer with a small initial capacity, then monitor its load status in real time, and automatically perform an expansion (or contraction) operation when a predefined trigger condition is met. During the expansion process, the integrity of existing data and the consistency of ongoing read and write operations are guaranteed.

[0035] Suppose the system is processing a customer's video command. Initially, the customer speaks calmly, but then becomes emotional and speaks faster, causing a surge in audio data. The method for dynamically constructing a circular buffer for this scenario is as follows: When a client session starts, the system allocates a ring buffer in memory with an initial capacity, e.g. 100 slots, each holding one packet. Both read and write pointers are initialized to the start position.

[0036] The buffer manager continuously calculates the current occupancy of the buffer, which is the ratio of the difference between the number of slots currently performing "write" operation and the number of slots currently performing "read" operation to the total number of slots.

[0037] Then, the current occupancy is compared to a dynamically set trigger threshold. When the current occupancy is greater than the trigger threshold, it indicates that the buffer is about to be saturated, and new data will overwrite old data if no intervention is made. In this case, the system triggers the expansion operation. Conversely, the system triggers the contraction operation. The trigger threshold in this application is dynamic and not a fixed value. The method of dynamically setting the trigger threshold is as follows: first, set a high water line threshold (e.g. 80%) and a low water line threshold (e.g. 30%), then the system dynamically adjusts the two water line thresholds according to the average size of the packet and the historical statistics of the processing delay. For example, if the average processing delay becomes longer, the expansion is triggered in advance (the high water line threshold is reduced to 70%) to leave enough space for longer processing time.

[0038] In this embodiment, the alignment time window is changed from a static configuration parameter to a variable dynamically calculated according to the real-time monitored network delay, processing delay and packet interval. The purpose is to enable the system to adaptively cope with changing network conditions and data processing load, rather than relying on a fixed value, to capture all associated modal data packets with the highest probability in the shortest possible time.

[0039] In step A12, the method of dynamically setting the alignment time window is as follows: The system continuously monitors and calculates the following indicators: Calculate the network jitter. For each modal data stream, the system calculates the standard deviation of its packet arrival time. For example, audio packets should theoretically arrive every 20ms, but due to network fluctuations, the arrival time interval may vary between 15ms and 25ms. The standard deviation of this range of variation is a measure of network jitter.

[0040] Evaluate the processing delay difference. The system evaluates the typical end-to-end processing delay (from acquisition to entry into the buffer) for different modal data streams. For example, through heartbeat packets or timestamp backtracking, the system knows that the processing delay of the video stream is 30ms slower than that of the audio stream on average.

[0041] Analyze the packet generation interval. The system identifies the native sampling period of different modalities, e.g. video frames are sampled every 40ms and audio frames are sampled every 20ms.

[0042] Then the alignment time window is dynamically calculated according to the following formula (1) : (1) In formula (1), represents the maximum network jitter value observed in all modalities, is used to amplify the influence of change; represents the absolute value of the processing delay difference between different modalities, is used to compensate for the deviation ; represents the maximum packet generation interval in all modalities, represents the packet generation interval of the type modality, is used to ensure that the window can cover at least one complete cycle; represents a safety constant used to deal with unmodeled sudden situations, so as not to . In order to prevent

[0043] the value from jumping too violently, preferably, the system does not recalculate for each packet, but uses a sliding window average or low-pass filtering method to smooth the calculated original value, and obtain a stable and practical value. Smoothing is not within the scope of protection claimed in this application, and therefore will not be specifically addressed.

[0044] In step A12, after the data packet is written into the ring buffer, the method of automatically triggering alignment checking is: When the data packet of any modality arrives, the buffer manager checks whether there is a data packet of other modalities generated within the time range ( representing the current time) in the ring buffer, if so, all modalities data packets within the time range are selected.

[0045] After the data packets of each modality are selected by step A12, go to step: A13, bind the data packets of each modality selected by step A12 as "alignment data group", and then release to the domain enhanced pre-training language model. For example, within the time window of the video frame , audio frames ,​​ wherein the first time instant and the second time instant are within a time window, the buffer binds the video frame, the audio frame, as an aligned data group. Then the aligned data group is released as a complete context unit to the subsequent domain enhanced pre-trained language model.

[0046] For example, when the system processes the customer saying "this one" (audio frame ), through step A1, the video frame with the timestamp aligned with the audio frame can be accurately obtained, although the customer's finger may point out of the picture in . This accurate alignment tells the model that "when the customer says 'this one', the visual focus is not in the current picture", thereby avoiding the model mistakenly identifying a certain item in the picture as "this one".

[0047] After obtaining the aligned data group as the input context unit of the domain enhanced pre-trained language model through step A1, the method of dynamically generating the operation semantic tree in step S1 proceeds to step: A2, while using the domain enhanced pre-trained semantic model to understand the intent of the aligned data group obtained through step A1, the identified preliminary entities are queried in real time with the background asset knowledge graph in parallel, and the query results are fed back to the domain enhanced pre-trained language model; Step A2 specifically includes the following steps: A21, the aligned data group obtained through step A1 is processed in parallel and asynchronously through a bypass and a main road, the bypass processing method is: The text data transcribed from the aligned data group is subjected to preliminary entity recognition through a pre-constructed dictionary, and the identified preliminary entities are formed into a preliminary entity list, then a batch query is initiated to the background asset knowledge graph, and the query results are structured into knowledge cues; The main road processing method is: determining whether to generate a knowledge cue for the text data, if yes, the knowledge cue is inserted into the specified position in the text data, and then proceed to step A22; if not, proceed to step A22; A22, input or feed back the same text data currently processed by the bypass or the text data processed by the bypass to the domain enhanced pre-trained language model, and the model predicts and outputs the structured understanding result.

[0048] ​​​​​For example, the aligned data set includes time-synchronized multi-modal data, such as a piece of audio transcription text and the corresponding video frame. Taking the text data of the audio transcription as an example, the method of bypass processing the text data is as follows: First, using the dictionary constructed by training, keywords or phrases pointing to domain entities (referred to as "preliminary entities") are extracted from the text data. For example, from the instruction "operate my sunflower painting", "sunflower painting" is identified as a preliminary entity.

[0049] Then, the system lists these preliminary entities and initiates a batch query to the background asset knowledge graph in parallel and asynchronously, and the query request is, for example: please return all asset entity IDs and their attributes related to "sunflower painting". For example, return the query record: 1, asset ID (such as A001), name "sunflower painting" original, status: in stock. Subsequently, the knowledge graph query result is structured and formatted into a piece of knowledge prompt, for example, formatted as: [knowledge back: the assets the user may refer to are: asset A001, sunflower painting original, status: in stock].

[0050] At the same time, the domain-enhanced pre-training language model on the main road processes the same text data, and the model is trained on professional corpus and understands the semantics of terms such as evaluation, pledge, and authorization. Since the training method of the domain-enhanced pre-training language model is not within the scope of the application for which protection is sought, it will not be described in detail. In short, model training is an optimization process of mapping relationship between preliminary entities and operation actions. The model outputs a context vector representation (i.e. a high-dimensional mathematical vector) of the text data as output, which contains the model's understanding of the text data based on the context. For example, the model input is "evaluate the value of the sunflower painting", and its output can be represented as "evaluate-sunflower painting".

[0051] In order to ensure that the model output is unambiguous, the knowledge prompt generated by the knowledge graph query is preferably inserted in front of the same text data being processed by the model to form an enhanced input sequence, and then the model is allowed to continue its subsequent reasoning and understanding based on this enhanced context. After receiving the knowledge back, the model continues to process the remaining instructions. At this time, the model's internal vector representation of "sunflower painting" has been anchored to the specific asset A001 from a vague concept. Finally, the model outputs a structured understanding result. At this time, the model output is no longer a vague "sunflower painting", but can directly output an unambiguous entity link pointing to a specific item in the knowledge graph, such as [sunflower painting, A001, evaluate].

[0052] After completing step A2, the system will perform the following steps: A3, dynamically generate operation semantic tree based on structured understanding result of model prediction output, specific method is: First, define domain syntax rules, such as defining a rule as: <operation> <action> <object> <parameter> <target>.

[0053] <operation> is the generated alignment data set; <action>: pre-defined action vocabulary, such as transfer, evaluation, authorization, pledge, etc. <object>: reference to specific nodes in asset knowledge graph, such as A001; <parameter>: key-value pair modifying action, such as ratio: 70%, period: 30 days; <target>: reference to specific nodes in customer knowledge graph or account knowledge graph, such as: pay to a specified account.

[0054] Then store the designed rules above in the syntax rule library.

[0055] Then, map the structured understanding result output by the language model to the above rules, and the mapping method is: Traverse the syntax rule library to find the production rule template that matches the action label of the input data, for example, for the labeled understanding result {action: authorization, object: A001, parameter: exhibition, cost ratio 70%, target: exhibition to achieve a certain target}, the found production rule template is expressed as: <operation> <action> <object> <parameter> <target>.

[0056] Subsequently, fill each data part in the understanding result into the corresponding slot in the matched production rule template, such as filling "authorization" into the <action> slot and "A001" into the <object> slot.

[0057] Finally, take "operation" as the root node of the semantic tree and the attributes of "operation" as the leaf nodes, and convert the production rule template into the operation semantic tree corresponding to the alignment data set. For example, the found production rule template is expressed as: <operation> <action> <object> <parameter> <target>, then create the <operation> slot as the root node of the semantic tree (such as "transfer"), then create the <action> leaf node and assign the attribute of the leaf node as "authorization", create the <object> leaf node and assign the attribute of the leaf node as "A001", and hang the key attributes of A001 (such as current evaluation) under the leaf node. Finally, a complete and hierarchical operation semantic tree corresponding to the alignment data set is generated.

[0058] After generating the operation semantic tree corresponding to the alignment data set through the multi-modal time alignment and domain knowledge injection of step S1, asFigure 1 As shown, the customer relationship management-oriented data processing method provided in the application comprises the following steps: S2, based on multi-source uncertainty reasoning, dynamically calculating the credibility probability of the customer instruction, and the calculation method comprises the following steps: B1, using the operation semantic tree, calculating the multi-modal behavior deviation score and the compliance conflict score; The calculation method of the multi-modal behavior deviation score is: First, the system continuously extracts low-dimensional and quantifiable behavior characteristics from each interaction with the same customer, rather than raw data. For example: the behavior characteristics of the text mode include: average sentence length, vocabulary complexity, specific domain term usage frequency, input speed; the behavior characteristics of the voice mode include: average speech speed, pitch range, fundamental frequency standard deviation, pause pattern; the behavior characteristics of the interaction mode include: login time point, commonly used device fingerprint characteristics, average interval time from login to operation initiation.

[0059] Then, a probability distribution model (such as a Gaussian distribution, recording the mean and standard deviation) is maintained for each behavior characteristic within a sliding time window (such as the last 90 days), which is the dynamic baseline for calculating the multi-modal behavior deviation score; Then, when the customer instruction comes, the current behavior characteristics of the customer are extracted in the above method, and then the standard score of the current behavior characteristics relative to the dynamic baseline is calculated, which is preferably the difference between the feature value of the current behavior characteristics and the mean value in the dynamic baseline, and the ratio of the difference to the standard deviation in the dynamic baseline.

[0060] Finally, the standard scores associated with different types of current behavior characteristics generated by the same customer at the current time are calculated to obtain the square root of the weighted sum, which is the multi-modal behavior deviation score of the customer at the current time.

[0061] The calculation method of the compliance conflict score is: First, match the nodes and edges in the operation semantic tree generated in step S1 with the nodes and edges in the dynamically updated compliance knowledge graph. The compliance knowledge graph includes: Entity nodes: such as customers, assets, accounts, countries; Rule nodes: such as constraint regulations, internal policies, laws and regulations; Relationship edges: such as restricted by, violated, allowed.

[0062] During the matching process, the system searches the compliance knowledge graph for paths that conflict with the operation semantic tree. For example, the operation semantic tree contains the operation "transfer" and the target is "account B". The compliance knowledge graph has an edge: account B→country of origin→country A, and another edge: country A→subject to→constraint regulation C. By traversing the compliance knowledge graph, the system finds a path from "account B" to "constraint regulation C" that conflicts with the execution of the customer instruction expressed by the operation semantic tree.

[0063] In this application, a conflict severity value is predefined for each conflict path, such as a conflict severity value of 0.6 for a path that violates a regulation, and a conflict severity value of 1 for a path that violates a law. After searching for each conflict path, the corresponding conflict severity value of each conflict path is queried, and the square root of the weighted sum of each conflict severity value is calculated as the customer's current compliance conflict score.

[0064] After step B1, the multi-modal behavior deviation score and the compliance conflict score of the customer instruction are obtained, the method of dynamically calculating the credibility probability of the customer instruction is transferred to step: B2, input the scores calculated in step B1 into the credibility probability inference model, and the credibility probability inference model dynamically adjusts the weight of the scores according to the environmental context, and finally predicts the credibility probability of the customer instruction expressed by the operation semantic tree.

[0065] Step B2 specifically includes the following steps: B21, define the risk nodes and node states of the credibility probability inference model, for example: Behavior abnormality node: represents the degree of deviation of the current operation behavior from the historical baseline; Instruction compliance node: represents the severity of the conflict between the current instruction and the compliance rules (represented by the conflict severity value); Identity authenticity node: represents the confidence of biometric authentication and other identity verification; Overall credibility node: the root node of the model, representing the overall credibility of the instruction.

[0066] The node state preferably includes "low", "medium" and "high" three states, respectively representing the risk level of the node.

[0067] Then, depending on the predefined risk nodes and node states, the network topology structure of the credibility probability inference model is constructed. For example, behavior abnormality and instruction compliance jointly affect overall credibility, so the network structure is designed as: behavior abnormality and instruction compliance as parent nodes in the model, pointing to the "overall credibility" sub-node.

[0068] B22, map the scores calculated in step B1 to the node states of the corresponding nodes in the network topology structure; For example, a threshold is set for the behavior abnormality node, such as a multimodal behavior deviation score less than a preset threshold, and the state of the behavior abnormality node is mapped to "low risk".

[0069] B23, dynamically adjusting the conditional probability table, the method comprising the steps of: B231, defining the environmental context; The system monitors a series of environmental context variables, such as monitoring whether the current time is in working hours or non-working hours, the asset sensitivity involved in the operation is "high" or "medium" or "low", the customer's own risk level, etc. B232, predefining a corresponding conditional probability table for each environmental context, and each set of conditional probability tables corresponds to a specific environmental context; For example, during working hours, the customer's behavior is more trusted, so even if the risk of behavior abnormality is "medium", as long as the risk of instruction compliance is "low", the overall trustworthiness probability is still high (such as 90%), and in this application, the environmental context "working hours" is defined as a separate conditional probability table as in the above example.

[0070] For example, at midnight, any abnormal behavior is more vigilant, and for the same behavior abnormality risk of "medium", the risk of instruction compliance is "low", the overall trustworthiness probability will be adjusted (such as adjusted to 60%), and the probability of overall trustworthiness "medium" will be increased accordingly. Therefore, the environmental context "midnight" is defined as a separate conditional probability table.

[0071] B233, according to the environmental context in which the customer instruction is currently located, matching the conditional probability table to be loaded from the pre-defined conditional probability tables; For example, the environmental context when the model predicts the trustworthiness probability is "midnight", the behavior abnormality risk obtained by analyzing the customer instruction is "medium", and the instruction compliance risk is "low", and the matching conditional probability table to be loaded is "the overall trustworthiness probability will be adjusted (such as adjusted to 60%), and the probability of overall trustworthiness "medium" will be increased accordingly" as in the above example.

[0072] After matching the conditional probability table to be loaded in step B23, the method for the model to predict the trustworthiness probability is transferred to step: B24, performing probability propagation on the conditional probability table to be loaded and the node state mapped by step B22, and calculating the trustworthiness probability of the overall trustworthiness node in the network topology structure constructed in step B21.

[0073] In the present application, the process of probability propagation is to calculate the probability of the leaf node (i.e. overall credibility) in each state by using the state of the parent node (such as the behavior abnormality node, the instruction compliance node) in the credibility probability reasoning model constructed according to the operation semantic tree generated according to the current customer instruction. For example, according to the behavior abnormality risk analyzed from the customer instruction, the risk is "medium", the instruction compliance risk is "low", and the environment context corresponding to the matched conditional probability table is "late night", and the overall credibility is, for example, a preset 60%, that is, the risk state of the leaf node (overall credibility) is, for example, "medium" (60% corresponds to the risk such as "medium").

[0074] Based on the conditional loading table and the node state mapped to the network topology structure in step B22, the method for calculating the credibility probability of the overall credibility node in the present application is briefly described as follows: For example, the risk state of the overall credibility node is respectively assigned to a corresponding numerical value representing its credibility, for example, "high risk = 1", "medium risk = 0.5", and "low risk = 0". The credibility probability = probability (high) x 1 + probability (medium) x 0.5 + probability (low) x 0. Probability (high), probability (medium), and probability (low) respectively represent the probability value of determining that the overall credibility node is a high risk, a medium risk, and a low risk after the probability propagation in step B24. It is emphasized here that there may be multiple execution paths from the initial node to the leaf node of "overall credibility node" in the credibility probability reasoning model, and therefore, there may be a case of simultaneously having probability (high), probability (medium), and probability (low).

[0075] After dynamically calculating the credibility probability of the customer instruction through step S2, as shown in Figure 1 The data processing method for customer relationship management provided by the present application enters step: S3, based on the pre-constructed credibility cost game model, dynamically generating the optimal security strategy for executing the customer instruction.

[0076] In the present embodiment, the credibility cost game model includes a multi-dimensional feature vector modeling model and a decision function, and the method for dynamically generating the optimal security strategy for executing the current customer instruction in step S3 includes the following steps: S31, multi-dimensional feature vector modeling is performed on the tentative execution path, and the modeling method is: Firstly, the key dimensions for evaluating a tentative execution path are determined, which include the cost and benefit of the path. In the present application, the dimensions for evaluating the path include: Time cost: the time required from initiating the instruction to completing the execution. For example, the automatic execution path takes 1 second, and the manual review path takes 1 hour.

[0077] Financial costs: The direct economic costs incurred in executing this path. For example, a manual review path requires the payment of agent manpower costs; an instruction simulation execution path requires additional computing resource costs, etc.

[0078] Risk cost: This represents the maximum potential loss incurred if the instruction is ultimately proven to be malicious or erroneous, and thus executed through this path. In this application, the risk cost is estimated based on the asset value in the operation semantic tree dynamically generated in step S1. For example, for a "transfer" operation, its risk cost might be the valuation of the asset associated with the "object" node in the operation semantic tree. In the automated execution path, this risk cost is high because a loss occurs once automated execution is performed. In the simulated execution path, the risk cost is close to zero because it is executed only in a simulated environment, and no real loss occurs.

[0079] Then, the values ​​of each feature dimension are normalized to a dimensionless uniform scale to eliminate the influence of different dimensions on subsequent analysis.

[0080] Finally, the normalized feature dimension values ​​are represented as vectors to form the multidimensional feature vector representation of the proposed execution path.

[0081] After expressing each proposed execution path in the credibility probability inference model as a multi-dimensional feature vector in step S31, proceed to the following step: S32, input the multi-dimensional feature vector and associated confidence probability of each proposed execution path into the decision function. The function solves the expected loss value of each proposed execution path and takes the proposed execution path with the minimum expected loss value as the optimal security strategy dynamically generated for the customer instructions obtained in step S1.

[0082] Specifically, first, we define the decision function. ,in, Indicates the first The probability of credibility of the proposed execution path. Indicates the first The risk and cost of the proposed implementation path Indicates the first The execution cost of the proposed execution path.

[0083] Then, for each path to be executed... , , Substitute it into the decision function and use the minimum The proposed execution path of the value is provided to the system as the final execution path.

[0084] To sum up, by setting the ring buffer, dynamically setting the alignment time window, adopting the back-annotation mechanism and other means, the present application expresses the customer instruction as a dynamic operation semantic tree, so that the domain enhanced pre-training language model can reason based on the most accurate and timely domain facts, thereby eliminating the ambiguity in the customer instruction. By dynamically adjusting the conditional probability table, the reasoning process of the credibility probability reasoning model can be adaptively adjusted according to the environmental context, greatly improving the expression ability of the model, and making the risk assessment result more accurate and fine. By taking the credibility probability and the multi-dimensional feature vector of the path to be executed as inputs of the decision function, the system realizes the transformation from the threshold-based conditional judgment to the quantitative decision based on the random optimization, and can effectively control the execution risk of the customer instruction.

[0085] It should be noted that the above specific embodiments are only the preferred embodiments of the present application and the technical principles applied. Those skilled in the art should understand that various modifications, equivalent replacements, changes, etc. can be made to the present application. However, as long as these changes do not deviate from the spirit of the present application, they should be within the protection scope of the present application. In addition, some terms used in the specification and claims of the present application are not limited, but only for the convenience of description.

Claims

1. A data processing method for customer relationship management, applied in a hybrid asset management scenario, characterized in that, Including the following steps: S1, based on multimodal temporal alignment and domain knowledge injection, dynamically generates an operational semantic tree; S2, based on multi-source uncertainty reasoning, dynamically calculates the credibility probability of customer instructions; S3, based on a pre-built trust cost game model, dynamically generates the optimal security strategy for executing the client's instructions.

2. The data processing method for customer relationship management according to claim 1, characterized in that, Step S1 specifically includes the following steps: A1 performs timing alignment and fusion of multimodal inputs; A2, while using the domain-enhanced pre-trained language model to perform intent understanding on the aligned data set obtained in step A1, simultaneously perform real-time queries on the identified preliminary entities and the background asset knowledge graph, and inject the query results back into the domain-enhanced pre-trained language model; A3 dynamically generates an operational semantic tree based on the structured understanding results of the model's predicted output.

3. The data processing method for customer relationship management according to claim 2, characterized in that, Step A1 specifically includes the following steps: A11 performs time-stamping and encapsulation on multimodal data; A12 creates a dynamically expanding and shrinking circular buffer and dynamically sets the alignment time window. Then, the time-stamped and encapsulated data packets are written to a circular buffer, and an alignment check is automatically triggered to filter out packets within the circular buffer. Data packets of all modalities within the time range, Indicates the current time; A13. Bind the data packets of each modality selected in step A12 into an "aligned data group" and then release them to the domain-enhanced pre-trained language model.

4. The data processing method for customer relationship management according to claim 2, characterized in that, Step A2 specifically includes the following steps: A21, the aligned data group obtained in step A1 is processed asynchronously in parallel via bypass and main paths. The bypass processing method is as follows: For the text data transcribed from the aligned data group, preliminary entity recognition is performed using a pre-built dictionary, and each identified preliminary entity is formed into a preliminary entity list. Then, a batch query is initiated to the asset knowledge graph in the background, and the query results are structured into knowledge prompts. The main road handling method is as follows: Determine whether the knowledge prompt has been generated for the text data. If so, insert the knowledge prompt at the specified position in the text data, and then proceed to step A22; If not, proceed to step A22; A22, the same text data currently being bypassed or the text data after bypassing is input or back-injected into the domain-enhanced pre-trained language model, and the model predicts and outputs the structured understanding result.

5. The data processing method for customer relationship management according to claim 2, characterized in that, Step A3 specifically includes the following steps: A31, based on the defined domain syntax rules, matches the structured understanding results of the model output with the corresponding production rule template, and fills the data in the understanding results into the corresponding slots in the template; A32, using the <operation> slot matched in step A31 as the root node of the semantic tree and the remaining slots as leaf nodes, the production rule template after filling the slot data is used to generate the operation semantic tree corresponding to the aligned data group.

6. The data processing method for customer relationship management according to claim 1, characterized in that, In step S2, the method for dynamically calculating the credibility probability of customer instructions includes the following steps: B1. Using the operational semantic tree, calculate the multimodal behavior deviation score and compliance conflict score; B2. The score calculated in step B1 is input into the credibility probability inference model. The credibility probability inference model dynamically adjusts the score weights according to the environmental context and finally predicts and outputs the credibility probability of the customer instruction expressed by the operation semantic tree.

7. The data processing method for customer relationship management according to claim 6, characterized in that, In step B1, the method for calculating the multimodal behavior deviation score includes the following steps: B11, the system continuously extracts behavioral features from each interaction with the same customer, and maintains a probability distribution model within a sliding time window for each behavioral feature, the probability distribution model serving as a dynamic baseline for calculating the deviation score; B12, extract the customer's current behavioral characteristics from the operational semantic tree, and calculate a standard score relative to the dynamic baseline; B13, calculate the square root of the weighted sum of squares of the standard scores associated with different types of current behavioral characteristics, and use it as the multimodal behavioral bias score.

8. The data processing method for customer relationship management according to claim 6, characterized in that, The method for calculating the compliance conflict score includes the following steps: C1, Match the operation semantic tree with the dynamically updated compliance knowledge graph by nodes and edges to search for whether there are any paths in the compliance knowledge graph that conflict with the operation semantic tree; C2, query the conflict severity value corresponding to each conflict path that conflicts with the operation semantic tree, calculate the square root of the weighted sum of squares of each conflict severity value, and use it as the customer's current compliance conflict score.

9. The data processing method for customer relationship management according to claim 6, characterized in that, The method for model prediction outputting the confidence probability includes the following steps: B21, relying on predefined risk nodes and node states, constructs the network topology of the credibility probability inference model; B22, map the score calculated in step B1 to the node state of the corresponding node in the network topology; B23, Dynamically adjusting the conditional probability table, the method includes the following steps: B231, defines the environment context; B232, predefines a corresponding conditional probability table for each of the aforementioned environmental contexts; B233, based on the current environment context of the client instruction, match the conditional probability table to be loaded from the predefined conditional probability tables; B24, perform probability propagation on the proposed conditional probability table and the node states mapped in step B22, and finally calculate and output the overall trustworthiness probability of the trustworthiness nodes in the network topology constructed in step B21.

10. The data processing method for customer relationship management according to any one of claims 1-9, characterized in that, Step S3 includes the following steps: S31, perform multi-dimensional feature vector modeling on the proposed execution path; S32, input the multi-dimensional feature vector of each proposed execution path and the associated confidence probability to the decision function, the function solves the expected loss value of each proposed execution path, and takes the proposed execution path with the minimum expected loss value as the optimal security strategy dynamically generated for the client instruction; The decision function Expressed as: ,in, Indicates the first The probability of credibility of the proposed execution path. Indicates the first The risk and cost of the proposed implementation path Indicates the first The execution cost of the proposed execution path.

Images