Financial institution customer service full-process digital system based on semantic understanding

Abstract

The invention relates to the technical field of financial intelligent customer service, and discloses a financial institution customer service full-process digital system based on semantic understanding. According to the system, an original request of a client is accessed through an interaction port, and a deep semantic kernel and interaction environment information are obtained through multi-level analysis. And the semantic mapping module associates the deep semantic kernel to the financial institution ontology concept network to generate a structured business intention framework. And the system synchronously calls a three-dimensional historical file of the customer, and in a context fusion layer, space-time alignment and logic stitching are carried out on the intention framework, the environment information and the historical file to generate an enhanced customer intention portrait with a multi-dimensional association tag. According to the method, deep semantic understanding and panoramic intention construction of complex and implicit appeals of customers are realized, so that core support is provided for accurate and personalized intelligent service response of the whole process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial intelligent customer service, in particular to a financial institution customer service whole-process digital system based on semantic understanding. BACKGROUND

[0002] The current intelligent customer service system of financial institutions generally adopts natural language processing technology combining intent classification and slot filling. This kind of technology classifies the natural language inquiry of the customer into a preset limited business category, and extracts the key words to fill in the fixed template. The historical information calling of the customer is mostly based on keyword query and list, and the combination of the current conversation context and the customer archives is relatively simple, which is often manifested as parallel display or shallow association of information.

[0003] The prior art scheme has defects in dealing with complex and professional financial business scenarios. The intent classification method is difficult to capture the deep business logic and multi-entity relationship implied in the user speech, resulting in that the understanding of the real demand of the customer stays on the surface and cannot be associated with the internal complex business rule network. In the information integration level, the system usually presents the customer basic data, product holding record, historical work order and other data in isolation, and lacks the mining of the internal logic and space-time association between different dimensions and different periods of information. This information stacking method cannot build a coherent, three-dimensional and causally reasoning customer intent view, resulting in mechanical service response, insufficient individualization and precision.

[0004] The present application needs to solve how to break through the surface semantics, deeply map the unstructured demands of the customer and the rigorous business concepts, rules and relationship network of the financial institution, so as to construct a computable business intent structure. At the same time, it needs to solve how to organically integrate the structured intent generated at this moment, the real-time interaction environment and the multi-dimensional historical archives of the customer, discover the time sequence association and logical connection therebetween, and dynamically synthesize a comprehensive, accurate and business insightful customer intent portrait. SUMMARY

[0005] The present application aims to provide a financial institution customer service whole-process digital system based on semantic understanding to solve the problems proposed in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a financial institution customer service whole-process digital system based on semantic understanding, which comprises:

[0007] An interaction port access module configured to receive the original interaction request of the customer from a plurality of customer service interaction ports, wherein the original interaction request carries unstructured service demand data and environmental background information;

[0008] a request analysis module arranged to perform multi-level analysis on the unstructured service appeal data in a primary processing layer to peel off a surface query sentence and a deep semantic kernel, and simultaneously peel off a customer state label and an interactive channel attribute from the environmental background information;

[0009] a semantic mapping module arranged to perform association matching of the deep semantic kernel with an ontology concept network of the financial institution in a semantic mapping layer, the ontology concept network being composed of business entities, business rules and business logic relationships, to generate a structured initial business intent framework;

[0010] a customer profile module configured to retrieve a stereoscopic historical profile associated with the customer according to a customer unique identification, the stereoscopic historical profile including a completed service event chain, an unresolved appeal record and a product holding relationship tree;

[0011] an intent portrait generation module arranged to perform spatio-temporal alignment and logic stitching of the structured initial business intent framework, the customer state label, the interactive channel attribute and the stereoscopic historical profile in a context fusion layer to generate an enhanced customer intent portrait with multi-dimensional associated labels.

[0012] Preferably, the multi-level analysis on the unstructured service appeal data in the primary processing layer includes:

[0013] receiving an audio stream from a voice interface or a text stream from a graphical interface;

[0014] performing speech-to-text operation on the audio stream to obtain an initial text, and performing encoding and unification processing on the text stream from the graphical interface;

[0015] calling a syntax analysis tool to perform dependency analysis on the unified text to identify core subject-predicate structures and modifying components, and separate out the surface query sentence directly expressed;

[0016] analyzing the implicit inclination, urgency and potential unclear needs in the text through an emotion dictionary and a context model, and condensing into the deep semantic kernel;

[0017] extracting device type, geographic location and access time from the environmental background information in parallel, combining into the interactive channel attribute, and extracting an identity token from a header of the original interactive request and analyzing the customer state label.

[0018] Preferably, the association matching of the deep semantic kernel with the ontology concept network of the financial institution in the semantic mapping layer includes:

[0019] splitting the deep semantic kernel into several discrete semantic segments;

[0020] calculating the similarity of each semantic segment with the business entity nodes in the ontology concept network, and activating the business entity nodes with similarity exceeding a threshold;

[0021] traversing the preset business logic relationship paths between the activated business entity nodes to form one or more potential business logic chains;

[0022] filling the necessary business parameter slots for each chain according to the standardized business procedures associated with the business logic chains, thereby assembling the structured initial business intent framework.

[0023] Preferably, the stereoscopic historical archives associated with the customer are retrieved according to the customer unique identifier, including:

[0024] sending a query instruction to the customer data center, the query instruction containing the customer unique identifier and time range parameters;

[0025] receiving the original historical data package from the customer data center, the original historical data package containing discrete interaction logs sorted by timestamp, account change events, and artificial service notes;

[0026] performing event extraction on the discrete interaction logs to identify the start, process node, and end marker of each independent service event, and linking them into the completed service event chain;

[0027] filtering out the open ticket records from the artificial service notes and organizing them into the unresolved appeal records;

[0028] summarizing the current and historical relationships between the customer and each financial product from the account change events to construct the product holding relationship tree with the customer as the root node and the product as the child node.

[0029] Preferably, in the context fusion layer, the structured initial business intent framework, the customer state label, the interaction channel attribute, and the stereoscopic historical archives are spatio-temporally aligned and logically stitched to generate an enhanced customer intent portrait with multi-dimensional association labels, including:

[0030] adding a timestamp to the structured initial business intent framework and comparing it with the time axis of the completed service event chain to find similar or related events occurring in adjacent time periods and obtain the comparison results;

[0031] analyzing the evolution pattern of the interaction channel attribute and the historical interaction channel to evaluate the stability or migratory nature of channel preference and obtain the evaluation results;

[0032] comparing the unresolved appeal record with the current initial business intent framework, identifying whether it is a continuation, derivative or repetition of a historical unresolved problem, and obtaining an identification result;

[0033] highlighting product nodes related to the current business intent in the product holding relationship tree, and loading detailed attribute rules thereof, and obtaining a marking result;

[0034] comprehensive analysis of the comparison result, the evaluation result, the identification result and the marking result, to generate an enhanced customer intent portrait containing intent categories, historical associations, channel backgrounds, product contexts and potential risk points.

[0035] Preferably, the system further comprises a process navigation engine, which receives the enhanced customer intent portrait and triggers dynamic path planning according to the enhanced customer intent portrait, including:

[0036] analyzing the intent categories and historical association strengths in the enhanced customer intent portrait, and screening all compliant basic service path templates from a pre-set process rule library;

[0037] According to the channel background and product context, adaptively filter the nodes in the basic service path template, remove the nodes that are not supported by the current channel or conflict with the current product rules;

[0038] According to the level of the potential risk points, insert mandatory verification nodes or risk confirmation nodes in the filtered path;

[0039] According to the efficiency preference indication in the customer state label, optimize the execution order of parallel nodes in the path, and finally generate a real-time dynamic service path containing specific operation nodes, decision branches and expected flow direction.

[0040] Preferably, the system further comprises a resource scheduler, which receives the real-time dynamic service path and allocates appropriate processing resources for each node in the path, including:

[0041] Identify the resource requirement type of each node in the real-time dynamic service path, including computing resources, knowledge resources, artificial agent expertise or approval authority level;

[0042] Real-time query of the idle state, load queue and estimated ready time of various resources;

[0043] According to the matching result of resource requirement type and resource state, and the emergency degree of the node, specify one or more candidate resource entities for each node;

[0044] A load balancing allocation strategy is made for nodes with multiple candidate resource entities, and the resource allocation result is attached to the real-time dynamic service path in the form of a resource binding table, forming a resource binding enhanced service path.

[0045] Preferably, the system further comprises an interactive execution and capture unit, which executes the resource binding enhanced service path and records the execution traces throughout, including:

[0046] According to the node order of the enhanced service path, the corresponding resource entities are sequentially driven to perform operations, including automatically calling application program interfaces, displaying information pages, initiating voice calls, or generating to-do tasks;

[0047] Before and after the execution of each node, the input data snapshot, output data snapshot, specific resource entity identifier used, start time, end time, and execution status code of the node are captured and recorded;

[0048] The feedback behavior of the customer during the execution of the node is monitored in real time, including confirmation, negation, follow-up questions, interruption, or channel switching, and the feedback behavior is stored in association with the corresponding node;

[0049] All node execution traces and customer feedback behaviors are woven into a complete process trace chain in chronological order.

[0050] Preferably, the system further comprises a path dynamic tuner, which makes real-time corrections to the executing enhanced service path according to the process trace chain, including:

[0051] The output data snapshot of the latest completed node in the process trace chain and the customer feedback behavior are compared with the expectations of the enhanced customer intent portrait, and the intent satisfaction deviation is calculated;

[0052] If the intent satisfaction deviation exceeds the allowed threshold, the process from semantic mapping to context fusion is retriggered based on the latest output data and customer feedback to generate a revised incremental intent portrait;

[0053] The revised incremental intent portrait is used to locally re-plan the sequence of nodes that have not been executed in the resource binding enhanced service path, including replacing, deleting, adding nodes, or adjusting the node order;

[0054] The re-planned node sequence and its resource allocation scheme are seamlessly connected to the current execution position to achieve dynamic switching of the service path.

[0055] Preferably, the system further comprises a session memory, which integrates the whole-process data of this time and structures the data after the single service ends, including:

[0056] receiving the whole-process data of this time, which includes the initial enhanced customer intention image, the resource-bound enhanced service path, the complete process traceability chain, and all path correction records;

[0057] extracting the core solution logic of this service and converting it into a reusable service mode script, which records the effective path, resource allocation and tuning record adopted under a specific intention image and specific situation;

[0058] associating and mounting the service mode script with the related business entity nodes in the ontology concept network;

[0059] updating the new customer feedback knowledge, resource performance data and path tuning experience generated in this service to the corresponding knowledge base, resource state library and process rule library, for optimizing the initialization process of subsequent similar services.

[0060] Compared with the prior art, the beneficial effects of the present application are:

[0061] associating and matching the deep semantic kernel of the customer query with the ontology concept network composed of business entities, rules and logical relationships to generate a structured initial business intention framework. This process realizes the deep decoding and formalized representation of the real business intention behind the customer's natural language expression. The generated intention framework not only clearly defines the core demand, but also associates related business entity constraints and rule premises, so that the system's understanding of the request goes from simple operation type recognition to the business logic level, providing a solid structured semantic foundation for subsequent accurate routing and processing.

[0062] spatial and temporal alignment and logical stitching of the structured initial business intention, real-time interaction environment tags and stereoscopic historical archives containing event chains, relationship trees, etc. to generate an enhanced customer intention image. The spatial and temporal alignment establishes a coherent view of the current request and historical behavior on the timeline. The logical stitching actively identifies and links the cause-and-effect, contradiction or progressive relationship between the current intention, customer state, historical events and asset relationships. This deep fusion mechanism converts the originally discrete data points into a self-consistent customer intention panorama with internal logic, enabling the system to respond based on comprehensive insights into the customer's long-term state, immediate situation and deep needs, improving the coherence, predictability and personalization level of the service. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1A working principle diagram of the financial institution customer service full-process digitalization system based on semantic understanding;

[0064] Figure 2 A flow chart for semantic mapping and initial business intent framework generation;

[0065] Figure 3 A flow chart for process navigation engine dynamic path planning;

[0066] Figure 4 A horizontal bar chart for parallel degree scoring;

[0067] Figure 5 A trend chart for monthly script sedimentation, reuse times and service time reduction rate. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0069] Please refer to Figure 1The application provides a financial institution customer service full-process digital system based on semantic understanding, which comprises: an interactive port access module as a unified entrance of the system, responsible for receiving original interactive requests of customers from voice, text, video and other interactive ports. The original interactive request received by the module carries unstructured service appeal data and environmental background information. A request analysis module constitutes a primary processing layer of the system, which is arranged to analyze the unstructured service appeal data in multiple levels, and the analysis process aims to strip the surface inquiry sentence directly expressed by the customer and the deep semantic kernel implied. The request analysis module simultaneously strips the customer state label representing the current state of the customer and the interactive channel attribute representing the source of the interaction from the environmental background information. A semantic mapping module constitutes a semantic mapping layer of the system, which is arranged to associate and match the deep semantic kernel output by the request analysis module with a predefined financial institution ontology concept network. The ontology concept network is composed of business entities, business rules and business logic relationships. Through the association and matching, the semantic mapping module generates a structured initial business intent framework. A customer profile module retrieves a stereoscopic historical profile associated with the customer from a background data center according to the customer unique identifier contained in or analyzed from the original interactive request. The stereoscopic historical profile contains multi-dimensional information such as the completed service event chain of the customer, the unresolved appeal record and the product holding relationship tree. An intent portrait generation module constitutes a context fusion layer of the system, which is arranged to fuse the structured initial business intent framework, the customer state label, the interactive channel attribute and the retrieved stereoscopic historical profile.

[0070] Embodiment 1: refer to Figure 2 The request analysis module analyzes the unstructured service appeal data in multiple levels in the primary processing layer. The interactive port access module receives an audio stream from a voice interface or a text stream from a graphical interface. For the audio stream, the request analysis module calls a speech recognition engine to perform speech-to-text operation to obtain an initial text; for the text stream from the graphical interface, character coding unification processing is performed. The request analysis module calls a syntax analysis tool to analyze the dependency relationship of the unified text, and identifies the core subject-predicate structure and the modification component in the text, so as to separate the surface inquiry sentence directly expressed by the customer. The request analysis module analyzes the implicit emotional tendency, problem urgency and potential unclear demand in the text through an integrated emotional dictionary and a pre-trained context model, and condenses these implicit information into a deep semantic kernel. The request analysis module extracts the device type, geographical location and access time from the environmental background information in parallel, and combines these information into the interactive channel attribute; at the same time, the identity token is extracted from the communication header of the original interactive request and the customer state label is analyzed.

[0071] The semantic mapping module associates and matches the deep semantic kernel with the ontology concept network of the financial institution in the semantic mapping layer. The semantic mapping module first splits the deep semantic kernel into several discrete semantic segments. The semantic mapping module performs vectorized similarity calculation on each semantic segment and the business entity nodes in the ontology concept network. The business entity nodes with similarity calculation results exceeding a preset threshold are activated. The semantic mapping module traverses the connection paths between the activated business entity nodes predefined by the business logic relationship to form one or more potential business logic chains. The semantic mapping module fills in the necessary business parameter slots for each chain according to the standardized business procedures associated with the business logic chain, thereby assembling a structured initial business intent framework.

[0072] In a specific implementation, the request parsing module performs multi-level parsing on unstructured service demand data in the primary processing layer. The work starts with receiving data streams from the interactive port access module. In a specific implementation, the interactive port access module receives an audio stream transmitted in pulse code modulation format from a voice interface or a text stream carried in the form of a hypertext transfer protocol message payload from a graphical interface. The request parsing module performs speech-to-text operation on the audio stream from the voice interface. This operation is implemented by calling a pre-trained end-to-end automatic speech recognition model to convert continuous audio signals into initial text composed of character sequences. For the text stream from the graphical interface, the request parsing module performs encoding unification processing to uniformly convert the existing multiple character encoding formats to the UTF-8 encoding format.

[0073] In some embodiments, the unified text is analyzed for dependency relationships by calling a syntax analysis tool. The syntax analysis tool can be a converter-based dependency syntax analyzer that receives the unified text as input and outputs a syntax dependency graph between words in a sentence. By analyzing the dependency graph, the core subject-predicate structure composed of a core predicate verb and its dominated object, as well as the modifying components such as attributive and adverbial attached to the core structure, are identified. Based on the separation of the core subject-predicate structure and the modifying components, the system extracts the surface inquiry sentence directly expressing the customer's question, for example, from "I want to inquire about the status of the transfer to account that I transferred through mobile banking yesterday", "inquire about the transfer to account status" is separated as the surface inquiry sentence.

[0074] In a specific implementation, the implicit sentiment, urgency, and potential unspoken needs in the text are analyzed through a sentiment dictionary and a context model. The sentiment dictionary contains a set of words with sentiment polarity and intensity annotations, and the context model is a language model trained on a large amount of customer service dialogue corpus. The system inputs the text into the context model to obtain the context vector representation of the words, and then matches and analyzes it with the sentiment dictionary to identify the implicit sentiment in the text, such as "anxious" and "unhappy", to judge the urgency of the problem, such as through the "as soon as possible" and "immediately" words or context, and to infer the potential needs of the customer that are not directly stated, such as the underlying semantic kernel behind "querying the transfer status" that implies "unhappiness with the transfer speed and a desire to urge or complain". These information is condensed into a structured deep semantic kernel representation.

[0075] It can be understood that the device type, geographic location, and access time are extracted from the environmental background information in parallel and combined into the interaction channel attributes. The environmental background information includes the user agent field in the network message header, the geographic location resolved from the global positioning system coordinates or the internet protocol address, and the server timestamp. From these information, the device type is "iOS mobile device", the geographic location is "Haidian District, Beijing", and the access time is "October 27, 2023, 14:30:00". These information is combined into the interaction channel attribute object. At the same time, the identity token is extracted from the header of the original interaction request. The token is an encrypted JSON web token. Parsing this token can obtain the customer's identity and status information, such as the customer level "gold card customer" and the account status "normal". These are parsed into specific customer status labels.

[0076] In a specific implementation, the semantic mapping module associates and matches the deep semantic kernel with the ontology concept network of the financial institution in the semantic mapping layer. The ontology concept network is composed of business entities, business rules, and business logic relationships, and is stored in the form of a graph. The business entity nodes include "personal customer", "transfer transaction", "account", and "arrival status". The business logic relationship edges include "initiate", "involve", "own", and "status as". The matching process first splits the deep semantic kernel into several discrete semantic segments, such as "customer-query", "transfer transaction-yesterday-mobile bank", "recipient account-someone", and "arrival status". Each semantic segment is represented as a vector in a high-dimensional space.

[0077] In some embodiments, the similarity between each semantic segment vector and the vector of all business entity nodes in the ontology concept network is calculated using the cosine similarity method, which is expressed as:

[0078]

[0079] where: Represents the similarity value. Represents a semantic fragment vector. Represents the vector of business entity nodes. This represents the magnitude of the vector. When the similarity value exceeds the preset threshold of 0.85, the corresponding business entity node is activated, such as the "transfer transaction", "account", and "acquisition status" nodes.

[0080] In practice, the semantic mapping module traverses the pre-defined business logic relationship paths between activated business entity nodes in the ontology concept network. It queries the graph database to find the path from the "Individual Customer" node to the "Transfer Transaction" node, then to the "Account" node, and finally to the "Receipt Status" node, thus forming a potential business logic chain. Based on the standardized business procedure "Personal Transfer Transaction Result Query" associated with this business logic chain, the necessary business parameter slots are filled into the chain, such as "Transaction Time: Yesterday", "Transaction Channel: Mobile Banking", and "Recipient Account Name: [Name]". After assembly, a structured initial business intent framework is generated. This framework includes the intent type "Query", the target business "Transfer Result", and a set of structured parameter key-value pairs.

[0081] Example 2: See Figure 3 The customer profile module retrieves a comprehensive historical record associated with each customer based on their unique identifier. It sends a query command to the financial institution's customer data center, containing the customer's unique identifier and a configurable time range parameter. The module receives raw historical data packets from the data center, including timestamp-sorted discrete interaction logs, account change events, and customer service notes. It extracts events from the discrete interaction logs, identifying the start marker, key process nodes, and end markers of each independent service event, linking these logs into a coherent chain of completed service events. From the customer service notes, it filters records with "unclosed" status, organizing them into unresolved request records. Finally, it summarizes the customer's current holdings and historical relationships with various financial products from account change events, constructing a product holding relationship tree with the customer as the root node and their held products as child nodes.

[0082] The intent portrait generation module performs spatio-temporal alignment and logical stitching of the structured initial business intent framework, customer state label, interaction channel attribute, and stereoscopic historical archive in the context fusion layer. The intent portrait generation module adds a current timestamp to the structured initial business intent framework and compares it with the timeline of the completed service event chain to find similar or related events occurring in adjacent time periods. The intent portrait generation module analyzes the evolution pattern presented by the current interaction channel attribute and historical interaction channel records to assess the stability or migration of customer channel preferences. The intent portrait generation module compares unresolved appeal records with the current initial business intent framework to identify whether the current intent is a continuation, derivative, or repetition of historical unresolved issues. The intent portrait generation module highlights product nodes related to the current business intent in the product holding relationship tree and loads detailed attributes and business rules of these products. The intent portrait generation module generates an enhanced customer intent portrait containing multiple-dimensional labels such as intent category, historical association strength, channel background, product context, and potential risk points by synthesizing the comparison results, evaluation results, identification results, and marking results.

[0083] In a specific implementation, the customer archive module retrieves the stereoscopic historical archive associated with the customer according to the customer unique identifier. The customer archive module sends a structured query language instruction to the customer data center of the financial institution, and the query instruction contains the parsed customer unique identifier, such as "CUST_ID_123456", and a preset time range parameter "past 365 days". After receiving the query instruction, the customer data center retrieves relevant data in the distributed database and returns an original historical data package. The original historical data package contains discrete interaction logs sorted in ascending order of timestamp, such as "[2023-10-26 10:00:00, client, operation: login], [2023-10-26 10:05:00, client, operation: query balance]", account change events, such as "[2023-10-25 15:30:00, event: transfer payment, amount: 5000], [2023-10-20 09:15:00, event: financial product purchase, amount: 20000]", and artificial service notes, such as "[service order ID: SR001, status: in processing, content: inquire about financial product redemption rules]".

[0084] In a specific implementation, the customer profile module performs event extraction on the discrete interaction logs in the original historical data package, the customer profile module applies a rule-based event boundary recognition algorithm, which merges log entries that are continuous and have a time interval less than a set threshold, such as five minutes, and have the same session identifier, and identifies the starting action of the merged log sequence as a starting flag, the core business operation as a process node, and the session end or timeout as an ending flag, thereby linking the discrete log entries into a completed service event chain, such as "Event Chain ID: E001, Start Time: 2023-10-26 10:00:00, Event Type: Login and Query, Process Nodes: Login System, Query Account A Balance, End Time: 2023-10-26 10:07:00", the customer profile module filters out work order records with a status field of "not closed", "processing" or "to be followed up" from the artificial service notes included in the original historical data package, and organizes these records into an unresolved appeal record list, such as "[Work Order ID: SR001, Creation Time: 2023-10-24, Topic: Financial Redemption Consultation, Latest Note: Rule has been informed, customer has not confirmed]".

[0085] It can be understood that the customer profile module induces the association between the customer and the financial product from the account change events included in the original historical data package, the customer profile module scans all account change events, extracts the product code field in the event, and queries the product master data dictionary to obtain the full name and type of the product, takes the customer as the root node, and takes each independent product instance held by the customer, such as "Current Account A", "Fixed Deposit B001", "Financial Product C_Share 001", as a child node, to build a product holding relationship tree, each product node in the tree is associated with its current state, such as "valid" and "expired", holding limit, and recent transaction time attribute, the customer profile module performs serialization operation on the tree structure to generate a structured document, such as JSON or XML format, containing node and edge relationships, to constitute a complete digital representation of the product holding relationship tree.

[0086] In some embodiments, the intent portrait generation module performs spatio-temporal alignment and logical stitching of the structured initial business intent framework, customer state label, interaction channel attribute, and stereoscopic historical archive in the context fusion layer. For the structured initial business intent framework, such as the intent type "query" and the target business "transfer result", the intent portrait generation module adds the current timestamp "2023-10-27 14:30:00" obtained from the interaction request. The intent portrait generation module compares this timestamp with the timeline of the completed service event chain, and finds a similar or related event that occurred within the adjacent time period, such as the past 24 hours. The comparison operation is completed by calculating the absolute difference between the current time and the historical event time, and comparing it with the preset time window threshold. The historical event record "event chain ID: E002, time: 2023-10-26 15:00:00, event type: transfer result query" is obtained as the comparison result.

[0087] In a specific implementation, the intent portrait generation module analyzes the interaction channel attribute "iOS mobile device client" and the historical interaction channel sequence recorded in the stereoscopic historical archive, which is "web page, iOS mobile device client, iOS mobile device client" in the past week. The intent portrait generation module calculates the frequency of the current channel appearing continuously in the historical sequence, evaluates the stability or migration of channel preference, and the evaluation result can be a quantitative stability score, which is expressed by the formula:

[0088]

[0089] wherein: represents the channel preference stability score, represents the channel of the i-th interaction in the historical sequence, represents the current interaction channel attribute, represents the length of the historical sequence, and the function returns 1 when the channels are the same, otherwise returns 0. The evaluation result is "channel preference stability", and the intent portrait generation module compares the unresolved appeal record with the current initial business intent framework. Through the natural language understanding model, the semantic relevance between the unresolved appeal record "financial redemption consultation" and the current intent "transfer result query" is calculated, and the recognition result is judged as "non-continuity problem".

[0090] In a specific implementation, when the natural language understanding model calculates the semantic relevance between the unresolved appeal record and the current initial business intent framework, first, the text description "financial redemption consultation" in the unresolved appeal record and the current intent "transfer result query" are preprocessed, including word segmentation, removal of stop words, and standardization of terms, to generate a unified text representation; then, the model calls a pre-trained context model trained based on a large amount of financial field dialogue corpus, which can map the text to a high-dimensional semantic vector space, thereby generating corresponding semantic vectors; the semantic relevance between the two semantic vectors is quantified by calculating the similarity of the two semantic vectors in the vector space, and the similarity value reflects the degree of association between the two intents in business logic; finally, based on the pre-set similarity threshold, the model automatically determines whether the semantic relevance meets the standard of continuity of the problem, and outputs the recognition result. The intent portrait generation module highlights the "account A" node involved in the current business intent framework and all account nodes related to the transfer function in the product holding relationship tree, and loads the detailed attribute rules of these nodes, such as "account A: single transfer limit 50000 yuan, daily cumulative limit 200000 yuan", and obtains the marking result.

[0091] It can be understood that the intent portrait generation module integrates the comparison result, i.e. "there is a similar query history within 24 hours", the evaluation result, i.e. "stable channel preference", the recognition result, i.e. "non-historical unresolved problem continuation", and the marking result, i.e. "associated account A and its limit rules", integrates these information as multi-dimensional association labels into a unified framework, and the intent portrait generation module generates an enhanced customer intent portrait data structure, which includes a core intent type field, a historical association strength field with a value of "high", a channel background field with a value of "stable mobile terminal", a product context field with a value of "involving account A and its limit rules", and a potential risk point field with a value of "repeated query in a short period of time, with anxiety about transaction security", and the enhanced customer intent portrait encapsulates the deep integration of the initial business logic and the customer context.

[0092] In an embodiment, the system further comprises a process navigation engine. The process navigation engine receives the enhanced customer intent profile. The process navigation engine parses the intent category and the historical correlation strength label in the enhanced customer intent profile, and screens all the basic service path templates that meet the regulatory and internal compliance requirements from a preset process rule library. The process navigation engine performs adaptive filtering on each operation node in the screened basic service path templates according to the channel background and the product context label in the enhanced customer intent profile, and removes the nodes that are not supported by the current interaction channel or conflict with the current involved product rules. The process navigation engine inserts mandatory identity verification nodes or risk prompt confirmation nodes in the filtered service path according to the potential risk points and their levels identified in the enhanced customer intent profile. The process navigation engine optimizes the execution order of the nodes that can be executed in parallel in the path according to the efficiency preference indication contained in the customer state label, and finally generates a real-time dynamic service path containing specific operation nodes, decision branch conditions and expected flow direction.

[0093] In a specific implementation, the process navigation engine receives the enhanced customer intent profile output by the intent profile generation module. The process navigation engine parses the intent category field and the historical correlation strength field in the enhanced customer intent profile. The value of the intent category field in the enhanced customer intent profile is “transfer result query”, and the value of the historical correlation strength field is “high”. The process navigation engine takes “transfer result query” and “high correlation strength” as a composite query condition to search in a preset process rule library. The process rule library is a graph database that stores various standardized service process templates. The process navigation engine screens all the basic service path templates related to the “transfer result query” intent and meeting the compliance requirements of the financial institution from the process rule library, such as “online self-service query process template”, “transfer to artificial auxiliary query process template” and “standard verification process template involving suspicious transactions”.

[0094] In some embodiments, the process navigation engine performs adaptive filtering on each operation node in the filtered base service path template according to the channel context field and the product context field in the enhanced customer intent profile, the channel context field value is "stable mobile terminal", and the product context field value contains "involving account A and its limit rules". The process navigation engine traverses the node sequence of each base service path template, checks the execution prerequisite condition of each node, and the execution prerequisite condition of the node "identity verification through SMS verification code" contains "channel supports SMS receiving and automatic filling". This condition is compatible with the channel background of "iOS mobile device client", so it is retained. The execution prerequisite condition of the node "direct to the counter" contains "channel attribute is offline network point". This condition is incompatible with the current channel background of "iOS mobile device client", so the process navigation engine removes this node from the current execution path. The execution prerequisite condition of the node "call for over-limit transfer approval record" contains "there is a history of over-limit transfer of the account involved". The process navigation engine queries the detailed attributes of account A in the product holding relationship tree and finds no over-limit transfer record. Therefore, the node is filtered and removed because it has no direct association with the current product context.

[0095] In a specific implementation, the process navigation engine inserts a mandatory verification node or a risk confirmation node in the filtered path according to the potential risk point field in the enhanced customer intent profile. The potential risk point field value is "repeated query in a short period of time, with anxiety about transaction security". The risk rule mapping table of the process navigation engine maps "repeated query" to "medium" level risk and configures a standardized risk confirmation node "voice verification or enhanced security question verification" for "medium" level risk. The process navigation engine inserts this risk confirmation node after the start of the filtered path sequence to ensure that additional customer identity and intent confirmation is completed before providing specific transfer results. The node content is set to "Dear customer, it is detected that you have recently frequently queried the transfer status. In order to protect your fund security, please confirm your identity again through voice or answer the preset security question".

[0096] It can be understood that the process navigation engine optimizes the execution order of parallel nodes in the path according to the efficiency preference indication implied in the customer state label in the enhanced customer intent profile. The customer state label contains an indication of "efficiency preference: high, inclined to self-service and parallel processing". The process navigation engine analyzes the remaining nodes in the current path, identifies that the three nodes "query core transaction system", "query asynchronous clearing system" and "generate query result summary" have no strict precedence relationship in logic and have parallel execution potential. The process navigation engine calculates the degree of improvement of overall efficiency by parallel execution and sorts the nodes. The sorting is based on a parallel degree score, which is expressed by the formula:

[0097]

[0098] wherein: represents the parallelism score of node i, used to determine its startup priority in parallel group, represents the direct relevance of node i to the core intent, defined by business rules, represents the dependency degree of node i to other nodes' output data, and are the relevance weight and dependency weight respectively, the process navigation engine arranges the parallel nodes in descending order according to the scores, and arranges high-score nodes to start execution first, after optimization, the process navigation engine integrates the node order, decision branch conditions, such as "if voice verification fails, jump to the human agent node", and the expected process flow direction, to generate a real-time dynamic service path, which is a structured instruction list sent to the downstream resource scheduling unit.

[0099] Referring to Figure 4 , this is a horizontal bar chart of parallelism score, used to show the priority ordering of different tasks. The objects shown are three tasks: generating query result summary, querying asynchronous clearing system, querying core transaction system; parallelism score is used to determine the startup priority of the task when executed in parallel; such charts are commonly used in process optimization and task scheduling scenarios (such as service processes in financial systems, task parallel processing in IT systems), which quantifies "parallelism" to reasonably arrange the execution order of tasks to improve overall efficiency.

[0100] In embodiment 4, the system further comprises a resource scheduler, which receives the real-time dynamic service path. The resource scheduler identifies the resource requirement type of each operation node in the real-time dynamic service path, which includes computing resources, knowledge base resources, human agents with specific expertise, or different levels of approval authority. The resource scheduler queries the current idle state, task load queue length, and estimated readiness time of each type of resource in the system in real time. The resource scheduler specifies one or more candidate resource entities for each node according to the matching results of the node's resource requirement type and the real-time state of the system resources, while considering the node's marked urgency. The resource scheduler formulates a load balancing allocation strategy for nodes with multiple candidate resource entities, and attaches the final determined resource allocation result in the form of a resource binding table to the real-time dynamic service path, forming an enhanced service path with resource binding.

[0101] ​The system further comprises an interactive execution and capture unit, which executes the resource-bound enhanced service path. The interactive execution and capture unit drives the corresponding resource entity to perform operations in sequence according to the node order of the enhanced service path, and the operation types include automatically calling internal application interfaces, showing information pages to customers, initiating voice calls to human agents, or generating background to-do tasks. The interactive execution and capture unit captures and records the input data snapshot, the output data snapshot, the specific resource entity identifier actually used, the start time, the end time, and the execution status code of each node before and after the execution of each node. The interactive execution and capture unit monitors the feedback behaviors of the customer in real time during the execution of each node, and the feedback behaviors include confirmation, negation, follow-up questions, interruption, or channel switching, and stores these feedback behaviors in association with the node that triggered the feedback. The interactive execution and capture unit weaves all the execution traces of the nodes and the customer feedback behaviors into a complete process trace chain in chronological order.

[0102] In a specific implementation, the resource scheduler receives the real-time dynamic service path generated by the flow navigation engine, parses the node sequence in the real-time dynamic service path, identifies the resource requirement type of each operation node, and the resource requirement type includes computing resources such as “high-performance transaction processing unit”, knowledge resources such as “cross-border transfer rule knowledge graph”, human agents with specific expertise such as “foreign exchange business agent group”, or different levels of approval authority such as “first-level authorization”. The resource scheduler sends a real-time query request to the resource state monitoring center of the financial institution, and the query request contains a list of the above-mentioned resource requirement types. The resource state monitoring center returns the current idle state, task load queue length, and estimated ready time of each type of resource, such as “artificial intelligence voice verification module: idle, load queue 0, estimated ready time 0 seconds”, “foreign exchange business agent group: busy, load queue 3, estimated ready time 120 seconds”.

[0103] In some embodiments, the resource scheduler specifies candidate resource entities for each node according to the matching results of the resource requirement type of the node and the real-time state of the system resources, and the emergency level marked by the node itself. The matching process involves calculating the matching score between the resource requirement and the resource state, and the matching score Φ is determined by the resource compliance coefficient C, the availability coefficient A, and the timeliness coefficient T, which is expressed by the formula:

[0104]

[0105] Wherein: represents the final matching score, respectively represent the weight coefficients of resource compliance, availability, and timeliness, represents the matching degree between the resource entity capability and the node requirement type, and the value is 0 to 1, representing the current idle probability of the resource entity, taking value from 0 to 1, representing the normalized ratio of the expected ready time of the resource entity relative to the tolerance time of the node, taking value from 0 to 1. The resource scheduler assigns a matching score resource entities exceeding the preset threshold are added to the candidate list, for the node with the label of high urgency, the resource scheduler preferentially selects the candidate resource entity with the minimum expected ready time, even if its load queue is slightly longer.

[0106] In a specific implementation, the resource scheduler formulates a load balancing allocation strategy for the node with multiple candidate resource entities, the load balancing allocation strategy is based on a round-robin algorithm or a weighted allocation algorithm based on the current load weight of the resource entity, the resource scheduler appends the final resource allocation result in the form of a resource binding table to the real-time dynamic service path, forming an enhanced service path with resource binding. The resource binding table is a one-to-one list corresponding to the path nodes, which clearly specifies which specific resource instance is responsible for the execution of each node. For example, for the "voice verification" node, it is bound to "artificial intelligence voice verification module_instance02"; for the "transfer details query" node, it is bound to "core transaction query service_cluster A". The enhanced service path with resource binding is sent to the interactive execution and capture unit.

[0107] It can be understood that the interactive execution and capture unit executes the enhanced service path with resource binding, and the interactive execution and capture unit drives the resource entity specified in the resource binding table to perform operations in sequence according to the node order of the enhanced service path, the operation type including automatically calling internal application program interfaces, such as calling "transaction query interface" and passing in parameters, displaying information pages, such as pushing "identity verification page" on the client side, initiating voice calls, such as calling "foreign exchange business seat number 1002", or generating background to-do tasks, such as creating "large transfer after-verification task" in the work order system. The driving process completes the transmission and triggering of instructions through the enterprise service bus or the remote procedure call framework, see Table 1.

[0108] Table 1: Node resource requirement and matching state table

[0109] Node ID Resource requirement type Matching result (candidate resource entity list) Designated resource entity ID Expected waiting time (seconds) N001 Voice verification module [AI voice module_02, AI voice module_05] AI voice module_02 0 N002 Transaction query interface [Query service_cluster A] Query service_cluster A 0 N003 Foreign exchange service representative [Representative group_B_employee ID 1002, Representative group_B_employee ID 1005] Representative group_B_employee ID 1002 45

[0110] In some embodiments, before and after the execution of each node, the interactive execution and capture unit captures and records the input data snapshot of the node, the output data snapshot of the node, the specific resource entity identifier used, the start time, the end time, and the execution status code. The input data snapshot records all the parameters passed in when the node is triggered, such as “{‘transfer serial number’:‘20231027123456’,‘verification method’:‘voice’}”. The output data snapshot records the result data returned after the node is executed, such as “{‘verification result’:‘success’,‘transaction status’:‘cleared’}”. The resource entity identifier comes from the resource binding table. The timestamp is accurate to the millisecond. The execution status code includes “success”, “failure”, “timeout”, and “interruption”. The interactive execution and capture unit monitors the feedback behavior of the customer in the process of node execution in real time. The feedback behavior includes clicking the “confirm” button on the page, saying “I disagree” in the voice call, inputting “why so long” in the message box, directly hanging up the phone, or switching from the mobile application to the phone channel. The monitoring is achieved by analyzing the client event log, real-time transcription of voice dialogue, or channel switching signaling. The captured feedback behavior is associated with the node that triggered the feedback and stored, such as associating the “customer clicks the negative button” event with the node “display transfer result page”.

[0111] In a specific implementation, the interactive execution and capture unit weaves the execution traces of all nodes and the customer feedback behavior into a complete process trace chain in chronological order. The process trace chain is a structured timeline data object. Each entry in the data object contains a timestamp, a node identifier, an action type, and an associated data index. The entries are arranged in ascending order of timestamp, forming a complete digital footprint from the beginning to the end of the service. The process trace chain is stored persistently in a distributed file system for subsequent path dynamic tuning and session memory usage.

[0112] Embodiment 5: The system further includes a path dynamic tuner. The path dynamic tuner modifies the enhanced service path being executed in real time according to the process trace chain. The path dynamic tuner compares the output data snapshot of the latest completed node in the process trace chain and the customer feedback behavior with the expected results of the enhanced customer intent profile used to generate the service path, and calculates the deviation value of the current intent satisfaction degree. If the calculated intent satisfaction degree deviation exceeds the threshold allowed by the system, the path dynamic tuner re-triggers the process from semantic mapping to context fusion based on the latest output data and customer feedback, and generates a revised incremental intent profile. The path dynamic tuner uses the revised incremental intent profile to locally re-plan the sequence of nodes that have not yet been executed in the enhanced service path of the resource binding, and the re-planning operation includes replacing nodes, deleting nodes, adding nodes, or adjusting the order between nodes.

[0113] The system further comprises a session memory which integrates the whole-process data of the current service after the single service is completed. The session memory receives the full amount of data of the current service, which includes the initial enhanced customer intent profile, the resource-bound enhanced service path, the complete process trace chain, and all path correction records. The session memory extracts the core solution logic of the current service and converts it into a structured reusable service mode script, which records the effective path, resource allocation scheme, and key tuning records adopted under a specific intent profile and specific context. The session memory associates and mounts the service mode script with the relevant business entity nodes in the ontology concept network. The session memory updates the new customer feedback knowledge, resource entity performance data, and path tuning experience generated in the current service into the corresponding knowledge base, resource state library, and process rule library, respectively.

[0114] In a specific implementation, the path dynamic tuner generates real-time corrections to the enhanced service path being executed according to the process trace chain generated by the interaction execution and capture unit. The path dynamic tuner subscribes to the data stream of the process trace chain. When a new node completion record is added to the process trace chain, the path dynamic tuner immediately reads the output data snapshot of the latest completion node and the associated customer feedback behavior. In a specific implementation, the latest completion node is the "show transfer result page", the output data snapshot is "{ 'transaction status': 'failed', 'failure reason': 'incorrect recipient account information'}", and the associated customer feedback behavior is that the customer clicked the "not recognized" button on the page and input the text "I confirm that the account number is correct". The path dynamic tuner compares the output data snapshot and the customer feedback text with the expectations of the enhanced customer intent profile on which the current enhanced service path is generated. The expectations of the enhanced customer intent profile are to successfully query the transfer status and provide clear results to the customer. The current actual output is a failed transaction and the customer's dissatisfaction with the failure reason. The path dynamic tuner calculates the intent satisfaction deviation, which involves comparing the expected output vector with the actual output vector and the feedback sentiment vector The formula is expressed as:

[0115]

[0116] Wherein: represents the intent satisfaction deviation value, whose value range is [0, 1], represents the expected output semantic vector extracted from the enhanced customer intent profile, represents the actual output semantic vector extracted from the output data snapshot, represents the sentiment and negation intensity vector extracted from the customer feedback behavior, representing the cosine similarity function, and respectively the weight coefficients of the output similarity and feedback intensity, representing the vector length, when the calculated intention satisfaction degree deviation value When the calculated intention satisfaction degree deviation value exceeds the system allowed threshold value 0.3, the path dynamic tuner determines that the subsequent part of the current service path cannot meet the customer's real intention.

[0117] In some embodiments, the path dynamic tuner re-triggers the processing procedure from semantic mapping to context fusion based on the latest output data "transaction failure" and the customer feedback "confirm the account number is correct". The path dynamic tuner combines the customer feedback "I confirm that the other party's account number is correct" with the original appeal as new unstructured service appeal data input, and requests the parsing module to perform multi-level parsing thereon to strip out a new deep semantic kernel "question and complaint on the reason for transaction failure being 'account error'". The semantic mapping module associates the deep semantic kernel with the ontology concept network and matches it to "transaction objection" and "error complaint" to generate a new structured initial business intention framework. The customer archive module retrieves the customer's stereoscopic historical archive, and the intention portrait generation module fuses the new framework with the archive to generate a revised incremental intention portrait. The incremental intention portrait contains a new intention category "transaction error complaint" and an enhanced potential risk point "customer dispute".

[0118] In a specific implementation, the path dynamic tuner uses the revised incremental intention portrait to locally re-plan the node sequence that has not yet been executed in the resource-bound enhanced service path. The path dynamic tuner reads the resource-bound enhanced service path, identifies the node position that has been executed, and the nodes that have not yet been executed in the original path include "end session" and "send satisfaction evaluation". The path dynamic tuner retrieves and inserts a new node sequence from the process rule library according to the "transaction error complaint" intention in the incremental intention portrait. The new sequence includes "initiate error complaint work order", "inform the customer of the complaint process and estimated time", and "transfer to dispute handling specialist". At the same time, the path dynamic tuner deletes the "end session" node that is no longer applicable in the original path. The path dynamic tuner adjusts the node order and places the "transfer to dispute handling specialist" node after the "inform the customer of the complaint process and estimated time" node. The path dynamic tuner calls the resource scheduler to re-allocate processing resources for the newly inserted nodes, such as allocating the "dispute handling specialist group" resource to the "transfer to dispute handling specialist" node.

[0119] It can be understood that the path dynamic tuner will re-plan the node sequence and its resource allocation scheme, and seamlessly dock to the current execution location. The path dynamic tuner sends a path switching instruction to the interactive execution and capture unit, which contains the new node sequence starting from the current node and the resource binding table. The interactive execution and capture unit suspends the advancement of the original path, loads the new node sequence and starts execution from the first node, i.e. “initiate error complaint work order”, realizes the dynamic switching of the service path, and the whole switching process is completed within milliseconds. The client's interactive interface smoothly transitions from the original result page to the complaint work order filling guide page.

[0120] After the single service is completed, the session memory integrates the whole process data and performs structured sedimentation. The session memory receives the full data of this service, including the initial enhanced customer intent image, the resource-bound enhanced service path, the complete process trace chain, and all path correction records. The path correction records include the intent satisfaction deviation value triggering correction, the regenerated incremental intent image, and the re-planned node sequence. The session memory extracts the core solution logic of this service, which is “when the transfer query result status is ‘failed’ and the reason is ‘account error’, but the customer feedback confirms that the account is correct, the transaction error complaint process should be triggered instead of ending the session”. The session memory converts this logic into a reusable service mode script. The service mode script is recorded in a structured script description language. The script defines the trigger condition, i.e. “the initial intent is query, the result is failure-account error, and the customer feedback is negative”, records the effective path used under this condition, i.e. “insert complaint work order creation and manual hotline transfer node”, the resource allocation scheme, i.e. “bind dispute handling special seat”, and the tuning record, i.e. “switch from the Nth node of the original path”.

[0121] In some embodiments, the session memory associates the service mode script with the relevant business entity nodes in the ontology concept network, and establishes a bidirectional link between the index identifier of the script and the "transfer transaction", "transaction status", "customer objection" business entity nodes in the ontology concept network. When subsequent semantic mapping is associated with these entities again, this service mode script can be quickly retrieved and recommended. The session memory updates the new customer feedback knowledge generated in this service to the corresponding knowledge base, for example, the "typical query of the customer to the failure code 'account information error'" is stored as a case knowledge. The session memory updates the resource performance data recorded in this service to the resource state library, for example, updates the average response time and success rate indicators of "artificial intelligence voice verification module_instance 02". The session memory updates the path tuning experience to the process rule library, for example, adds a decision branch rule pointing to the "transaction error complaint" sub-process in the "transfer result query" basic path template, and the branch condition is set as "the output result contains the failure code 'X' and the customer feedback emotion is negative", which is used to optimize the initialization processing of subsequent similar services.

[0122] Referring to Figure 5 This is a monthly script sedimentation, reuse times and service time reduction rate trend chart, which mainly shows the change relationship between "script sedimentation quantity", "script reuse times" and "average service time reduction rate". The increase of script sedimentation and reuse directly promotes the improvement of service efficiency. With more reusable service mode scripts accumulated and called more frequently, the standardization degree of customer service process is improved, and the service time is naturally reduced, which reflects the positive promotion effect of "session memory sedimentation" on service efficiency. Such charts are commonly used in performance analysis of enterprise digital systems (such as financial customer service systems) to evaluate the optimization effect of "knowledge sedimentation-reuse" mechanism on service efficiency, and help enterprises adjust process rule library and resource scheduling strategy.

[0123] It should be noted that, in this text, relational terms such as first and second are used merely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such process, method, article or apparatus.

[0124] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A semantic understanding-based financial institution customer service full-process digital system, characterized in that, The system comprises: An interactive port access module configured to receive an original interaction request of a customer from a plurality of customer service interactive ports, the original interaction request carrying unstructured service appeal data and environmental background information; A request analysis module arranged to perform multi-level analysis on the unstructured service appeal data in a primary processing layer to peel off a surface inquiry statement and a deep semantic kernel, and simultaneously peel off a customer state label and an interactive channel attribute from the environmental background information; A semantic mapping module arranged to associate and match the deep semantic kernel with an ontology concept network of a financial institution in a semantic mapping layer, the ontology concept network being composed of business entities, business rules and business logic relationships, to generate a structured initial business intent framework; A customer profile module configured to retrieve a stereoscopic historical profile associated with the customer according to a customer unique identifier, the stereoscopic historical profile including a completed service event chain, an unresolved appeal record and a product holding relationship tree; An intent portrait generation module arranged to perform spatio-temporal alignment and logical stitching on the structured initial business intent framework, the customer state label, the interactive channel attribute and the stereoscopic historical profile in a context fusion layer to generate an enhanced customer intent portrait with multi-dimensional associated labels.

2. The system for semantic understanding based financial institution customer service full-process digitalization according to claim 1, characterized in that, The multi-level analysis on the unstructured service appeal data in the primary processing layer comprises: Receiving an audio stream from a voice interface or a text stream from a graphical interface; Performing a speech-to-text operation on the audio stream to obtain an initial text, and performing encoding and unification processing on the text stream from the graphical interface; Calling a syntax analysis tool to perform dependency analysis on the unified text to identify core subject-predicate structures and modifying components, and separate the surface inquiry statement directly expressed; Analyzing the implicit inclination, urgency and potential unclear needs in the text through an emotional dictionary and a context model, and condensing into the deep semantic kernel; Parallelly extracting device type, geographic location and access time from the environmental background information, combining into the interactive channel attribute, and extracting an identity token from the header of the original interaction request and analyzing the customer state label. 3.The semantic understanding based financial institution customer service full-process digitalization system according to claim 1, characterized in that, The association and matching of the deep semantic kernel with the ontology concept network of the financial institution in the semantic mapping layer, the ontology concept network being composed of business entities, business rules and business logic relationships, to generate a structured initial business intent framework, comprises: Splitting the deep semantic kernel into a plurality of discrete semantic segments; Calculating the similarity of each semantic segment with a business entity node in the ontology concept network, and activating the business entity node with a similarity exceeding a threshold value; Traversing the preset business logic relationship paths between the activated business entity nodes to form one or more potential business logic chains; According to the standardized business procedures associated with the business logic chains, filling necessary business parameter slots for each chain to assemble into the structured initial business intent framework.

4. The system for semantic understanding based financial institution customer service full-process digitalization according to claim 1, characterized in that, The retrieval of the stereoscopic historical profile associated with the customer according to the customer unique identifier comprises: sending a query instruction to a customer data center, the query instruction containing a customer unique identifier and a time range parameter; receiving raw historical data packets from the customer data center, the raw historical data packets containing a timestamp-ordered discrete interaction log, account change events, and manual service notes; performing event extraction on the discrete interaction log to identify the start, process node, and end marker of each independent service event, and linking them into a completed service event chain; screening open ticket records from the manual service notes and organizing them into unresolved appeal records; summarizing the current and historical relationship between the customer and each financial product from the account change events to construct a product holding relationship tree with the customer as the root node and the product as the child node. 5.The semantic understanding based financial institution customer service full-process digitalization system according to claim 1, characterized in that, In the context fusion layer, the structured initial business intent framework, the customer state label, the interaction channel attribute, and the stereoscopic historical archive are spatiotemporally aligned and logically stitched to generate an enhanced customer intent portrait with multi-dimensional association labels, including: adding a timestamp to the structured initial business intent framework and comparing it with the timeline of the completed service event chain to find similar or related events occurring in adjacent time periods and obtain the comparison result; analyzing the interaction channel attribute and the evolution pattern of historical interaction channels to assess the stability or migration of channel preference and obtain the evaluation result; comparing the unresolved appeal records with the current initial business intent framework to identify whether they are a continuation, derivative, or repetition of historical unresolved problems and obtain the identification result; highlighting the product nodes related to the current business intent in the product holding relationship tree and loading their detailed attribute rules to obtain the marking result; integrating the comparison result, evaluation result, identification result, and marking result to generate an enhanced customer intent portrait containing intent categories, historical associations, channel backgrounds, product contexts, and potential risk points.

6. The system for semantic understanding based financial institution customer service full-process digitalization according to claim 5, characterized in that, The system further includes a process navigation engine that receives the enhanced customer intent portrait and triggers dynamic path planning based on the enhanced customer intent portrait, including: analyzing the intent categories and historical association strengths in the enhanced customer intent portrait to filter all compliant basic service path templates from a pre-set process rule library; adaptively filtering the nodes in the basic service path templates based on the channel background and product context to remove nodes that are not supported by the current channel or conflict with the current product rules; inserting mandatory verification nodes or risk confirmation nodes in the filtered path based on the level of potential risk points; optimizing the execution order of parallel nodes in the path based on the efficiency preference indication in the customer state label to finally generate a real-time dynamic service path containing specific operation nodes, decision branches, and expected flow direction.

7. The semantic understanding based financial institution customer service whole-process digitalization system according to claim 6, characterized in that, The system further includes a resource scheduler that receives the real-time dynamic service path and allocates appropriate processing resources for each node in the path, including: identifying resource requirement types of each node in the real-time dynamic service path, the resource requirement types including computing resource, knowledge resource, human agent expertise, or approval level; real-time querying of idle state, load queue, and estimated ready time of each type of resource; assigning one or more candidate resource entities to each node according to matching results of resource requirement types and resource states, and emergency level of the node; formulating load balancing allocation strategy for the node with multiple candidate resource entities, and appending resource allocation results in the form of resource binding table to the real-time dynamic service path, forming an enhanced service path with resource binding.

8. The system for semantic understanding based financial institution customer service full-process digitalization according to claim 7, characterized in that, The system further comprises an interactive execution and capture unit, which executes the enhanced service path with resource binding and records execution traces throughout the process, including: driving corresponding resource entities to perform operations in sequence according to the node order of the enhanced service path, the operations including automatic calling of application program interface, display of information page, initiation of voice call, or generation of to-do task; capturing and recording input data snapshot, output data snapshot, specific resource entity identifier used, start time, end time, and execution status code of each node before and after execution; real-time monitoring of feedback behaviors of the customer during node execution, the feedback behaviors including confirmation, negation, follow-up question, interruption, or channel switching, and storing the feedback behaviors in association with corresponding nodes; weaving execution traces of all nodes and customer feedback behaviors into a complete process trace chain in chronological order. 9.The semantic understanding based financial institution customer service full-process digitalization system according to claim 8, characterized in that, The system further comprises a path dynamic tuner, which makes real-time corrections to the executing enhanced service path according to the process trace chain, including: comparing output data snapshot of the latest completed node and customer feedback behaviors in the process trace chain with expectations of the enhanced customer intent profile, and calculating intent satisfaction deviation; if the intent satisfaction deviation exceeds the allowed threshold, re-triggering the process from semantic mapping to context fusion based on the latest output data and customer feedback, and generating a revised incremental intent profile; using the revised incremental intent profile to locally re-plan the sequence of nodes not yet executed in the enhanced service path with resource binding, including replacing, deleting, adding nodes, or adjusting node order; seamlessly connecting the re-planned node sequence and its resource allocation scheme to the current execution position, and realizing dynamic switching of the service path.

10. The semantic understanding based financial institution customer service whole-process digitalization system according to claim 9, characterized in that, The system further comprises a session memory, which integrates and structures the data of the whole process after the single service is completed, including: receiving the full data of the service, including the initial enhanced customer intent profile, the enhanced service path with resource binding, the complete process trace chain, and all path correction records; extracting the core solution logic of the service, and converting it into a reusable service mode script, which records the effective path, resource configuration, and tuning records adopted under a specific intent profile and specific situation; The service mode script is associated with a relevant business entity node in the ontology concept network; New customer feedback knowledge, resource performance data and path tuning experience generated in the service are updated to the corresponding knowledge base, resource state library and process rule library for optimizing the initialization processing of subsequent similar services.

Images