Compliance Management System and Method Based on Corpus Data of Financial Business Operations

Abstract

This invention relates to the field of financial information system technology, specifically to a compliance management system and method for corpus data based on financial business operations. The system includes a corpus sequence management module, a role behavior extraction module, a risk chain identification module, a semantic trend identification module, and a violation path identification module. In this invention, an interaction cycle structure is constructed by using the role, sequence, and time information of statements to clarify context and temporal relationships, thereby improving the completeness of semantic recognition. Subject-object structures and verb combinations are extracted around business objects to construct semantic action chains, identifying continuous changes in operational methods and role behaviors. By analyzing the evolution of tone, modality, and modifiers, changes in expression tendencies are judged, revealing potential violation intentions. Combining speech frequency, object focus, and content organization characteristics, repeatedly mentioned and guiding content is accurately screened, achieving risk identification under the semantic chain and improving the accuracy and coverage of review.

Description

Technical Field

[0001] This invention relates to the field of financial information system technology, and in particular to a compliance management system and method based on corpus data from financial business operations. Background Technology

[0002] The field of financial information systems technology involves the collection, transmission, processing, storage, and management of information in financial business scenarios. Its core aspects include compliance management of financial data, risk control, business process optimization, and information system security construction. This technology field is based on data-driven approaches, combined with industry rules and regulations, and achieves full-process support and supervision of financial business operations through the construction of an information platform. Among them, the traditional compliance management system based on corpus data in financial business refers to a system that identifies and processes unstructured corpus data, including customer communication content, business records, and internal communications, in the daily operations of financial institutions. The technical issue it addresses is how to extract information related to financial compliance from a large amount of corpus data involved in the financial service process and review and label it. The traditional method uses rule dictionary construction combined with keyword matching to screen and label corpus data. This method, based on regulatory requirements, uses manually maintained terminology databases or language fragment sets to perform semantic judgment on text to identify illegal expressions or sensitive content.

[0003] Traditional corpus-based compliance management methods, which rely on rule dictionaries and keyword matching, depend on manually maintained terminology databases for semantic judgment. Their processing mechanism is biased towards static and surface-level matching, lacking the ability to track the contextual relationships and dynamic evolution of dialogue. In practice, this method often only captures explicit violations, failing to effectively identify potentially risky expressions such as subtle guidance, shifting stances, and repetitive inducements. For example, in multi-round interactions between a customer and sales personnel regarding a financial product, even if the characters' wording doesn't explicitly violate the rule dictionary, but their tone gradually shifts or repeats to create a guiding trend, this method struggles to accurately identify them. Furthermore, this method processes individual sentences, failing to establish semantic extensions based on speaking rounds and behavioral chains. This results in violation clues being scattered across multiple dialogue rounds, making them difficult to connect and identify, thus impacting the overall review effectiveness. The heavy burden of rule maintenance and poor adaptability to new business expressions limit the scope and real-time performance of identification, posing significant time-related risks and potential omissions. Summary of the Invention

[0004] To address the technical problems existing in the prior art, embodiments of the present invention provide a compliance management system and method based on corpus data from financial business operations. The technical solution is as follows:

[0005] On the one hand, a compliance management system based on corpus data from financial business operations is provided. The system includes:

[0006] The corpus sequence management module extracts the statements spoken by customers and business personnel in financial customer service conversations, obtains role, sequence number and time information, identifies continuous interaction groups, constructs round structure, and generates a corpus round structure set.

[0007] The role behavior extraction module extracts verbs, subjects, and business objects from sentences based on the corpus round structure set, filters financial project object words such as financial management and loans, records the usage form, verb combination and modification structure, and generates a list of subject-object semantic actions.

[0008] The risk chain identification module analyzes the changes in the purpose and role of the actions based on the main and guest semantic action list, constructs action paths, identifies operation chains around the same object, filters out paths that continue the trend, and generates a set of suspected illegal semantic chain structures.

[0009] The semantic direction recognition module analyzes the suspected illegal semantic chain structure set, extracts modal words, tone and modifiers, compares the modal changes at the beginning and end of the link, judges the word trend changes, and generates a semantic change path marker set.

[0010] The violation path identification module determines whether there is repeated mention of a single business content based on the semantic change path tag set, verifies whether it constitutes a dominant tendency, and generates a set of content fragments with violation tendency by combining the role, object and behavior.

[0011] As a further aspect of the present invention, the corpus round structure set includes roles, sequence numbers, time information, and round numbers; the subject-object semantic action list includes verbs, subjects, business object items, usage forms, verb combinations, and modifying structures; the suspected violation semantic chain structure set includes action purpose, role orientation, action continuity path, operation chain, and path sequence; the semantic change path marker set includes modal words, tone structures, modifying items, modal change status, and direction adjustment features; and the violation-prone content fragment set includes role information, object items, repeated mention behavior, content organization pattern, and dominant tendency.

[0012] As a further aspect of the present invention, the corpus sequence management module includes:

[0013] The statement annotation submodule obtains the statements spoken by customers and business personnel in financial customer service conversations, extracts the speaking role, sequence number and start time corresponding to each statement, sorts them based on the sequence number and start time, constructs the statement annotation structure by combining the speaking role information, and generates statement number index values.

[0014] The statement aggregation submodule calls the statement number index value to identify statement groups that speak continuously from the same role, determines whether the speaking roles of adjacent statements are consistent and whether the time interval is lower than the conversation continuity threshold, filters statement groups that meet the conditions for aggregation, organizes and merges the statement group numbers and start time order, and generates a continuous statement sequence.

[0015] The round generation submodule identifies the switching points and time sequence of speaking roles based on the continuous sentence sequence quantity, determines the arrangement relationship between sentence groups, constructs a sorted list according to round logic, processes the round numbers uniformly, and generates a corpus round structure set.

[0016] As a further aspect of the present invention, the role behavior extraction module includes:

[0017] The semantic element recognition submodule obtains the sentence content annotated in the round structure of the corpus, identifies the verbs, subjects and object words of the characters speaking in the sentences, extracts the combination relationship between semantic elements, marks the correspondence between subjects and verbs according to the word order structure, and generates subject-verb pairing quantity.

[0018] The financial object screening submodule extracts the object word content based on the subject-verb pairing quantity, classifies and identifies all object words, determines whether they belong to financial products or loan service related items, filters semantic combinations that match the business type, and generates financial verb-object combinations.

[0019] The semantic structure summarization submodule calls the financial verb-object combination, extracts the modifiers and collocation structures in the verb combination, sorts out the collocation forms between the subject, verb, and object, summarizes the usage order and structural style of the modifiers, and generates a list of subject-object semantic actions.

[0020] As a further aspect of the present invention, the risk chain identification module includes:

[0021] The action purpose identification submodule obtains the verb and business object combination recorded in the subject-object semantic action list, identifies the action purpose of the same object item in the different statements, determines whether the same role uses different action intentions in the different statements, and generates the purpose change status.

[0022] The object operation chain construction submodule extracts all verb combinations corresponding to the associated object items according to the change state of the purpose, identifies the action sequence of the role around the same object, arranges the operation nodes according to the statement rounds, marks the correspondence between the role and the object in the path, and generates an object operation path set.

[0023] The semantic chain filtering submodule calls the object operation path set, detects the connection mode of continuous actions in the path in terms of semantic structure, filters path sequences with a continuous trend in verb combination, judges the continuity of actions between roles and the number of object co-occurrences in the path, and generates a set of suspected illegal semantic chain structures.

[0024] As a further aspect of the present invention, the semantic direction recognition module includes:

[0025] The segment extraction submodule calls the sentence segments in the suspected illegal semantic chain structure set, extracts the starting and ending segments of the chain, identifies modal words, mood structures and modifiers in the segment, extracts the grammatical position and sentence structure relationship of language components, and generates pragmatic structure extraction values.

[0026] The word usage variation recognition submodule extracts values ​​from the pragmatic structure, identifies the variation patterns of modal word categories and mood structures in the starting and ending segments, extracts the continuity features of modifier combination methods, determines whether word usage changes form a semantic shift in a consistent direction, and generates a word usage trend change rate.

[0027] The directional feature annotation submodule calls the word tendency change rate to extract the semantic direction information of continuously changing paragraphs in the semantic chain, identify the structural offset trend in the path nodes, annotate the semantic offset position and direction type, and generate a semantic change path tag set.

[0028] As a further aspect of the present invention, the illegal path identification module includes:

[0029] The role extraction submodule calls the sentence segments in the semantic change path marker set, extracts the role information and object item content in the segment, identifies the collocation relationship between the role and the object according to the sentence structure, obtains the distribution characteristics of the semantic function of the role in the syntactic position, and generates the corresponding value of the role object.

[0030] The tendency verification submodule identifies the repeated pointing of a character to an object item in the difference segment based on the corresponding value of the character object, detects the recurrence pattern of character association in the content structure, determines whether the content involved by the character forms a continuous organizational tendency, and generates a character dominance coefficient.

[0031] The content generation submodule calls the role dominance coefficient to filter sentence segments in the structural combination where roles and objects are frequently bound, extracts the behavioral terms and grammatical patterns of the corresponding roles, identifies stable arrangements of semantic direction and structural type, and generates a set of content fragments with violation tendencies.

[0032] As a further aspect of the present invention, the spoken statements refer to natural language text formed by customers or business personnel in financial service dialogues. The statements have clear syntactic structures and semantic subject-predicate relationships, and belong to unstructured language data.

[0033] The round structure refers to a two-dimensional structure constructed based on the chronological order of statements and the alternation order of speaking roles;

[0034] The business object refers to the specific products, services, or operational content in financial communication, including identifiable core business entities such as wealth management products, loan applications, and account management, which serve as the target of semantic behavior.

[0035] The terminology refers to the linguistic structure of a sentence, which consists of a subject, a verb, and modifiers.

[0036] As a further aspect of the present invention, the action path refers to a sequence of actions consisting of multiple semantic verbs used by a role around the same business object item, and the path is used to identify the continuity of operation logic under the same target;

[0037] The operation chain refers to a semantic behavior connection sequence constructed in the corpus by multiple roles or the same role based on a unified business object item;

[0038] The terminology trend refers to the direction of language use reflected by modal words or verb modifiers in continuous sentences, including the evolution from scheme-type words to commitment-type words;

[0039] The dominant tendency refers to the linguistic behavior characteristic in which a single role controls the direction of information transmission by continuously using leading verbs and repeatedly pointing to the same business object in a sentence sequence.

[0040] On the other hand, a compliance management method based on corpus data from financial business operations, which is executed based on the aforementioned compliance management system for corpus data from financial business operations, includes the following steps:

[0041] S1: Obtain the role tags, timestamps and sentence sequence numbers from the financial customer service conversation corpus, identify the sentence rounds according to the speaking time order and role alternation points, organize and number each group of consecutive sentences, and generate a corpus round structure set;

[0042] S2: Call the role statements in the corpus round structure set, extract the subject, verbs and business object items, filter the financial products and loan service items that appear in the object items, mark the verb collocation structure and modifying elements, and generate a list of subject-object semantic actions;

[0043] S3: Based on the verb and object combination in the subject-object semantic action list, identify the changes in the role and use of the same object item in different statements, construct an operation path based on the action continuation structure, and generate a set of suspected illegal semantic chain structures.

[0044] S4: Call the sentence fragments in the suspected illegal semantic chain structure set, extract modal words, tone structure and modifiers, compare the changes in word usage in the beginning and end segments of the link, determine whether the semantic content direction has been adjusted, and generate a semantic change path marker set;

[0045] S5: Based on the role and object information extracted from the semantic change path marker set, identify the repeated use of the same object in the scenario of financial products or loan services, call the role and verb usage in the subject-object semantic action list, analyze the sentence structure organization pattern and object word distribution frequency, calculate the role content coverage degree, and generate a set of content fragments with violation tendency.

[0046] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0047] In this invention, an interactive turn structure is constructed by using the roles, order, and time information of statements to clarify the context and temporal relationships, thereby improving the completeness of semantic recognition. Subject-object structures and verb combinations are extracted around business objects to construct semantic action chains, identify continuous changes in operation methods and role behaviors, and determine changes in expression tendencies by analyzing the evolution of tone, modality, and modifiers to reveal potential violations. By combining speaking frequency, object focus, and content organization characteristics, repeatedly mentioned and guiding content is accurately screened, thereby achieving risk identification under the semantic chain and improving the accuracy and coverage of review. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of a compliance management system based on corpus data from financial business operations provided in an embodiment of the present invention;

[0050] Figure 2 This is a system flowchart of the present invention;

[0051] Figure 3 This is a flowchart of the corpus sequence management module in this invention;

[0052] Figure 4 This is a flowchart of the role behavior extraction module in this invention;

[0053] Figure 5 This is a flowchart of the risk chain identification module in this invention;

[0054] Figure 6This is a flowchart of the semantic direction recognition module in this invention;

[0055] Figure 7 This is a flowchart of the illegal path identification module in this invention;

[0056] Figure 8 This is a flowchart of the method steps in this invention. Detailed Implementation

[0057] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0058] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0059] This invention provides a compliance management system based on corpus data from financial business operations, such as... Figure 1 The diagram shown illustrates a compliance management system based on corpus data from financial business operations. The system includes:

[0060] The corpus sequence management module acquires the spoken statements of customers and business personnel in financial customer service conversations, extracts the corresponding roles, sequence numbers and time information of the statements in the call, identifies continuous interactive statement groups, numbers them according to statement time and role switching order, constructs a round sorting structure, and generates a corpus round structure set.

[0061] Spoken statements refer to the natural language text formed by customers or business personnel in financial service dialogues. Spoken statements have clear syntactic structures and semantic subject-verb relationships. They belong to unstructured language data and are used for corpus recognition and semantic analysis tasks.

[0062] Turn ordering structure refers to a two-dimensional structure constructed based on the chronological order of statements and the alternation order of speaking roles. It is used to reconstruct the actual communication order of statements in a dialogue and is widely used in dialogue turn modeling and language sequence analysis.

[0063] The role behavior extraction module extracts verbs, subjects, and business object items from the sentences annotated in the round structure of the corpus. It filters object words that appear in financial projects such as wealth management products and loan services, records the word forms, verb combinations, and modifier structures used by the role for the object items, and generates a list of subject-object semantic actions.

[0064] Business object items refer to specific products, services, or operational content in financial communication, including wealth management products, loan applications, account management, and other identifiable core business entities, which serve as the receiving targets of semantic behavior;

[0065] The terminology refers to the linguistic structure of a sentence, consisting of a subject, verb, and modifiers. It is used to express a role's operation or intention towards a business object and has typical applications in language structure annotation and behavior determination.

[0066] The risk chain identification module identifies whether the purpose and role of the action in the statement change in different statements based on the verb and business object combination recorded in the subject-object semantic action list, constructs a continuous action path, identifies the operation chain between roles around the same object item, filters the path sequence with the action usage method has a continuous trend, and generates a set of suspected illegal semantic chain structures.

[0067] Action continuity path refers to the sequence of actions formed by multiple semantic verbs used by a role around the same business object. The path is used to identify the continuity of operational logic under the same goal and is used for behavior pattern tracking in semantic link analysis.

[0068] An operation chain refers to a semantic behavior connection sequence constructed in a corpus by multiple roles or the same role based on a unified business object item, used to reveal the behavioral interaction relationship of roles in multi-turn communication;

[0069] The semantic direction recognition module calls the sentence segments in the suspected illegal semantic chain structure set, extracts modal words, tone structure and modifiers in the language form, compares the modal changes between the beginning and end segments of the link, identifies whether there is a continuous change trend in the word trend, identifies the direction adjustment features, and generates a semantic change path marker set.

[0070] Lexical trends refer to the direction of language use reflected by modal words or verb modifiers in continuous sentences, including the evolution from scheme-type words to commitment-type words, and are often used to determine changes in language intent;

[0071] The violation path identification module extracts role information and object items from the semantic change path tag set to determine whether there is repeated mention of a single business content, identifies the content organization pattern of the role in multiple sentence segments, verifies whether it constitutes a dominant tendency, and generates a set of violation tendency content fragments by combining role characteristics, object characteristics and behavior patterns.

[0072] Dominant tendency refers to the linguistic behavioral characteristics in which a single role controls the direction of information transmission by continuously using leading verbs and repeatedly pointing to the same business object in a sequence of statements. These characteristics are used to identify the role's proactive guiding behavior.

[0073] Content organization patterns refer to the way characters construct sentences in multi-turn dialogues, the order of information presentation, structural integrity, keyword distribution, and language arrangement, which are used to analyze information delivery strategies and tendencies.

[0074] The corpus round structure set includes roles, sequence numbers, time information, and round numbers; the subject-object semantic action list includes verbs, subjects, business object items, usage forms, verb combinations, and modifying structures; the suspected violation semantic chain structure set includes action purpose, role orientation, action continuity path, operation chain, and path sequence; the semantic change path marker set includes modal words, tone structure, modal items, modal change status, and direction adjustment features; and the violation-prone content fragment set includes role information, object items, repeated mention behavior, content organization pattern, and dominant tendency.

[0075] Specifically, such as Figure 2 , 3 As shown, the corpus sequence management module includes:

[0076] The statement annotation submodule obtains the statements spoken by customers and business personnel in financial customer service conversations, extracts the speaking role, sequence number and start time corresponding to each statement, sorts them based on the sequence number and start time, constructs the statement annotation structure by combining the speaking role information, and generates statement number index values.

[0077] Extracting each spoken statement from the raw voice or text records parsed from financial customer service conversations involves reading the speech content, speaking time, and speaker information from the log file. Common conversation records fields such as `speaker_id` (speaker ID), `utterance_id` (statement ID), and `timestamp_start` (start time) are then used to arrange all statements in chronological order. If the log file lacks explicit numbering, a sequential numbering is constructed based on the ascending timestamp order. Speaker role identification can be aided by channel identifier fields, role label fields, or semantic keywords. Some systems provide channel number markers, such as channel 0 representing customers and channel 1 representing sales personnel. In the absence of channel information, keywords such as "please" can also be used. Phrases like "Ask," "I'll help you," and "Hello" are categorized as statements made by sales personnel. Alternatively, the role can be determined by the logical relationship between interrogative and answering sentences within a series of statements. For example, in five statements, the sequence is: "Hello," "I lost my card," "When did you find out?" "10 AM," and "Then I'll help you." After initial labeling, statements 1, 3, and 5 can be identified as being issued by sales personnel, while statements 2 and 4 are issued by customers. These statements are chronologically labeled as U1 to U5, with start times T ranging from T1 to T5. This results in a statement labeling structure such as (U1, sales personnel, T1), (U2, customer, T2), (U3, sales personnel, T3), (U4, customer, T4), and (U5, sales personnel, T5). This structure is then organized into a statement number index list and stored for use by subsequent modules.

[0078] The statement aggregation submodule calls the statement number index value to identify statement groups that speak consecutively from the same role, determines whether the speaking roles of adjacent statements are consistent and whether the time interval is lower than the conversation continuity threshold, filters statement groups that meet the conditions for aggregation, organizes and merges the statement group numbers and start time order, and generates a continuous statement sequence.

[0079] The system sequentially scans the statement list using statement number indexes, identifying whether the speaking roles are consecutive and determining if the time interval is less than a threshold. If the current statement and the next statement have the same role, the system calculates whether their time difference is less than a set session continuity threshold. For example, if the threshold is set to 20 seconds, and both U3 and U4 are spoken by business personnel, and T4 minus T3 equals 10 seconds, then U3 and U4 meet the aggregation condition and are grouped into the same statement group. If the two statements have different roles or the time difference is greater than the threshold, they are not aggregated. Aggregation begins with a new set of statements. After traversing all statements, all statement groups are generated. Each statement group records the role, starting statement number, and starting time. For example, statement group G1 includes U3 and U4, with a starting time of T3. If G2 is U5, with a starting time of T5, then the continuous statement sequence is structured data containing multiple elements. Each element records the aggregated statement number set, role, and time information, which are sorted by time and used as input for subsequent modules.

[0080] The round generation submodule identifies the switching points and time sequence of speaking roles based on the number of consecutive sentence sequences, determines the arrangement relationship between sentence groups, constructs a sorted list according to round logic, processes the round numbers uniformly, and generates a corpus round structure set.

[0081] Read all statement groups from the continuous statement sequence, sort them in ascending order by start time, and then traverse the statement groups. Determine if the speaking role of adjacent statement groups has changed. If the role changes from customer service to customer or vice versa, it is considered a speaking round switch. The two sets of statements before and after the switch point belong to different rounds. Assign a number to each round, such as Round 1, Round 2, etc. The round numbers should be consistent in chronological order. The numbering format can be uniformly set to Round_n, where n is the round number value, forming a round structure set. Each element in the set records the round number, the corresponding role, the list of statement numbers contained in that round, and the earliest start time. For example, Round_1 contains U1 and U3, the speaking role is customer service, and the start time is T1; Round_2 contains U2, the role is customer, and the time is T2; Round_3 contains U4 and U5, the role is customer, and the time is T4. All structure set contents can be stored in the form of field identifiers, supporting retrieval and calling.

[0082] Specifically, such as Figure 2 , 4 As shown, the character behavior extraction module includes:

[0083] The semantic element recognition sub-module obtains the statement content annotated in the corpus turn structure set, recognizes the verbs, subjects, and object words in the statements of the roles, extracts the combination relationships between the semantic elements, marks the corresponding situations of the subjects and verbs according to the word order structure, and generates the subject-verb pairing quantity.

[0084] Read the statement text and speaker role annotation information in the corpus turn structure set, process them in sequence according to the order of appearance of the statements in the structure, perform word segmentation operations on each statement to disassemble it into word units, and use a Chinese word segmentation tool to disassemble a statement such as "The customer hopes to apply for a credit loan" into "customer", "hopes", "apply for", "credit loan". Then, identify the pairing relationship between the subject and the verb through the word order structure, and judge that the subject is at the beginning of the sentence and is mostly a noun or a pronoun, and the verb is mostly a word expressing an action or intention. If there is a compound verb structure such as "hopes to apply for", it is regarded as a verb phrase as a whole, and the structure is analyzed in combination with the actual statement. For example, "customer", "hopes to apply for", "credit loan" are used as the subject, verb, and object respectively. In another example, the statement is "The user intends to apply for a financial product", and after disassembling, it can be obtained that "the user" is the subject, "intends to apply for" is the compound verb, and "financial product" is the object word. In the annotation structure, they are recorded as the subject S1, the verb V1, and the object O1 in sequence. Continue to process more sample statements such as "I need to open a savings account", "The account holder intends to transfer funds", etc., extract the corresponding subject-verb-object triples, and at the same time number them according to the order of appearance of the statements to generate a pairing quantity list. For example, in the first statement, the subject is numbered S1 and the verb is V, and in the second statement, the subject is S2 and the verb is V2. The number information is arranged according to the actual corpus order, and a pairing quantity containing the subject number and the verb number is generated to support the screening and induction of the subsequent modules.

[0085] The financial object screening sub-module extracts the object word content according to the subject-verb pairing quantity, classifies and identifies all the object words, judges whether they belong to the related projects of financial products or loan services, screens the semantic combinations that meet the business type, and generates financial verb-object combinations.

[0086] The system receives all object words extracted from subject-verb pairings and identifies and classifies these words item by item. First, a financial domain object dictionary is constructed, containing terms representing financial products and loan services, such as "credit card," "fund," "money market fund," "mortgage loan," and "fixed deposit." For each object word, it determines whether it belongs to a preset financial category through exact or fuzzy matching. If the object word is "mortgage application," it is determined to belong to the loan category through dictionary matching; if the object word is "Yu'ebao," it is determined to be a financial product. The selection criteria are based on whether the classification results belong to the two core business categories of "financial management" or "loan." The judgment logic is set so that if the object word's category tag is identified as financial management or loan, that word is retained. Semantic combination is used to select the best combination, otherwise it is discarded. This method can retain combinations such as "customer-application-credit card" and "user-purchase-fund", while filtering out combinations such as "customer-consultation-address change" where the object word "address change" does not belong to the target business scope. For ambiguous references in the object word classification, such as "this product" and "this service", it is necessary to search the original sentence context for semantic backtracking. For example, if there is "this product is a fixed-term wealth management product launched by the bank", then "this product" can be replaced with "fixed-term wealth management product" and classified as wealth management. During the classification process, all object words should retain their original text, corresponding subject and verb, and business category information. Finally, all financial verb-object combinations that meet the business scope are summarized and output.

[0087] The semantic structure summarization submodule calls the financial verb-object combination, extracts the modifiers and collocation structures in the verb combination, sorts out the collocation forms between the subject, verb, and object, summarizes the usage order and structural style of the modifiers, and generates a list of subject-object semantic actions.

[0088] Based on the three elements of subject, verb, and object in financial verb-object combinations, this study identifies the modifiers and collocations in each verb combination, categorizes the modifiers structurally, and identifies whether they appear before or after the verb, used to limit the attribute of the action, express the degree of intent, or indicate time characteristics. For example, in the sentence "Quickly apply for a credit card," "quickly" is an adverbial modifier, paired with the verb "apply," forming a fixed collocation structure. In another sentence, "Successfully open a wealth management account," "successfully" is a verb modifier, forming a verb phrase with the verb "open." During the processing, all verb modifiers can be grouped into one category, and their frequency can be counted. For example, when processing 1000 sample sentences, the verb "apply" is often preceded by the words "quickly" and "apply." The words "immediately" and "as soon as possible" were used, with "quickly" appearing 280 times (28%), marking it as a high-frequency modifier. Further analysis of subject-verb combinations was conducted, such as "users hope to activate" and "customers intend to purchase." "Hope" and "intend" were intentional modifiers, not core verbs but forming compound actions, categorized as subject + intentional verb + action verb + object. The object words were also analyzed for collocation; for example, "credit card" often combines with "apply" and "activate," and "financial products" often combines with "purchase" and "subscribe." This resulted in a subject-object semantic action list, including subject number, verb and its modifiers, object word, collocation order, and pattern. Each structure recorded its actual sentence source and collocation frequency, used for summarizing semantic templates and constructing knowledge extraction models.

[0089] Specifically, such as Figure 2 , 5 As shown, the risk chain identification module includes:

[0090] The action purpose identification submodule obtains the verb and business object combination recorded in the subject-object semantic action list, identifies the action purpose of the same object item in the different statements, determines whether the same role uses different action intentions in the different statements, and generates the purpose change status.

[0091] The system retrieves verb combinations and business object combinations from the subject-object semantic action list. Each combination is processed sequentially. By establishing a unified object identification rule, objects expressing the same meaning but with different wording are standardized. For example, "regular wealth management," "wealth management products," and "fund account" are uniformly identified as "wealth management." The system then clusters and statistically analyzes the actions associated with each object. For instance, actions associated with the "wealth management" object include "purchase," "view," and "redemption." Subsequently, each action group is categorized according to its semantic meaning. For example, "purchase" is categorized as "add operation," "view" as "information access," and "redemption" as "cancel operation." An intent category label sequence is constructed, and the verb category sequences used by the same role in different statements surrounding the same object are compared based on subject encoding. If inconsistencies in categories occur, it is considered that the object has different uses. When continuing processing, it is necessary to set the judgment criteria for intention category change. For example, if two actions belong to different dimensions in the category label, or if a character uses different category actions to point to the same object in a chronological order, it is marked as a usage change state. For example, if character A uses the intention of "apply" in statement 1 and the intention of "cancel" in statement 4, it constitutes a usage change. Referring to the conversion logic between different usage category labels, each category jump is regarded as a change event. If the same object has two or more category jumps, it is marked as a significant usage difference. The output data should include the character number, object name, verb sequence, usage category, and change label to form a usage change state set.

[0092] The object operation chain construction submodule extracts all verb combinations corresponding to the associated object items based on the change of state according to the purpose, identifies the action sequence of the role around the same object, arranges the operation nodes according to the statement rounds, marks the correspondence between the role and the object in the path, and generates an object operation path set;

[0093] Based on the combination of object items and verbs in the state of changing usage, the action content appearing in the corresponding sentences is extracted sequentially. The action combinations of the same object in different sentences are arranged based on time or word order numbering. Each action is transformed into an operation node, which records the verb itself, role number, and original sentence position code. An action path chain is constructed by linking all operation nodes involving the object. For example, the operation chain for the object "loan" includes operations such as "application," "approval," "contracting," "disbursement," and "query," linked sequentially as nodes 1 to 5 according to the order of sentence appearance. Roles between nodes are matched; if the role numbers of consecutive nodes are the same, the chain is a single-role path; if different role numbers appear alternately, a multi-role path is constructed. The role path chain requires marking the role corresponding to each node during processing. For example, if nodes 1 to 3 are operated by role A and node 4 is operated by role B, then the role allocation table should be marked. In addition, during the construction process, it is also necessary to handle the situation where multiple paths reference the same object. For example, if role A and role C both perform multi-step operations on "credit card", the system should split it into multiple paths. Each path should separately record the order of operation nodes, participating roles, and verb content. Furthermore, a round number field should be added to the path information to identify the temporal information of the statement in the text. All path chain structures are uniformly stored as an object operation path set, and the action nodes in each path are sorted according to the statement number, and the path number and object label are marked as the basis for the next stage of input filtering.

[0094] The semantic chain filtering submodule calls the object operation path set, detects the connection of continuous actions in the path in terms of semantic structure, filters path sequences with a continuous trend in verb combination, judges the continuity of actions between roles and the number of object co-occurrences in the path, and generates a set of suspected illegal semantic chain structures.

[0095] The system calls the object operation path set and has already constructed path information. It analyzes the semantic continuation features of consecutive actions in each path, and initially establishes a verb semantic continuation rule table. Verb combinations indicating logical order are categorized; for example, "application" and "approval" form a sequential pair, and "signing" and "loan disbursement" form an operation continuation pair. Verb pairs that meet the continuation rules are marked as "continuous" in the path, while verb combinations without logical order are marked as "interrupted." For example, "application" followed by "cancellation" does not constitute continuation. In the continuation judgment, each pair of adjacent verbs is checked. If there are three or more continuable combinations of consecutive verbs in a path, the path is judged to have a semantic continuation trend. The system continues to detect changes in role numbers within the path. If each operation node in the path consists of two or more role numbers, and the number of role switching exceeds 50% of the total number of nodes, it is considered a multi-role chain. In addition, the frequency of repeated occurrences of the same object in the path is statistically analyzed. For example, if the object "financial product" appears more than four times in the path, it is classified as a high-frequency co-occurrence object path. The path is filtered by combining three parameters: continuity trend, role switching frequency, and object repetition frequency. Path structures that simultaneously possess semantic coherence and operation concentration are retained. The path number, number of participating roles, object co-occurrence frequency, and length of the continuation paragraph are marked. The system then summarizes and outputs a set of suspected non-compliant semantic chain structures composed of all paths that meet the conditions.

[0096] Specifically, such as Figure 2 , 6 As shown, the semantic direction recognition module includes:

[0097] The segment extraction submodule calls the sentence segments in the suspected illegal semantic chain structure set, extracts the starting and ending segments of the chain, identifies modal words, mood structures and modifiers in the segment, extracts the grammatical position and sentence structure relationship of language components, and generates pragmatic structure extraction values;

[0098] After retrieving the sentence segments involved in the suspected non-compliant semantic chain structure, the starting and ending segments of the chain are extracted. Language elements are extracted from these two parts sequentially. First, modal words such as "should," "can," "must," and "may" are identified and classified using a semantic dictionary, categorized into "mandatory," "possible," "selective," and "suggestive" types. Next, the tone structure is identified and classified to determine whether it belongs to a declarative, imperative, interrogative, or rhetorical question. Modifiers are located and extracted, identifying adverbial modification structures and adjective combinations in phrases, such as expressions like "complete quickly," "may be delayed," and "seriously delayed." The grammatical roles of modifier phrases in the segment are analyzed, marking the positional relationships of subject, predicate, object, attributive, and adverbial. Simultaneously, the grammatical roles of modifier phrases are identified. By combining the subordinate clause structure and nested sentence patterns with the results of part-of-speech tagging and dependency parsing in natural language processing, the core semantic structure and modification relationships are extracted. For example, in the sentence "Loan applications should be processed immediately", "should" is identified as a mandatory modal word, "immediately" is an adverb modifier, "process" is a verb predicate, and "loan application" is the object. The sentence structure can be labeled as the order of "modal word + adverb + verb + noun phrase". For the sentence "Business matters may be approved at a later date", "maybe" is extracted as a possible modal word, "at a later date" is an adverb of time, "approval" is a verb predicate, and "business matters" is the object structure. The above content is converted into structured language component extraction data, named pragmatic structure extraction value, which includes dimensional fields such as part-of-speech tagging, sentence structure type, modifier combination, and grammatical position relationship.

[0099] The word usage variation recognition submodule extracts values ​​based on pragmatic structure, identifies the variation patterns of modal word categories and mood structures in the beginning and end segments, extracts the continuity features of modifier combination methods, determines whether word usage changes form a semantic shift in a consistent direction, and generates word usage trend change rate.

[0100] Based on the pragmatic structure extraction values ​​generated in the previous stage, a differential analysis was conducted on the modal words, modifiers, and mood structures of the initial and final paragraphs. First, modal words were categorized; for example, if the initial paragraph contained "must" and the final paragraph contained "may," they were marked as "mandatory" and "selective," respectively, belonging to the semantic shift type "mandatory to selective." Next, it was identified whether the mood structure had changed, such as from a declarative sentence to a selective interrogative sentence, which constitutes a weakening of the mood pattern. Then, a structural analysis was performed on the modifiers, extracting the modifying phrase combinations in the initial paragraph and comparing them with the modifying expressions in the final paragraph. The degree of consistency of the terms and the differences in the arrangement of the modifier structures were statistically analyzed. For example, "complete quickly" changed to "consider appropriately," indicating that the modifier shifted from a time-sensitive orientation to a vague suggestion orientation, thus determining the change in the modifier. When considering continuity features, by comparing the modification combinations between adjacent segments, if most segments show a trend of weakening modification or strengthening judgment in the same direction, they are marked as semantic shifts with "consistent direction". These shift phenomena are further classified and statistically analyzed. If more than 60% of the segments in the total segments show modification or modal word changes in the same direction, it is defined as a high degree of word usage trend change rate. For example, if 10 groups of segments are processed, and 7 of them show a trend of change from affirmative and mandatory to suggestive and ambiguous, then the trend change rate of this group is 70%, which can be classified as a significant shift group. The modal word category change, tone structure change category, continuous change pattern of modification items, and statistically consistent trend ratio in each pair of segments are combined together to output a word usage trend change rate dataset, which is used to mark the continuity of semantic shift trends.

[0101] The directional feature annotation submodule calls the word trend change rate to extract the semantic direction information of continuously changing paragraphs in the semantic chain, identify the structural offset trend in the path nodes, annotate the semantic offset position and direction type, and generate a semantic change path tag set.

[0102] After obtaining the rate of change in word usage trends, the paths of continuously changing segments contained in the semantic chain are identified. First, the direction of change of modal words, modifiers, and tone structures in each segment of each path is summarized and categorized into semantic offset direction labels such as "enhancing," "weakening," and "shifting." For example, if multiple segments in a path gradually evolve from "must be handled" to "can be considered," then the overall direction of the path is defined as "forced selection." When marking the offset position, the segment number, the specific location of the word change, and the offset direction type of each offset node must be recorded. For example, node 1 is marked as "modal word offset - forced to suggestion," and node 2 is marked as "modifier offset - reduced timeliness." By traversing the offset types of each node in the path, it is identified whether a continuous chain is formed. Continuing the offset trend, if three or more consecutive nodes in a path exhibit the same offset direction, such as all being of the "semantic weakening" type, they are marked as a continuous directional offset chain. At the same time, consistency calculation is performed on all offset nodes in the path, and the proportion of offset nodes with the same direction in the total offset nodes is counted. For example, if there are 6 total offset nodes in the path, and 5 of them belong to the "weakening" type, then the offset direction consistency is 83%. If it is higher than 70%, it is considered a directional chain with strong consistency. Each path that meets the consistency threshold is numbered and labeled with the structural offset trend type. The list of structural offset nodes, direction labels, and consistency values ​​of each path are combined into structural data, and the output forms a semantic change path label set, which is used as input for compliance detection or language behavior analysis models in subsequent stages.

[0103] Specifically, such as Figure 2 , 7 As shown, the illegal path identification module includes:

[0104] The role extraction submodule calls the sentence segments in the semantic change path tag set, extracts the role information and object item content in the segment, identifies the collocation relationship between the role and the object according to the sentence structure, obtains the distribution characteristics of the semantic function of the role in the syntactic position, and generates the corresponding value of the role object.

[0105] After receiving sentence segments from the semantic change path tag set, the roles and objects of each segment are extracted sequentially. First, a natural language processing model identifies words with action functions as candidate roles, such as "administrator," "auditor," and "system." Simultaneously, the objects acted upon by the verbs are identified, such as "application," "record," "information," and "data." Dependency parsing is used to obtain the relationship between the verb and the subject and object, forming a preliminary match between roles and objects. For example, if the segment is "administrator deletes user information," then "administrator" is the role, and "user information" is the object. Next, the collocation structure between roles and objects is analyzed, recording their positional distribution in the grammatical structure, and determining whether they conform to subject-verb-object, subject-verb structure, or other similar structures. The subject structure is omitted. For example, the phrase "submit report" omits the role, but the role can be identified as "review system" through context tracking. Then, the position of the role in multiple segments is statistically analyzed to identify its typical distribution pattern in the sentence structure. For example, in "user operation interface submit feedback", "user" is the subject role at the beginning of the sentence. The frequency of occurrence of the combination of role and object is extracted by combining multiple sentence segment samples. For example, if "system" as a role and "data" as an object item appear 62 times in 100 sentence segments, the corresponding value of role and object is marked as 62%, and its syntactic structure is recorded as "subject + verb + noun". Through this process, a correspondence matrix of role and object items in each segment is constructed, covering semantic function, lexical collocation and sentence structure features.

[0106] The tendency verification submodule identifies repeated pointing of a character to an object item in the difference segment based on the corresponding value of the character object, detects the recurrence pattern of the character association in the content structure, determines whether the content involved by the character forms a continuous organizational tendency, and generates a character dominance coefficient.

[0107] Based on the corresponding values ​​of the roles and objects, we analyze whether the pointing of each role in different paragraphs is repetitive. First, we group each role and count the number and frequency of occurrence of the objects it is bound to in different paragraphs. For example, if the role "reviewer" points to "violation content," "text fragments," and "data fields" in multiple paragraphs, with an occurrence rate exceeding 60%, it is initially identified as a repetitive pointing. Then, we compare the content structure of these paragraphs to determine whether there is a recurring pattern in the expression. For example, if the same combination of sentence structure for the same role and object appears repeatedly, such as "reviewers identify sensitive content" and "reviewers record sensitive content," although the verbs are different, the main structure is the same, constituting expression repetition. Subsequently, we determine whether it constitutes a persistent organizational tendency. By analyzing the consistency and temporal continuity of semantic content, we can identify whether a role has long dominated the semantic change path of a specific object. If a role appears as the subject or agent in most segments and the matched objects are concentrated in a few types of content, it can be judged as a continuous organizational tendency. We further generate a role dominance coefficient, which represents the degree of control of the role in the distribution of objects. For example, if "review module" and "content paragraph" appear 140 times in 200 segments, the role dominance coefficient is 70%. If this value is higher than the set threshold of 60%, it is marked as a "high dominance role" as a key reference for subsequent content screening. We output the dominance ratio corresponding to each role and store it in the role tendency statistics table for later use.

[0108] The content generation submodule calls the role dominance coefficient, filters the sentence segments in the structural combination where roles and object items are frequently bound, extracts the behavioral terms and grammatical patterns of the corresponding roles, identifies the stable arrangement of semantic direction and structural type, and generates a set of content fragments with violation tendency.

[0109] After receiving the role dominance coefficient, all sentence segments are screened to extract expressions with high frequency of role-object binding and strong structural stability. First, a set of segments is selected from roles with a dominance coefficient higher than 60%. Then, the action verbs used by that role in all segments are analyzed, such as "review," "process," "record," and "mark," and matched with their corresponding object items such as "text content," "user behavior," and "risk data." Common sentence structures for role-object combinations are statistically analyzed, such as "role + verb + object" or "system module + behavior + content unit." By clustering the repetition of structures and verb categories in multiple segments, frequently occurring structures are identified. Combinations with consistent semantics are marked as semantically stable combinations. For example, if the action verbs in most sentences point to risk suppression operations, they are classified as "restricted behavior structures". If the structural forms are concentrated in the two categories of verb-object structures and compound predicate structures, they are marked as "stable structural arrangement". High-frequency roles, high-frequency objects, action verbs, grammatical templates, semantic directions, etc. are further combined into a set of content fragments. For example, if a role repeatedly appears phrases such as "identify risky content", "report violation fragments", and "submit monitoring records" in multiple sentences, it can be summarized as violation-prone content fragments. This information is organized into fragment data and used for the construction of compliance detection tasks and the input content of semantic trigger behavior analysis.

[0110] Please see Figure 8 The compliance management method for corpus data based on financial business operations is executed based on the aforementioned compliance management system for corpus data based on financial business operations, and includes the following steps:

[0111] S1: Obtain the role tags, timestamps and sentence sequence numbers from the financial customer service conversation corpus, identify the sentence rounds according to the speaking time order and role alternation points, organize and number each group of consecutive sentences, and generate a corpus round structure set;

[0112] S2: Call the role statements in the corpus round structure set, extract the subject, verbs and business object items, filter the financial products and loan service items that appear in the object items, mark the verb collocation structure and modifiers, and generate a list of subject-object semantic actions;

[0113] S3: Based on the verb and object combination in the subject-object semantic action list, identify the changes in the role and use of the same object item in different sentences, construct an operation path based on the action continuation structure, and generate a set of suspected illegal semantic chain structures.

[0114] S4: Call the sentence fragments in the suspected illegal semantic chain structure set, extract modal words, mood structure and modifiers, compare the changes in word usage in the beginning and end of the link, determine whether the direction of semantic content has been adjusted, and generate a semantic change path tag set;

[0115] S5: Based on the role and object information extracted from the semantic change path tag set, identify the repeated use of the same object in the scenario of financial products or loan services, call the role and verb usage in the subject-object semantic action list, analyze the sentence structure organization pattern and object word distribution frequency, calculate the role content coverage, and generate a set of content fragments with violation tendency.

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A compliance management system for corpus data based on financial business operations, characterized by, The system comprises: The corpus sequence management module extracts the speech statements of the customer and the service personnel in the financial customer service conversation, obtains the roles, sequence numbers and time information, identifies continuous interaction groups, constructs round structures, and generates a corpus round structure set; The role behavior extraction module extracts verbs, subjects and business objects in the statements based on the corpus round structure set, screens financial, loan financial project object words, records the form of language, verb combination and modification structure, and generates a subject-predicate semantic action list; The risk chain identification module analyzes the action path and the change of the role pointing according to the subject-predicate semantic action list, constructs an action path, identifies an operation chain around the same object, screens a path with a continuation trend, and generates a suspected illegal semantic chain structure set; The risk chain identification module comprises: The action purpose identification submodule obtains the verb and business object combination recorded in the subject-predicate semantic action list, identifies the action purpose of the same object item in different statements, judges whether the same role uses different action intentions in different statements, and generates a use change state; The object operation chain construction submodule extracts all verb combinations corresponding to the associated object item according to the use change state, identifies the action sequence of the role around the same object, arranges the operation nodes according to the statement rounds, marks the corresponding relationship between the role and the object in the path, and generates an object operation path set; The semantic chain screening submodule calls the object operation path set, detects the connection mode of continuous actions in the path in the semantic structure, screens path sequences with a continuation trend in the verb combination mode, judges the action continuity between the roles and the object co-occurrence quantity in the path, and generates a suspected illegal semantic chain structure set; The semantic trend identification module analyzes the suspected illegal semantic chain structure set, extracts mood words, tone and modification items, compares the mood changes at the beginning and end of the link, judges the word trend change, and generates a semantic change path marker set; The semantic trend identification module comprises: The sentence extraction submodule calls the sentence paragraphs in the suspected illegal semantic chain structure set, extracts the beginning and end of the sentence paragraph, identifies the mood words, tone structure and modification items in the sentence paragraph, extracts the grammatical position and sentence relationship of the language components, and generates a pragmatic structure extraction value; The word change identification submodule identifies the change mode of the mood word category and the tone structure in the beginning and end of the sentence paragraph according to the pragmatic structure extraction value, extracts the continuity characteristics of the modification item combination mode, judges whether the word change forms a consistent semantic shift, and generates a word trend change rate; The direction feature labeling submodule calls the word trend change rate, extracts the semantic direction information of the continuously changed paragraphs in the semantic chain, identifies the structure shift trend in the path node, labels the semantic shift position and direction pointing type, and generates a semantic change path marker set; The illegal path calibration module judges whether there is a repeated mentioning behavior for a single business content according to the semantic change path marker set, verifies whether it constitutes a dominant tendency, combines the role, object and behavior mode, and generates an illegal tendency content segment set. 2.The compliance management system for corpus data based on financial services according to claim 1, wherein: The corpus turn structure set includes roles, sequence numbers, time information, and turn numbers, the subject-predicate semantic action list includes verbs, subjects, business object items, word forms, verb combinations, and modification structures, the suspected violation semantic chain structure set includes action purposes, role orientations, action continuous paths, operation chains, and path sequences, the semantic change path marker set includes modal words, mood structures, modification items, modal change conditions, and direction adjustment features, and the violation tendency content segment set includes role information, object items, repeated mentioning behaviors, content organization modes, and dominant tendencies. 3.The compliance management system for corpus data based on financial services according to claim 1, wherein: The corpus sequence management module includes: A sentence labeling submodule obtains speaking sentences of customers and business personnel in a financial customer service session, extracts speaking roles, sequence numbers, and start times corresponding to each sentence, sorts the sentences based on the sequence numbers and start times, constructs a sentence labeling structure in combination with the speaking role information, and generates a sentence number index value; A sentence aggregation submodule calls the sentence number index value, identifies a sentence group of continuous speaking of the same role, judges whether the speaking roles between adjacent sentences are consistent and whether the time interval is lower than a session continuity threshold, aggregates the sentence group that meets the conditions, arranges the numbers and start times of the combined sentence group, and generates a continuous sentence sequence amount; A turn generation submodule identifies speaking role switching points and time sequences based on the continuous sentence sequence amount, judges the arrangement relationship between the sentence groups, constructs a sorting list according to the turn logic, uniformly processes the turn numbers, and generates a corpus turn structure set. 4.The compliance management system for corpus data based on financial services according to claim 3, wherein: The role behavior extraction module includes: A semantic element identification submodule obtains the sentence content labeled in the corpus turn structure set, identifies the verbs, subjects, and object words of the role speaking in the sentence, extracts the combination relationship between the semantic elements, marks the correspondence between the subjects and verbs according to the word order structure, and generates a subject-verb pairing amount; A financial object filtering submodule extracts the object word content based on the subject-verb pairing amount, classifies and identifies all object words, judges whether they belong to financial product or loan service related items, filters the semantic combinations that meet the business type, and generates a financial subject-object combination; A semantic structure induction submodule calls the financial subject-object combination, extracts the modification words and collocation structures in the verb combination, sorts the collocation forms among the subject, verb, and object, induces the use order and structure style of the modification components, and generates a subject-predicate semantic action list. 5.The compliance management system for corpus data based on financial services according to claim 1, wherein: The violation path calibration module includes: A role extraction submodule calls the sentence paragraphs in the semantic change path marker set, extracts the role information and object item content in the sentence paragraphs, identifies the collocation relationship between the roles and objects according to the sentence structure, obtains the distribution features of the role semantic functions in the syntactic positions, and generates role-object corresponding values; A tendency verification submodule identifies the repeated pointing situation of the role to the object item in the difference sentence based on the role-object corresponding values, detects the recurrence mode of the role association form in the content structure, judges whether the role involved content forms a continuous organization tendency, and generates a role dominant degree coefficient; The content generation sub-module calls the role leading degree coefficient, filters the sentence segments in which the role and the object item are frequently bound in the structure combination, extracts the behavior language and the syntax pattern of the corresponding role, identifies the stable arrangement form of the semantic direction and the structure type, and generates a set of illegal tendency content segments. 6.The compliance management system for corpus data based on financial services according to claim 1, wherein: The speaking sentence refers to the natural language text formed by the customer or the service personnel in the financial service dialogue, the sentence has a clear syntax structure and semantic subject-predicate relationship, and belongs to unstructured language data. The turn structure refers to a two-dimensional structure constructed according to the time sequence of the sentence and the alternating sequence of the speaking role. The business object refers to a specific product, service or operation content in the financial communication, including a recognizable business core entity such as a financial product, a loan application, and account management, which is the receiving target of the semantic behavior. The language form refers to the language expression structure composed of the subject, verb and modifier in the sentence. 7.The compliance management system for corpus data based on financial services according to claim 1, wherein: The action path refers to a behavior sequence composed of a plurality of semantic verbs used by the role around the same business object item, and the path is used to identify the continuity of the operation logic under the same target. The operation chain refers to a connection sequence of semantic behaviors constructed by a plurality of roles or the same role based on a unified business object item in the corpus. The word usage trend refers to the language usage direction embodied by the modal word or the verb modifier in the continuous sentence, including the evolution process from the scheme type word to the commitment type word. The leading tendency refers to the language behavior feature of the single role in the sentence sequence, which forms the control of the information transmission direction by continuously using the leading verb and repeatedly pointing to the same business object item.

8. A compliance management method of corpus data based on financial business, characterized by, The compliance management system for corpus data based on financial business development according to any one of claims 1-7 comprises the following steps: S1: obtaining the role label, timestamp and sequence number in the financial customer service session corpus, identifying the sentence turn according to the speaking time sequence and the role alternating point, arranging and numbering each group of continuous sentences, and generating a set of corpus turn structures; S2: calling the role sentence in the set of corpus turn structures, extracting the subject, verb and business object item, filtering the financial products and loan service items appearing in the object item, marking the verb collocation structure and modifier component, and generating a list of subject-predicate semantic actions; S3: identifying the change of the role usage of the same object item in the difference sentence according to the verb and object combination in the list of subject-predicate semantic actions, constructing an operation path based on the action continuation structure, and generating a set of suspected illegal semantic chain structures; obtaining the verb and business object combination recorded in the list of subject-predicate semantic actions, identifying the action usage of the same object item in the difference sentence, judging whether the same role uses different action intentions in the difference sentence, and generating a usage change state; according to the usage change state, extracting all verb combinations corresponding to the associated object item, identifying the action sequence of the role around the same object, arranging the operation nodes according to the sentence turn, marking the corresponding relationship between the role and the object in the path, and generating a set of object operation paths; The object operation path set is called, the connection mode of continuous actions in the path on the semantic structure is detected, the path sequence with a continuous trend of verb combination mode is screened, the action continuity between roles and the number of object co-occurrence in the path are judged, and a suspected illegal semantic chain structure set is generated; S4: The sentence fragments in the suspected illegal semantic chain structure set are called, mood words, mood structures, and modification components are extracted, the word usage mode changes in the link start and end sentences are compared, it is judged whether the semantic content direction is adjusted, and a semantic change path marker set is generated; The sentence paragraphs in the suspected illegal semantic chain structure set are called, the start and end sentences of the link are extracted, mood words, mood structures, and modification items in the sentences are identified, the grammatical position and sentence structure relationship of language components are extracted, and a pragmatic structure extraction value is generated; According to the pragmatic structure extraction value, the change mode of mood word categories and mood structures in the start and end sentences is identified, the continuity characteristics of the modification item combination mode are extracted, it is judged whether the word changes form a direction-consistent semantic shift, and a word trend change rate is generated; The word trend change rate is called, the word meaning direction information of the continuously changed paragraphs in the semantic chain is extracted, the structure shift trend in the path node is identified, the semantic shift position and direction pointing type are marked, and a semantic change path marker set is generated; S5: According to the role and object item information extracted in the semantic change path marker set, the reuse situation of the same object item in the financial product or loan service scene is identified, the role and verb usage mode in the subject-object semantic action list are called, the sentence structure organization mode and object word distribution frequency are analyzed, the role content coverage degree is calculated, and an illegal tendency content fragment set is generated.

Images