Financial institution system information conflict detection method, device, equipment and medium

Abstract

The application discloses a conflict detection method, device and equipment of financial institution system information and a medium, relates to the technical field of computers, and comprises the following steps: when a system update event occurs in a financial institution, performing optical character recognition on new system information corresponding to the system update event based on a preset data processing strategy to determine a character recognition result; constructing a system clause tree based on the character recognition result, performing document fragmentation using the system clause tree to determine a first fragmentation result; determining current reference system information from old system information in a system knowledge base based on the first fragmentation result and a preset similarity fusion matching algorithm; and performing conflict detection on the first fragmentation result and the reference system information based on a preset large model to determine a conflict detection result corresponding to the system update event. The application can effectively solve the problems in the prior art, and improves the comprehensiveness, reliability, accuracy, flexibility and applicability of conflict detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer technology, and in particular to a conflict detection method, device and equipment for financial institution system information and a medium. BACKGROUND

[0002] In the operation of a financial institution, compliance is the key to stable development, and it is crucial to ensure that there is no conflict between new and old systems. However, in the existing large model and conflict detection related solutions, the following problems exist:

[0003] 1) Directly using a large model to compare two files, but the large model will ignore the part that exceeds the context window length, causing the comparison result to be incorrect, and the large model conflict checking effect is poor due to the large number of tokens input at one time;

[0004] 2) The traditional keyword or rule matching based solution relies on manual detection rules, which lacks intelligence and can only identify surface contradictions, making it difficult to capture semantic level or cross-clause implicit conflicts. In addition, this method has poor scalability and needs to be constantly updated by humans, which is inefficient and has limited accuracy, prone to missed and false detections, and cannot meet the needs of automated conflict detection of large-scale and diverse system files. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a conflict detection method, device and equipment for financial institution system information, which can effectively solve the problems existing in the related solutions, improve the comprehensiveness, reliability, accuracy, flexibility and applicability of conflict detection, and thus improve the user experience. The specific scheme is as follows:

[0006] In a first aspect, the present application provides a conflict detection method for financial institution system information, comprising:

[0007] When a system update event occurs in a financial institution, based on a preset data processing strategy, performing optical character recognition on new system information corresponding to the system update event to determine a character recognition result;

[0008] Based on the character recognition result, constructing a system clause tree, and using the system clause tree to perform document fragmentation to determine a first fragmentation result;

[0009] Based on the first fragmentation result and a preset similarity fusion matching algorithm, determining the current reference system information from the old system information in the system knowledge base;

[0010] Based on a preset large model, performing conflict detection on the first fragmentation result and the reference system information to determine a conflict detection result corresponding to the system update event.

[0011] Optionally, the new institutional information corresponding to the institutional update event is subjected to optical character recognition based on a preset data processing strategy, including:

[0012] The new institutional information corresponding to the institutional update event is obtained through a preset interface; wherein the new institutional information includes institutional files in several formats;

[0013] The new institutional information is subjected to skew correction to determine corrected information;

[0014] The corrected information is subjected to denoising to determine denoised information;

[0015] The denoised information is subjected to contrast enhancement to determine enhanced information;

[0016] The enhanced information is processed based on a multi-modal optical character recognition algorithm to determine a character recognition result.

[0017] Optionally, the institutional clause tree is constructed based on the character recognition result, and the document is fragmented using the institutional clause tree, including:

[0018] The character recognition result is subjected to identification and matching of structural identifiers through a regular expression template library and a sequence labeling model based on a pre-trained language model to determine an identifier matching result;

[0019] The parent-child hierarchy is constructed based on the identifier matching result to determine a hierarchy construction result;

[0020] Whether there is an identifier jump in the character recognition result is determined based on the identifier matching result and a preset model to determine a jump determination result;

[0021] If the jump determination result is yes, it is determined whether there is a cross-page break or optical character recognition error, and when the corresponding target determination result is yes, a corresponding jump processing operation is triggered to construct the institutional clause tree using the obtained new character recognition result;

[0022] If the jump determination result is no, the tree-shaped semantic structure is reconstructed based on the identifier matching result and the hierarchy construction result to determine the institutional clause tree;

[0023] The semantic association degree between adjacent sentences is analyzed based on the length information and semantic vectors corresponding to each sentence in the character recognition result to determine the semantic continuity degree corresponding to each sentence;

[0024] The potential fragmentation boundary is identified based on the semantic continuity degree and the text length corresponding to the sub-leaf in the institutional clause tree to determine a target split point;

[0025] Based on the target segmentation point, the character recognition result and the system clause tree, document segmentation is performed to determine a first segmentation result.

[0026] Optionally, based on the first segmentation result and a preset similarity fusion matching algorithm, current reference system information is determined from old system information in a system knowledge base, including:

[0027] Based on the old system information in the system knowledge base, document segmentation is performed to determine a second segmentation result;

[0028] For any system information segment in the first segmentation result and the second segmentation result, based on a pre-trained language model, semantic vector representations corresponding to the current two system information segments are determined respectively;

[0029] Based on the cosine similarity between the semantic vector representations, a corresponding semantic similarity is determined;

[0030] Each word segmentation code in the current two system information segments is subjected to word vector conversion to determine a conversion result;

[0031] Each word vector in the conversion result is configured with a corresponding information entropy weight;

[0032] Based on the information entropy weight and the conversion result, an entropy weight similarity between the current two system information segments is determined;

[0033] Based on the semantic similarity and the entropy weight similarity, weighted fusion is performed to determine a target similarity corresponding to the current two system information segments;

[0034] According to the size of the value, the target similarity is sorted to determine a sorting result;

[0035] Based on the sorting result, reference system information is determined.

[0036] Optionally, based on a preset large model, conflict detection is performed on the first segmentation result and the reference system information to determine a conflict detection result corresponding to the system update event, including:

[0037] A prompt word template corresponding to the system update event is determined;

[0038] Through a preset large model and in combination with the prompt word template, each system information segment of the first segmentation result and the reference system information are subjected to conflict, consistency and missing relationship detection to determine a conflict detection result.

[0039] Optionally, after the determination of the conflict detection result corresponding to the system update event, the method further comprises:

[0040] If the conflict detection result indicates that a plurality of system information segments in the first segmentation result are determined as conflict segments;

[0041] Based on the large model annotation information corresponding to the conflict segments, it is determined whether a preset segmentation backtracking condition is triggered to determine a condition determination result;

[0042] If the condition determination result is yes, based on a preset segmentation backtracking strategy and the conflict segments, the step of jumping back to the document segmentation using the system clause tree is performed.

[0043] Optionally, the method further comprises:

[0044] In the process of determining the current reference system information, based on a first preset algorithm, keyword extraction is performed on the current two system information segments to determine a keyword extraction result;

[0045] Based on the keyword extraction result, a keyword similarity corresponding to the current two system information segments is determined;

[0046] Based on a second preset algorithm, time expression extraction is performed on the current two system information segments to determine a timing information extraction result;

[0047] The timing information extraction result is normalized, and based on the corresponding normalized result, comparison is performed to determine a timing similarity corresponding to the current two system information segments;

[0048] Based on a third preset algorithm, entity extraction is performed on the current two system information segments to determine an entity extraction result;

[0049] Based on a preset knowledge graph and the entity extraction result, comparison is performed to determine an entity similarity corresponding to the current two system information segments;

[0050] Based on the keyword similarity, the timing similarity, the entity similarity, the semantic similarity, and the entropy weight similarity, weighted fusion is performed to determine the target similarity.

[0051] In a second aspect, the application provides a conflict detection device for financial institution system information, comprising:

[0052] A character recognition module is configured to, when a financial institution has a system update event, perform optical character recognition on new system information corresponding to the system update event based on a preset data processing strategy to determine a character recognition result;

[0053] a document fragmentation module configured to construct a regulation clause tree based on the character recognition result, and perform document fragmentation using the regulation clause tree to determine a first fragmentation result;

[0054] a fusion matching module configured to determine, based on the first fragmentation result and a preset similarity fusion matching algorithm, current reference regulation information from old regulation information in a regulation knowledge base;

[0055] a conflict detection module configured to perform conflict detection on the first fragmentation result and the reference regulation information based on a preset large model to determine a conflict detection result corresponding to the regulation update event.

[0056] In a third aspect, the present application provides an electronic device, comprising:

[0057] a memory configured to store a computer program;

[0058] a processor configured to execute the computer program to implement the steps of the conflict detection method for financial institution regulation information.

[0059] In a fourth aspect, the present application provides a computer readable storage medium configured to store a computer program, which, when executed by a processor, implements the steps of the conflict detection method for financial institution regulation information.

[0060] As can be seen, in the present application, when a regulation update event occurs in a financial institution, optical character recognition is performed on new regulation information corresponding to the regulation update event based on a preset data processing strategy to determine a character recognition result; a regulation clause tree is constructed based on the character recognition result, and document fragmentation is performed using the regulation clause tree to determine a first fragmentation result; current reference regulation information is determined from old regulation information in a regulation knowledge base based on the first fragmentation result and a preset similarity fusion matching algorithm; and conflict detection is performed on the first fragmentation result and the reference regulation information based on a preset large model to determine a conflict detection result corresponding to the regulation update event. That is, in the present application, new regulation information corresponding to a regulation update event is first obtained, optical character recognition is performed, a regulation clause tree is constructed using the determined character recognition result to complete document fragmentation and obtain a first fragmentation result. Then, current reference regulation information is determined from old regulation information in a regulation knowledge base based on the first fragmentation result and a preset similarity fusion matching algorithm, and conflict detection is performed on the first fragmentation result and the reference regulation information using a combination of a preset large model to determine a conflict detection result corresponding to the regulation update event. In this way, the problems existing in the prior related solutions can be effectively solved, the comprehensiveness, reliability, accuracy, flexibility and applicability of conflict detection are improved, and the user experience is improved. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute a part of the embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0062] Figure 1 A conflict detection method flow chart of financial institution system information provided by the present application;

[0063] Figure 2 A specific conflict detection method flow chart of financial institution system information provided by the present application;

[0064] Figure 3 A conflict detection device structure schematic diagram of financial institution system information provided by the present application;

[0065] Figure 4 An electronic equipment structure diagram provided by the present application. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute a 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 skilled in the art without creative labor fall within the scope of protection of the present application.

[0067] In the prior large model and conflict detection related solutions, the following problems exist: 1) directly using a large model to compare two files, but the large file will ignore the part exceeding the context window length of the large model, causing the comparison result to be wrong, and the large model conflict checking effect is poor due to too many tokens input at one time; 2) the traditional solution based on keyword or rule matching relies on manual detection rules, and the intelligent degree is insufficient, and can only identify surface contradictions, and it is difficult to capture semantic level or cross-clause implicit conflicts. In addition, this method has poor scalability, and needs to be constantly updated by manual rules, which is low in efficiency and limited in accuracy, and is prone to miss detection and false detection problems, and cannot adapt to the automatic conflict detection needs of large-scale and diversified system files.

[0068] Therefore, the present application provides a conflict detection solution of financial institution system information, which can effectively solve the problems existing in the prior related solutions, improve the comprehensiveness, reliability, accuracy, flexibility and applicability of conflict detection, and further improve the user experience.

[0069] Referring toFigure 1 As shown, the embodiment of the present application discloses a conflict detection method of financial institution system information, comprising:

[0070] Step S11, when a system update event occurs in a financial institution, based on a preset data processing strategy, performing optical character recognition on new system information corresponding to the system update event to determine a character recognition result.

[0071] In combination with Figure 2 As shown, in the embodiment, when a system update event occurs in a financial institution, the system of the present scheme first receives a system file in Word, PDF or other formats uploaded by a user, and then performs OCR (Optical Character Recognition) recognition on the file, that is, through a preset interface, new system information corresponding to the system update event is obtained; wherein the new system information includes system files in several formats; the new system information is subjected to inclination correction to determine corrected information; the corrected information is subjected to denoising to determine denoised information; the denoised information is subjected to contrast enhancement to determine enhanced information; based on a multi-modal optical character recognition algorithm, the enhanced information is processed to determine a character recognition result. It can be understood that in the embodiment, the multi-modal OCR technology is used to recognize the characters in the file, and the adaptive image preprocessing algorithm is used in the recognition process, including inclination correction, denoising, contrast enhancement and other processes, to ensure the accuracy of character recognition.

[0072] Step S12, constructing a system clause tree based on the character recognition result, and using the system clause tree to perform document fragmentation to determine a first fragmentation result.

[0073] In this embodiment, after character recognition is completed, the recognition result is optimized according to the characteristics of the regulations and systems, a regular expression template library is used to match the structural identifiers, a system clause tree is constructed, and document segmentation is completed, that is, the character recognition result is identified and matched with the structural identifiers by using the regular expression template library and a sequence labeling model based on a pre-trained language model to determine the identifier matching result; the parent-child hierarchy is constructed based on the identifier matching result to determine the hierarchy construction result; whether there is an identifier jump in the character recognition result is determined based on the identifier matching result and a preset model to determine the jump determination result; if the jump determination result is yes, it is determined whether there is a cross-page break or an optical character recognition error, and when the corresponding target determination result is yes, a corresponding jump processing operation is triggered to construct a system clause tree using the obtained new character recognition result; if the jump determination result is no, the tree-shaped semantic structure is reconstructed based on the identifier matching result and the hierarchy construction result to determine the system clause tree; the semantic association degree between adjacent sentences in the character recognition result is analyzed based on the length information and semantic vectors corresponding to each sentence to determine the semantic continuity degree corresponding to each sentence; the potential segmentation boundary is identified based on the semantic continuity degree and the text length corresponding to the sub-leaf in the system clause tree to determine the target segmentation point; and the document is segmented based on the target segmentation point, the character recognition result, and the system clause tree to determine the first segmentation result. The matching of the structural identifiers can be as follows:

[0074] ^Chapter [one two three four five six seven eight nine hundred thousand 0-9]+;

[0075] ^Article [one two three four five six seven eight nine hundred thousand 0-9]+;

[0076] ^([one two three four five six seven eight nine hundred 0-9]+)$;

[0077] ^[0-9]+.

[0078] It can be understood that, in constructing the system clause tree, in addition to regular expression matching, a sequence labeling model based on a pre-trained language model (such as BERT) can also be introduced to identify chapters, clauses, items, and other structures to improve the processing capability of non-standard format files (tables, flowcharts, etc.) and enhance the structural identification capability.

[0079] It needs to be understood that, regarding the construction of the system clause tree, after completing the structure identifier matching, the embodiment automatically identifies the hierarchical structure of the chapters, clauses, items, etc. of the regulations and rules, constructs parent-child relationship rules according to the detected hierarchical numbers, for example, “chapter” is a first-level node; “clause” is a second-level node; “item” and “clause” are third-level nodes. Finally, the parent-child relationship rules are reconstructed into a tree-shaped semantic structure that can be understood by machines. If a number jump is encountered, a semantic continuity model can be used to assist in determining whether there is a cross-page break or an OCR recognition error, and when there is, a corresponding jump processing operation is triggered to use the obtained new character recognition result to construct the system clause tree.

[0080] Further, regarding document segmentation, in the embodiment, the character recognition result is segmented according to the child leaves in the system clause tree, so as to divide the document fragments corresponding to different hierarchical nodes into different child leaves, and if the total length of the texts of all sentences in a child leaf in the clause tree exceeds 512, the following steps are used to segment the child leaf:

[0081] (1) The length of each sentence in the child leaf is defined as , and the semantic vector is ;

[0082] (2) The semantic continuity of each sentence is calculated as , which is used to measure the semantic association strength between the current sentence and the adjacent sentence:

[0083] ;

[0084] In the formula, represents the modulus of the vector; i+1 represents the adjacent sentence of the current sentence (i). A lower indicates a larger semantic jump, which is a potential segmentation boundary.

[0085] .

[0086] In the formula, is a value used to measure the semantic integrity of the current sentence; is the target length of the segmentation; represents the length of the first sentence in the child leaf, j=1; represents the total length of the texts from the first sentence of the child leaf to the current sentence (i); represents the absolute value symbol, which can be adjusted dynamically by adjusting the weights and to achieve a bias towards semantic integrity or length balance. Finally, the code is traversed to select the boundary with the maximum score:

[0087] .

[0088] That is, by comparing each sentence in the sub-leaf , the sentence with the largest semantic jump and the most balanced length (i.e. The largest sentence) can be selected as the cutting point at the end punctuation mark, and the two document segments obtained after cutting will be recorded on two different sub-leaves. In this way, by constructing the institutional clause tree and the semantic continuity-based segmentation algorithm, the large file is divided into segments less than 512 characters, ensuring that the text content segments on each sub-leaf can adapt to the context window of the large model. This avoids errors caused by ignoring the super-long part when directly inputting a large file. By calculating the semantic continuity, the segmentation boundary is ensured at the semantic jump point, reducing the risk of cutting off semantic coherence, thereby improving the accuracy of conflict detection. Thus, the context window limitation of the large model when processing large files and the performance degradation caused by excessive tokens are effectively solved.

[0089] Step S13, based on the first segmentation result and the preset similarity fusion matching algorithm, determining the current reference institutional information from the old institutional information in the institutional knowledge base.

[0090] In this embodiment, after the segmentation is completed, the obtained first segmentation result is subjected to multi-dimensional similarity matching with the existing institutions in the institutional knowledge base, including semantic similarity and entropy weight similarity, the matching degree of each piece with the institutions in the knowledge base is calculated, and the institutional piece with the highest matching degree is selected as the reference, that is, the document is segmented based on the old institutional information in the institutional knowledge base to determine the second segmentation result; for any institutional information piece in the first segmentation result and the second segmentation result, based on the pre-trained language model, the semantic vector representation corresponding to the current two institutional information pieces is determined; based on the cosine similarity between the semantic vector representations, the corresponding semantic similarity is determined; the word vector transformation is performed on each word segmentation code in the current two institutional information pieces to determine the transformation result; the corresponding information entropy weight is configured for each word vector in the transformation result; based on the information entropy weight and the transformation result, the entropy weight similarity between the current two institutional information pieces is determined; the semantic similarity and the entropy weight similarity are weighted and fused to determine the target similarity corresponding to the current two institutional information pieces; the target similarity is sorted according to the value to determine the sorting result; based on the sorting result, the reference institutional information is determined.

[0091] It can be understood that the embodiment proposes a multi-dimensional similarity fusion algorithm, which comprehensively considers the semantic similarity and the entropy weight similarity, and realizes more accurate evaluation of the similarity of two system information segments by using weighted fusion. The algorithm first calculates the similarity of two dimensions respectively, then gives corresponding weights according to the importance of each dimension, and finally fuses the weighted similarity to obtain the final similarity score. Then, according to the score, one or more with the maximum score is selected as the reference system information.

[0092] (1) Semantic similarity calculation.

[0093] The pre-trained language model is used to obtain the semantic vector representation of the two system information segments, and then the cosine similarity between the two semantic vectors is calculated as the semantic similarity.

[0094] Suppose the semantic vectors of the two system information segments obtained by the pre-trained language model are and , then the calculation formula of the semantic similarity is:

[0095] ;

[0096] Among them, represents the dot product of the vector, and the modulus of the vector is multiplied.

[0097] (2) Entropy weight similarity calculation.

[0098] Each word segmentation in the two system information segments is converted into a corresponding word vector: and .

[0099] Each word vector is assigned a weight of information entropy.

[0100] ;

[0101] Among them, is the probability of the word appearing in the corpus, and -log indicates that the rarer the word is, the more important it is.

[0102] The similarity of the two system information segments :

[0103] .

[0104] Among them, represents semantic consistency; represents semantic distance penalty; represents importance weighting (rare words have greater influence).

[0105] (3) Multi-dimensional similarity fusion.

[0106] According to the importance of semantic similarity, document similarity and keyword similarity, they are respectively given corresponding weights , , and satisfy + =1. The final similarity score S is calculated by the formula:

[0107] .

[0108] The value of the weight can be adjusted according to the specific application scenario and actual demand, using methods such as expert experience or data-driven method.

[0109] In summary, it can capture the implicit conflict at the semantic level and across clauses, not just relying on surface keywords or rules. Entropy weight similarity emphasizes the importance of rare words, which helps to identify subtle differences in key clauses.

[0110] In addition, in a specific embodiment, additional similarity dimensions can also be introduced to help more comprehensively and multi-dimensionally compare two system documents, so as to capture more potential conflict points, including but not limited to keyword similarity, time sequence similarity, and entity similarity, that is, in the process of determining the current reference system information, based on a first preset algorithm, keyword extraction is performed on the current two system information segments to determine a keyword extraction result; based on the keyword extraction result, the keyword similarity corresponding to the current two system information segments is determined; based on a second preset algorithm, time expression extraction is performed on the current two system information segments to determine a time information extraction result; the time information extraction result is normalized, and comparison is made based on the corresponding normalized result to determine the time sequence similarity corresponding to the current two system information segments; based on a third preset algorithm, entity extraction is performed on the current two system information segments to determine an entity extraction result; comparison is made based on a preset knowledge graph and the entity extraction result to determine the entity similarity corresponding to the current two system information segments; based on the keyword similarity, the time sequence similarity, the entity similarity, the semantic similarity, and the entropy weight similarity, weighted fusion is performed to determine the target similarity. Wherein, the first preset algorithm can be a term frequency-inverse document frequency algorithm or a TextRank algorithm (a keyword extraction technology derived based on a PageRank algorithm), the second preset algorithm can be a regular expression or a named entity recognition algorithm, and the third preset algorithm can be a named entity recognition algorithm. The entity includes but is not limited to an institutional person and a place, and the calculation of the entity similarity not only involves comparing whether the entities are consistent, but also involves comparing whether the relationships between the entities are consistent. Further, in addition to the above similarity dimensions, a structure similarity can also be added, which compares whether the hierarchical structures of the two documents are similar. The difference in structure can imply that the logic of system arrangement is different, which can further lead to conflicts; a logic similarity can also be added, which compares two documents by analyzing the logical relationship in the clauses. The inconsistency of the logical structure can directly lead to conflicts.

[0111] Step S14, based on a preset large model, performing conflict detection on the first segmentation result and the reference system information to determine a conflict detection result corresponding to the system update event.

[0112] In this embodiment, after determining the reference system information, conflict detection is performed in combination with the large model and the first fragmentation result, that is, a prompt word template corresponding to the system update event is determined; the conflict, consistency and missing relationship of each system information segment of the first fragmentation result and the reference system information is detected through a preset large model in combination with the prompt word template to determine a conflict detection result. It can be understood that in this embodiment, a customized prompt word project is designed, the fragmented content and the matched system segment are input into the large model, the system prompt word is used to clearly indicate the task requirements, the large model is guided to judge whether there is a conflict, consistency or missing relationship between the two, and the conflict reason when the conflict occurs is output.

[0113] Further, in this embodiment, after determining the conflict detection result, if the conflict detection result indicates that several system information segments in the first fragmentation result are determined as conflict segments; based on the large model annotation information corresponding to the conflict segments, it is judged whether a preset fragmentation backtracking condition is triggered to determine a condition judgment result; if the condition judgment result is yes, then based on a preset fragmentation backtracking strategy and the conflict segments, the step of utilizing the system clause tree to fragment the document is re-visited. That is, if a certain fragment is specially marked by the large model in conflict detection, for example: 'ambiguous' or 'need more context', backtracking and re-fragmentation are allowed, for example, by expanding the fragmentation boundary or merging adjacent segments, to ensure that the key context is not lost.

[0114] In addition, in this embodiment, when a conflict is detected, iterative detection can be triggered and started, that is, re-fragmentation or expansion of the context range is performed to confirm whether the conflict really exists, reduce false positives, and the final output result contains specific conflict types, such as term conflict, authority conflict, time conflict, etc., for subsequent processing and reporting.

[0115] As can be seen, in this application, new system information corresponding to the system update event is first acquired, optical character recognition is performed, system clause tree construction is performed using the determined character recognition result to complete document fragmentation and obtain a first fragmentation result. Then, based on the first fragmentation result and a preset similarity fusion matching algorithm, reference system information is determined from old system information in the system knowledge base, and then conflict detection is performed on the first fragmentation result and the reference system information in combination with a preset large model to determine a conflict detection result corresponding to the system update event. In this way, the problems existing in the related solutions can be effectively solved, the comprehensiveness, reliability, accuracy, flexibility and applicability of conflict detection are improved, and the user experience is improved.

[0116] Referring to Figure 3 The application embodiment also discloses a financial institution system information conflict detection device, which comprises:

[0117] a character recognition module 11, configured to perform optical character recognition on new institution information corresponding to the institution update event based on a preset data processing strategy to determine a character recognition result when an institution update event occurs in a financial institution;

[0118] a document fragmentation module 12, configured to construct an institution clause tree based on the character recognition result and perform document fragmentation using the institution clause tree to determine a first fragmentation result;

[0119] a fusion matching module 13, configured to determine current reference institution information from old institution information in an institution knowledge base based on the first fragmentation result and a preset similarity fusion matching algorithm;

[0120] a conflict detection module 14, configured to perform conflict detection on the first fragmentation result and the reference institution information based on a preset large model to determine a conflict detection result corresponding to the institution update event.

[0121] In some embodiments, the character recognition module 11 can specifically include:

[0122] an information receiving unit, configured to acquire new institution information corresponding to the institution update event through a preset interface; the new institution information includes institution files in several formats;

[0123] a correction unit, configured to perform skew correction on the new institution information to determine corrected information;

[0124] a denoising unit, configured to perform denoising on the corrected information to determine denoised information;

[0125] a contrast enhancement unit, configured to perform contrast enhancement on the denoised information to determine enhanced information;

[0126] a recognition unit, configured to process the enhanced information based on a multi-modal optical character recognition algorithm to determine a character recognition result.

[0127] In some embodiments, the document fragmentation module 12 can specifically include:

[0128] an identifier matching unit, configured to perform structural identifier identification and matching on the character recognition result through a regular expression template library and a sequence labeling model based on a pre-trained language model to determine an identifier matching result;

[0129] a hierarchical construction unit, configured to perform parent-child hierarchical construction based on the identifier matching result to determine a hierarchical construction result;

[0130] The jump determination unit is configured to determine, based on the identifier matching result and a preset model, whether there is an identifier jump in the character recognition result, to obtain a jump determination result.

[0131] The jump processing unit is configured to, if the jump determination result is yes, determine whether there is a cross-page break or an optical character recognition error, and trigger a corresponding jump processing operation when a corresponding target determination result is yes, to construct a regulation clause tree by using a new character recognition result.

[0132] The reconstruction unit is configured to, if the jump determination result is no, reconstruct a tree-shaped semantic structure based on the identifier matching result and the hierarchical construction result, to determine a regulation clause tree.

[0133] The continuity determination unit is configured to analyze the semantic association degree between adjacent sentences based on the length information and the semantic vectors corresponding to each sentence in the character recognition result, to determine the semantic continuity of each sentence.

[0134] The split point determination unit is configured to identify a potential segmentation boundary based on the semantic continuity and the length of the text corresponding to a sub-leaf in the regulation clause tree, to determine a target split point.

[0135] The first segmentation unit is configured to segment a document based on the target split point, the character recognition result, and the regulation clause tree, to determine a first segmentation result.

[0136] In some embodiments, the fusion matching module 13 can specifically include:

[0137] The second segmentation unit is configured to segment a document based on old regulation information in a regulation knowledge base, to determine a second segmentation result.

[0138] The vector determination unit is configured to determine, for any regulation information segment in the first segmentation result and the second segmentation result, a semantic vector representation corresponding to the two regulation information segments respectively based on a pre-trained language model.

[0139] The semantic similarity determination unit is configured to determine a semantic similarity between the semantic vector representations based on a cosine similarity.

[0140] The conversion unit is configured to convert each word segmentation in the two regulation information segments into a word vector, to determine a conversion result.

[0141] The weight configuration unit is configured to configure an information entropy weight for each word vector in the conversion result.

[0142] An entropy weight similarity determination unit is configured to determine an entropy weight similarity between the two current institutional information segments based on the information entropy weight and the conversion result.

[0143] A first weighted fusion unit is configured to perform weighted fusion based on the semantic similarity and the entropy weight similarity to determine a target similarity corresponding to the two current institutional information segments.

[0144] An ordering unit is configured to order the target similarities according to the values to determine an ordering result.

[0145] A reference institutional information determination unit is configured to determine reference institutional information based on the ordering result.

[0146] In some embodiments, the conflict detection module 14 can specifically include:

[0147] A template determination unit is configured to determine a prompt word template corresponding to the institutional update event.

[0148] A conflict detection unit is configured to detect conflicts, consistencies and missing relationships among the institutional information segments in the first segmentation result and the reference institutional information by using a preset large model and in combination with the prompt word template to determine a conflict detection result.

[0149] In some embodiments, the conflict detection apparatus for the financial institution institutional information can further include:

[0150] A condition judgment unit is configured to judge whether a preset segmentation backtracking condition is triggered based on large model annotation information corresponding to the conflict segments if the conflict detection result indicates that a plurality of the institutional information segments in the first segmentation result are determined as conflict segments to determine a condition judgment result.

[0151] A step jumping unit is configured to jump back to the step of segmenting the document by using the institutional clause tree based on a preset segmentation backtracking strategy and the conflict segments if the condition judgment result is yes.

[0152] In some embodiments, the conflict detection apparatus for the financial institution institutional information can further include:

[0153] A keyword extraction unit is configured to extract keywords from the two current institutional information segments based on a first preset algorithm to determine a keyword extraction result in the process of determining the reference institutional information.

[0154] A keyword similarity determination unit is configured to determine a keyword similarity corresponding to the two current institutional information segments based on the keyword extraction result.

[0155] a time sequence information extraction unit configured to extract time sequence information from the two current system information segments based on a second preset algorithm to obtain time sequence information extraction results;

[0156] a time sequence similarity determination unit configured to normalize the time sequence information extraction results and compare the normalized results to determine time sequence similarities between the two current system information segments;

[0157] an entity extraction unit configured to extract entities from the two current system information segments based on a third preset algorithm to obtain entity extraction results;

[0158] an entity similarity determination unit configured to compare the entity extraction results based on a preset knowledge graph to determine entity similarities between the two current system information segments;

[0159] a second weighted fusion unit configured to perform weighted fusion based on the keyword similarity, the time sequence similarity, the entity similarity, the semantic similarity, and the entropy weight similarity to determine the target similarity.

[0160] Further, the embodiment of the present application further discloses an electronic device, Figure 4 is an electronic device 20 structure diagram shown according to an exemplary embodiment, the contents in the figure cannot be considered as any limitation on the use range of the present application.

[0161] Figure 4 A structure diagram of an electronic device 20 provided by the embodiment of the present application. The electronic device 20 can specifically include at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25 and a communication bus 26. Wherein, the memory 22 is used to store computer programs, the computer programs are loaded and executed by the processor 21 to realize the related steps in the financial institution system information conflict detection method disclosed by any of the preceding embodiments. In addition, the electronic device 20 in the embodiment can be an electronic computer.

[0162] In the embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol followed by the communication interface 24 can be any communication protocol applicable to the technical solution of the present application, which is not limited here; the input / output interface 25 is used to obtain external input data or output data to the outside world, and the specific interface type can be selected according to the specific application needs, which is not limited here.

[0163] In addition, the memory 22 can be a read-only memory, a random access memory, a magnetic disk or an optical disk, etc. as a carrier for storing resources, and the resources stored thereon can include an operating system 221, a computer program 222, etc. The storage mode can be temporary storage or permanent storage.

[0164] The operating system 221 is used to manage and control each hardware device on the electronic device 20 and the computer program 222, and can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program capable of completing the conflict detection method of the financial institution system information disclosed in any of the preceding embodiments and executed by the electronic device 20, the computer program 222 can further include a computer program capable of completing other specific work.

[0165] Further, the present application also discloses a computer readable storage medium for storing a computer program; wherein the computer program is executed by a processor to implement the conflict detection method of the financial institution system information disclosed above. For the specific steps of the method, please refer to the corresponding content disclosed in the preceding embodiments, which will not be described here.

[0166] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. For the same or similar parts between each embodiment, please refer to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and please refer to the method part for the relevant part.

[0167] The skilled person can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of both. In order to clearly show the interchangeability of hardware and software, the composition and steps of each example have been described in the above description. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0168] The steps of the method or algorithm described in combination with the embodiments disclosed herein can be directly implemented by hardware, a software module executed by a processor, or a combination of both. The software module can be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0169] Finally, it needs to be pointed out that in this document, relational terms such as first and second and the like can only be intended to distinguish one entity or operation from another entity or operation without necessarily requiring or implying any actual such relationship or order between such entities or operations. Moreover, the terms "comprising", "including", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the statement "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus including the stated element.

[0170] The above detailed description of the technical solutions provided by the present application has been provided, and the principles and implementation modes of the present application have been described by applying specific examples. The above description of the examples is only for the purpose of helping to understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed, and the above description of the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method of conflict detection of financial institution regime information, characterized by, The method comprises the following steps: When a financial institution has a system update event, based on a preset data processing strategy, optical character recognition is performed on new system information corresponding to the system update event to determine a character recognition result; Based on the character recognition result, a system clause tree is constructed, and document segmentation is performed using the system clause tree to determine a first segmentation result; Based on the first segmentation result and a preset similarity fusion matching algorithm, current reference system information is determined from old system information in a system knowledge base; Based on a preset large model, conflict detection is performed on the first segmentation result and the reference system information to determine a conflict detection result corresponding to the system update event; Wherein, based on the character recognition result, a system clause tree is constructed, and document segmentation is performed using the system clause tree, comprising: Through a regular expression template library and a sequence labeling model based on a pre-trained language model, the character recognition result is identified and matched with a structure identifier to determine an identifier matching result; Based on the identifier matching result, a parent-child hierarchy is constructed to determine a hierarchy construction result; Based on the identifier matching result and a preset model, it is determined whether there is an identifier jump in the character recognition result to determine a jump determination result; If the jump determination result is yes, it is determined whether there is a cross-page break or optical character recognition error, and when the corresponding target determination result is yes, a corresponding jump processing operation is triggered to construct a system clause tree using the obtained new character recognition result; If the jump determination result is no, based on the identifier matching result and the hierarchy construction result, a tree-shaped semantic structure is reconstructed to determine a system clause tree; Based on the length information and semantic vectors corresponding to each sentence in the character recognition result, the semantic association degree between adjacent sentences is analyzed to determine the semantic continuity degree corresponding to each sentence; Based on the semantic continuity degree and the text length corresponding to the sub-leaf in the system clause tree, a potential segmentation boundary is identified to determine a target split point; Based on the target split point, the character recognition result, and the system clause tree, document segmentation is performed to determine a first segmentation result; Based on the first segmentation result and a preset similarity fusion matching algorithm, current reference system information is determined from old system information in a system knowledge base, comprising: Based on the old system information in the system knowledge base, document segmentation is performed to determine a second segmentation result; Based on the pre-trained language model, the semantic vector representation corresponding to the current two system information segments is determined for any system information segment in the first segmentation result and the second segmentation result; Based on the cosine similarity between the semantic vector representations, the corresponding semantic similarity is determined; Each word segmentation code in the current two system information segments is converted into a word vector to determine a conversion result; Each word vector in the conversion result is configured with a corresponding information entropy weight; Determine an entropy weight similarity between the two current institutional information segments based on the information entropy weight and the conversion result; Perform weighted fusion based on the semantic similarity and the entropy weight similarity to determine a target similarity corresponding to the two current institutional information segments; Sort the target similarity according to the size of the value to determine a sorting result; Determine reference institutional information based on the sorting result; The identification of the potential segmentation boundary based on the semantic continuity and the length of the text corresponding to the sub-leaf in the institutional clause tree comprises: ； wherein, is the semantic continuity degree for each sentence in the sub-leaf; is a value for measuring the semantic completeness degree of the current sentence; is the target length of division; represents the length of the first sentence in the sub-leaf, j = 1; represents the total length of characters accumulated from the first sentence of the sub-leaf to the current sentence (i); represents the absolute value symbol; and is the weight; The determination of the entropy weight similarity between the two current institutional information segments based on the information entropy weight and the conversion result comprises: ； wherein, is the entropy weight similarity between the current two said institutional information segments; denotes semantic consistency; denotes semantic distance penalty; denotes importance weighting, i.e. the information entropy weight; , is the transformation result corresponding to the current two said institutional information segments, and , denotes the word vector.

2. The conflict detection method of financial institution system information according to claim 1, characterized in that, The optical character recognition of the new institutional information corresponding to the institutional update event based on the preset data processing strategy comprises: Obtain new institutional information corresponding to the institutional update event through a preset interface; wherein the new institutional information comprises institutional files in several formats; Perform skew correction on the new institutional information to determine corrected information; Perform denoising on the corrected information to determine denoised information; Perform contrast enhancement on the denoised information to determine enhanced information; Process the enhanced information based on a multi-modal optical character recognition algorithm to determine a character recognition result.

3. The financial institution system information conflict detection method of claim 1, wherein The conflict detection based on the preset large model and the first segmentation result and the reference institutional information to determine a conflict detection result corresponding to the institutional update event comprises: Determine a prompt word template corresponding to the institutional update event; Perform conflict, consistency and missing relationship detection on each institutional information segment of the first segmentation result and the reference institutional information through a preset large model combined with the prompt word template to determine a conflict detection result.

4. The financial institution system information conflict detection method according to claim 3, characterized by, After determining the conflict detection result corresponding to the institutional update event, the method further comprises: If the conflict detection result indicates that several institutional information segments in the first segmentation result are determined as conflict segments; Determine whether a preset segmentation backtracking condition is triggered based on the large model annotation information corresponding to the conflict segments to determine a condition determination result; If the condition determination result is yes, then based on a preset segmentation backtracking strategy and the conflict segments, jump back to the step of performing document segmentation using the institutional clause tree.

5. The method of conflict detection of financial institution system information according to claim 1, wherein, Further comprising: In the process of determining the current reference institutional information, extract keywords from the two current institutional information segments based on a first preset algorithm to determine a keyword extraction result; Determine a keyword similarity corresponding to the two current institutional information segments based on the keyword extraction result; Extract time expressions from the two current institutional information segments based on a second preset algorithm to determine a time sequence information extraction result; Normalize the time sequence information extraction result and compare the corresponding normalized results to determine a time sequence similarity corresponding to the two current institutional information segments; based on a third preset algorithm, entity extraction is performed on the current two system information segments to determine an entity extraction result; based on a preset knowledge graph and the entity extraction result, comparison is performed to determine an entity similarity corresponding to the current two system information segments; based on the keyword similarity, the time sequence similarity, the entity similarity, the semantic similarity, and the entropy weight similarity, weighted fusion is performed to determine the target similarity.

6. An apparatus for conflict detection of financial institution system information, characterized by comprising: Comprise: a character recognition module, configured to, when a system update event occurs in a financial institution, perform optical character recognition on new system information corresponding to the system update event based on a preset data processing strategy to determine a character recognition result; a document segmentation module, configured to construct a system clause tree based on the character recognition result, and perform document segmentation using the system clause tree to determine a first segmentation result; a fusion matching module, configured to determine, based on the first segmentation result and a preset similarity fusion matching algorithm, reference system information from old system information in a system knowledge base; a conflict detection module, configured to perform conflict detection on the first segmentation result and the reference system information based on a preset large model to determine a conflict detection result corresponding to the system update event; wherein the document segmentation module comprises: an identifier matching unit, configured to identify and match structural identifiers in the character recognition result based on a regular expression template library and a sequence labeling model based on a pre-trained language model to determine an identifier matching result; a hierarchical construction unit, configured to construct parent-child hierarchies based on the identifier matching result to determine a hierarchical construction result; a jump determination unit, configured to determine whether there is an identifier jump in the character recognition result based on the identifier matching result and a preset model to determine a jump determination result; a jump processing unit, configured to, if the jump determination result is yes, determine whether there is a cross-page break or optical character recognition error, and when the corresponding target determination result is yes, trigger a corresponding jump processing operation to construct a system clause tree using the obtained new character recognition result; a reconstruction unit, configured to, if the jump determination result is no, reconstruct a tree-shaped semantic structure based on the identifier matching result and the hierarchical construction result to determine a system clause tree; a continuity determination unit, configured to analyze the semantic association degree between adjacent sentences based on the length information and semantic vectors corresponding to each sentence in the character recognition result to determine the semantic continuity of each sentence; a cut point determination unit, configured to identify a potential segmentation boundary based on the semantic continuity and the character length corresponding to a sub-leaf in the system clause tree to determine a target cut point; a first segmentation unit, configured to perform document segmentation based on the target cut point, the character recognition result, and the system clause tree to determine a first segmentation result; the fusion matching module comprises: a second segmentation unit, configured to perform document segmentation based on old system information in a system knowledge base to determine a second segmentation result; A vector determination unit is configured to determine semantic vector representations corresponding to any two pieces of the institutional information in the first and second pieces of the split results based on a pre-trained language model. A semantic similarity determination unit is configured to determine semantic similarity between the semantic vector representations based on cosine similarity. A transformation unit is configured to perform word vector transformation on each word in the two pieces of the institutional information to determine a transformation result. A weight configuration unit is configured to configure an information entropy weight for each word vector in the transformation result. An entropy weight similarity determination unit is configured to determine entropy weight similarity between the two pieces of the institutional information based on the information entropy weight and the transformation result. A first weighted fusion unit is configured to perform weighted fusion based on the semantic similarity and the entropy weight similarity to determine target similarity corresponding to the two pieces of the institutional information. A sorting unit is configured to sort the target similarity according to the value to determine a sorting result. A reference institutional information determination unit is configured to determine reference institutional information based on the sorting result. The entropy weight similarity determination unit comprises: ； wherein, is the semantic continuity degree for each sentence in the sub-leaf; is a value for measuring the semantic completeness degree of the current sentence; is the target length of division; represents the length of the first sentence in the sub-leaf, j = 1; represents the total length of characters accumulated from the first sentence of the sub-leaf to the current sentence (i); represents the absolute value symbol; and is the weight; The entropy weight similarity determination unit comprises: ； wherein, is the entropy weight similarity between the current two said institutional information segments; denotes semantic consistency; denotes semantic distance penalty; denotes importance weighting, i.e. the information entropy weight; , is the transformation result corresponding to the current two said institutional information segments, and , denotes the word vector.

7. An electronic device, comprising: The computer program is stored in the memory and executed by the processor to implement the conflict detection method of the institutional information of the financial institution. The computer program is stored in the memory and executed by the processor to implement the conflict detection method of the institutional information of the financial institution. ​ 8. A computer-readable storage medium, characterized in that, ​

Images