A text review method based on deep learning

Abstract

The application discloses a kind of text censorship methods based on deep learning, comprising the following steps: S1, obtaining text to be examined and block coding, establish text block vector index;S2, function modeling is carried out to censorship rule, and rule parameter package is generated;S3, rule example text is encoded and example vector index is established;S4, evidence text block is retrieved based on example vector and text block vector, and evidence feature is generated;S5, evidence feature is input into KAN determination network with function cluster parameter, and censorship event flow is generated;S6, the frequency of veto trigger is counted and artificial bee colony reverse search is triggered, and buffer parameter package is generated;S7, according to censorship situation, update function cluster parameter and operating parameter vector, and output censorship result set.The application realizes that text censorship rule can be calculated and determination process is self-adaptively adjusted, and improves censorship stability and continuous operation ability.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a text review method based on deep learning. Background Technology

[0002] Existing text review technologies are mainly applied to scenarios such as internet content governance, information release review, and compliance risk control. They typically achieve automated review through keyword matching, rule engines, or deep learning-based text classification models. Keyword matching and rule engines rely on manually formulated review rules, which are simple to implement but have limited ability to understand contextual semantics, making them prone to missed detections or false positives in cases of complex semantics, implicit expressions, or contextual shifts. To improve semantic recognition capabilities, deep learning-based text review methods have emerged in recent years. These methods train on large amounts of labeled data to automatically identify inappropriate text. However, these methods are highly dependent on the quality of training samples and the consistency of annotations; their stability is insufficient when rules change frequently or sample distributions drift.

[0003] In actual review processes, review rules often exist in the form of clauses, and different rules have priorities and conflicting relationships. Existing deep learning models typically struggle to directly express the rule structure and to finely control the rule judgment process. When the model's judgment results deviate from the rule's decision, existing technologies mostly correct this by adjusting the threshold as a whole or retraining the model, lacking fine-grained analysis of specific rule-triggered behaviors. In scenarios without clear right or wrong labels or in online operation, existing technologies struggle to identify whether the judgment status is abnormal based on the statistical characteristics of review behavior, and also lack adaptive adjustment mechanisms for abnormal judgment behaviors.

[0004] Therefore, how to provide a deep learning-based text censorship method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a text review method based on deep learning. This invention combines review behavior trend analysis with an adaptive parameter adjustment mechanism to achieve dynamic adjustment and stable operation of the text review judgment process. This invention completes rule-level judgment control without the need for correct / incorrect labels, and possesses advantages such as strong interpretability, high operational stability, and strong rule adaptability.

[0006] A text review method based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] S1. Obtain the text to be reviewed, divide it into blocks according to sentence boundaries to obtain a set of text blocks, encode the set of text blocks to obtain a set of text block vectors, and establish a text block vector index.

[0008] S2. Obtain the review rules, input the review rules into the KAN rule function network, and generate a rule parameter package;

[0009] S3. Read the retrieval example text from the rule parameter package, encode it to obtain an example vector set, and establish an example vector index;

[0010] S4. For each rule element, perform similarity retrieval on the text block vector index using the example vector index, select text blocks to form an evidence text block set, and calculate the similarity sequence, hit density and position distribution of the evidence text block set to generate evidence features.

[0011] S5. Input the evidence features and function cluster parameters into the KAN decision network, output the review event flow and record the running parameter vector;

[0012] S6. Count the number of rejection triggers and the number of reviews for the review event flow using a sliding window and calculate the rejection trigger frequency. Compare the rejection trigger frequency with the trigger threshold. When the comparison result meets the trigger condition, construct ABC. Initialize the running parameter vector to the individual positions of the beehive. Perform a reverse diffusion search on the positions corresponding to high rejection trigger frequencies to generate a buffer parameter package.

[0013] S7. The review event flow statistics include the frequency of rejection triggers, the proportion of reviews, and the number of conflicts to form a review status quantity. The review status quantity is mapped to function drift parameters and the function cluster parameters are updated. The buffer parameter package is updated to the running parameter vector, and the KAN judgment network is called to output the text review result set.

[0014] Optionally, S2 specifically includes:

[0015] S21. Obtain the structured representation of the review rule text, input the review rule text into the KAN rule function network, perform sequence parsing and rule unit segmentation on the review rule text to obtain the rule element sequence;

[0016] S22. Perform a functional mapping operation on the rule element sequence, mapping each rule element to a set of one-dimensional function representations, and assigning veto priority parameters to the one-dimensional function representations to generate function cluster parameters.

[0017] S23. Aggregate and bind the parameters of the function cluster to form a regular parameter package.

[0018] Optionally, the allocation of veto priority parameters specifically includes:

[0019] When the ABC algorithm in S6 performs backdiffusion search, for each search iteration, the rule element identifier that causes the current running parameter vector to enter the veto trigger state is recorded.

[0020] Accumulate the count of the rule element identifiers corresponding to the veto trigger state to obtain the failure count value of each rule element in the failed path of the back diffusion search.

[0021] The failure count is mapped to a set of discrete priority numbers, and the failure counts of all rule elements are sorted, with the sorting number corresponding one-to-one with the priority number.

[0022] The priority number is used as the veto priority parameter, written into the one-dimensional function representation of the corresponding rule element, and bound to the function cluster parameter.

[0023] Optionally, S4 specifically includes:

[0024] S41. For each rule element in the rule parameter package, read the set of example vectors corresponding to the rule element in the example vector index, and generate a retrieval query vector sequence according to the storage order of the example vectors;

[0025] S42. For each search query vector in the search query vector sequence, call the text block vector index to perform similarity search, obtain the candidate text block set corresponding to each search query vector, and record the text block identifier and similarity score of the candidate text blocks.

[0026] S43. Perform a merging process on several candidate text block sets corresponding to the same rule element, aggregate the similarity scores according to the text block identifier, generate a text block aggregation score, sort the text block aggregation scores, and select the text blocks in the sorting results to form an evidence text block set.

[0027] S44. For each evidence text block in the evidence text block set, calculate the similarity score between the evidence text block and each retrieval query vector in the retrieval query vector sequence, and generate a similarity sequence according to the order of the retrieval query vector sequence.

[0028] S45. Perform a hit density calculation on the evidence text block set, count the number of times the same evidence text block appears in the candidate text block set, and calculate the hit density by combining the number of occurrences with the length of the retrieval query vector sequence.

[0029] S46. Read the start and end positions of the evidence text block in the text to be examined, calculate the position distribution of the evidence text block in the text to be examined, and combine the similarity sequence, hit density and position distribution to generate evidence features.

[0030] Optionally, the rejection trigger frequency statistics and trigger determination in S6 include the following steps:

[0031] The review event stream is divided into consecutive sliding windows in chronological order, and each sliding window contains a fixed number of review event records;

[0032] Within each sliding window, count the number of veto triggers corresponding to the rule elements marked as veto trigger states, and count the total number of review events within the sliding window;

[0033] The veto trigger frequency corresponding to the current sliding window is obtained by calculating the number of veto triggers and the total number of review events.

[0034] The rejection trigger frequency is compared with a preset trigger threshold. When the rejection trigger frequency reaches or exceeds the trigger threshold, a search trigger flag is generated.

[0035] Optionally, the reverse search process based on the ABC algorithm in S6 specifically includes:

[0036] After generating the search trigger marker, the ABC algorithm is constructed to map the running parameter vector recorded in S5 to the individual positions of the bee colony, thus forming the initial bee colony state;

[0037] During the search process of the Artificial Bee Colony Algorithm (ABC), candidate parameter positions are generated for the position of each individual bee colony, and the candidate parameter positions are substituted into the KAN decision network to perform decision calculations to obtain the corresponding review event stream.

[0038] For each candidate parameter position, the corresponding review event stream calculates the rejection trigger frequency. When the rejection trigger frequency meets the search trigger condition in all sliding windows involving the candidate parameter position, the current candidate parameter position is determined to be an infeasible parameter position.

[0039] Perform a retrospective analysis on the review event flow corresponding to the infeasible parameter position, extract the set of rule element identifiers triggered during the rejection triggering process, and calculate the failure contribution vector of the rule element to the current parameter position based on the set of rule element identifiers.

[0040] In the artificial bee colony algorithm, the reverse direction of the failure contribution vector is used as the search guidance direction, and the position of individual bees in the colony is updated by reverse diffusion.

[0041] The locations of infeasible parameters and their corresponding failure contribution vectors generated during the backdiffusion search are collected to generate a buffer parameter package.

[0042] Optionally, the formation of the review status quantity in S7 specifically includes:

[0043] Within a predetermined statistical period, the review events marked as veto-triggered in the review event stream are statistically analyzed, the number of veto-triggers corresponding to each rule element is calculated, and combined with the total number of review events within the statistical period, a veto-trigger frequency sequence is formed.

[0044] The review events that were manually or by rules reviewed after entering the rejection decision path in the review event flow are marked and statistically analyzed. The ratio between the number of events whose rejection decision results were corrected by review and the number of rejection decision events is calculated to obtain the review ratio.

[0045] The number of conflicts is obtained by statistically analyzing the cases where the same text block or the same evidence feature is simultaneously marked as a veto trigger state by more than one rule element in the review event flow, and calculating the number of times the veto judgment results between rule elements are inconsistent.

[0046] The rejection trigger frequency sequence, review ratio, and conflict number are aligned according to a unified time dimension and combined to form a review status metric.

[0047] Optionally, the update function cluster parameters in S7 specifically include:

[0048] For each rule element, the situation change vector corresponding to the rule element is calculated based on the difference results of the veto trigger frequency sequence between adjacent statistical periods, the difference results of the review ratio, and the difference results of the number of conflicts.

[0049] The situation change vector is merged to obtain the drift control quantity corresponding to the rule element, and the drift control quantity is mapped to the value range of a one-dimensional function parameter associated with the rule element in the operating parameter vector to determine the function parameter sub-range.

[0050] Perform directional update operations on the one-dimensional function parameters within the function parameter sub-interval, perform reduction update operations on the amplitude control parameters in the one-dimensional function parameters, and perform interval offset update operations on the input interval control parameters in the one-dimensional function parameters;

[0051] For one-dimensional function parameters that do not fall within the function parameter subinterval, keep the current parameter value unchanged, and restrict parameter updates to only occur within the function parameter subinterval;

[0052] For the one-dimensional function representations used for similarity sequences, hit densities, and positional distributions under the same rule element, targeted update operations are performed according to the function parameter sub-intervals, and the updated one-dimensional function parameter set is written into the function cluster parameters, and the rule parameter package is updated synchronously.

[0053] Optionally, the buffer parameter packet update process in S7 specifically includes:

[0054] For the set of infeasible parameter locations recorded in the buffer parameter package, the running parameter vector components corresponding to the infeasible parameter locations are clustered according to the rule element identifier to form a rule element-level infeasible parameter subset;

[0055] For each rule element, the value distribution interval of each one-dimensional function parameter in the infeasible parameter subset is statistically analyzed, and the value distribution interval is marked as the infeasible parameter interval;

[0056] When updating the running parameter vector, an avoidance update operation is performed on the one-dimensional function parameters in the current running parameter vector that fall into the infeasible parameter interval. The avoidance update operation is completed by adjusting the current parameter value to outside the boundary of the infeasible parameter interval.

[0057] For one-dimensional function parameters in the current running parameter vector that do not fall into the infeasible parameter interval, keep their current values ​​unchanged;

[0058] The updated running parameter vector after the avoidance is completed will be used as the new running parameter vector for subsequent decision calculations.

[0059] The beneficial effects of this invention are:

[0060] (1) This invention models the review rules as functions and introduces a one-dimensional function representation of the rule elements in the judgment process, so that the review rules can directly participate in the text judgment process in a calculable and adjustable form. This achieves a close integration of rule semantics and model calculation process. Compared with the existing technology that only relies on keyword matching or black box deep models, it improves the precision of rule expression and the controllability of the judgment process.

[0061] (2) In the process of review operation, the present invention introduces a statistical analysis mechanism of review event flow and review status quantity. By continuously monitoring the operation status indicators such as rejection trigger frequency, review ratio and conflict number, the dynamic perception of review judgment status is realized. Compared with the existing technology that relies on static threshold or offline evaluation, the stability of the system under the conditions of rule change and text distribution fluctuation is enhanced.

[0062] (3) The present invention further combines artificial bee colony algorithm to form a buffer parameter package and performs constrained update and avoidance operations at the function parameter level, so that the running parameter vector can be adjusted outside the identified infeasible parameter region. Compared with the overall retraining or global threshold adjustment method, the localization and continuity of parameter update are realized, which improves the long-term operation capability and adaptive adjustment capability of the text review system. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a flowchart of a deep learning-based text review method proposed in this invention;

[0065] Figure 2 This is a reverse search loop graph for a deep learning-based text review method proposed in this invention.

[0066] Figure 3 This is a schematic diagram illustrating the situation-driven update of function cluster parameters in a deep learning-based text review method proposed in this invention. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0068] refer to Figure 1-3 A deep learning-based text review method includes the following steps:

[0069] S1. Obtain the text to be reviewed, divide it into blocks according to sentence boundaries to obtain a set of text blocks, encode the set of text blocks to obtain a set of text block vectors, and establish a text block vector index.

[0070] S2. Obtain the review rules, input the review rules into the KAN rule function network, and generate a rule parameter package;

[0071] S3. Read the retrieval example text from the rule parameter package, encode it to obtain an example vector set, and establish an example vector index;

[0072] S4. For each rule element, perform similarity retrieval on the text block vector index using the example vector index, select text blocks to form an evidence text block set, and calculate the similarity sequence, hit density and position distribution of the evidence text block set to generate evidence features.

[0073] S5. Input the evidence features and function cluster parameters into the KAN decision network, output the review event flow and record the running parameter vector;

[0074] S6. Count the number of rejection triggers and the number of reviews for the review event flow using a sliding window and calculate the rejection trigger frequency. Compare the rejection trigger frequency with the trigger threshold. When the comparison result meets the trigger condition, construct ABC. Initialize the running parameter vector to the individual positions of the beehive. Perform a reverse diffusion search on the positions corresponding to high rejection trigger frequencies to generate a buffer parameter package.

[0075] S7. The review event flow statistics include the frequency of rejection triggers, the proportion of reviews, and the number of conflicts to form a review status quantity. The review status quantity is mapped to function drift parameters and the function cluster parameters are updated. The buffer parameter package is updated to the running parameter vector, and the KAN judgment network is called to output the text review result set.

[0076] In this embodiment, S1 specifically includes:

[0077] S11. Use word segmentation algorithms to divide the text into meaningful basic units, namely words or phrases; for Chinese text, use a dictionary-based word segmentation method; while for Western languages ​​such as English, basic word segmentation can be completed by using spaces as delimiters, and finally obtain word sequences.

[0078] S12. Divide the word sequence into blocks; generate a series of non-overlapping or partially overlapping blocks;

[0079] S13. For each piece of text, use the text2vec-type text embedding model to convert the word sequence into a vector.

[0080] In this embodiment, S2 specifically includes:

[0081] S21. Obtain the structured representation of the review rule text. Input the review rule text into the KAN rule function network to perform sequence parsing and rule unit segmentation on the review rule text to obtain a rule element sequence. Read the original review rule text and process each review rule as an independent rule record. Perform field splitting operation on each rule record to split the rule content into a rule identifier field, a rule constraint description field, and a retrieval example text field. The rule identifier field is used to uniquely identify the rule unit, the rule constraint description field is used to represent the constraint semantics of the rule, and the retrieval example text field is used to store the original text example content corresponding to the rule constraint. Perform sequential processing on the rule constraint description field to convert the rule constraint description field into a text sequence arranged in semantic order. Combine the rule identifier field, rule constraint description field, and retrieval example text field into a unified data structure to generate the structured representation of the review rule text.

[0082] S22. Perform a functional mapping operation on the rule element sequence, mapping each rule element to a set of one-dimensional function representations, and assigning a rejection priority parameter to the one-dimensional function representation to generate function cluster parameters; the functional mapping operation is to perform feature embedding and function parameter generation processing on each rule element in the rule element sequence; the one-dimensional function representation establishes a mapping relationship with the rule element as the index and writes it into the function parameter set corresponding to each rule element.

[0083] S23. Aggregate and bind function cluster parameters to form a rule parameter package. Aggregation and parameter binding refer to performing unified encapsulation and association processing on function cluster parameters, binding function cluster parameters corresponding to the same review rule to rule element identifiers one-to-one, and writing function cluster parameters and veto priority parameters into the same parameter structure according to a fixed field order; after completing the binding, the parameter structures corresponding to multiple rule elements are aggregated according to the rule identifier to form a data set that can be sequentially read and indexed, and the data set constitutes the rule parameter package.

[0084] In this embodiment, the allocation of veto priority parameters specifically includes:

[0085] When the ABC algorithm in S6 performs backdiffusion search, for each search iteration, the rule element identifier that causes the current running parameter vector to enter the veto trigger state is recorded.

[0086] Accumulate the count of the rule element identifiers corresponding to the veto trigger state to obtain the failure count value of each rule element in the failed path of the back diffusion search.

[0087] The failure count is mapped to a set of discrete priority numbers, and the failure counts of all rule elements are sorted, with the sorting number corresponding one-to-one with the priority number.

[0088] The priority number is used as the veto priority parameter, written into the one-dimensional function representation of the corresponding rule element, and bound to the function cluster parameter.

[0089] In this embodiment, S3 specifically includes:

[0090] S31. Perform vector encoding processing on each of the retrieved example text sequences to obtain a set of example vectors that correspond one-to-one with the retrieved example texts, and maintain the correspondence between the example vectors and the rule element identifiers;

[0091] S32. Group and store the example vector set according to the rule element identifier, and construct an example vector index structure that can be accessed by the rule element identifier.

[0092] In this embodiment, S4 specifically includes:

[0093] S41. For each rule element in the rule parameter package, read the set of example vectors corresponding to the rule element in the example vector index, and generate a retrieval query vector sequence according to the storage order of the example vectors;

[0094] S42. For each search query vector in the search query vector sequence, call the text block vector index to perform similarity search, obtain the candidate text block set corresponding to each search query vector, and record the text block identifier and similarity score of the candidate text blocks; the similarity score adopts Euclidean distance;

[0095] S43. Perform a merging process on several candidate text block sets corresponding to the same rule element, aggregate the similarity scores according to the text block identifier, generate a text block aggregate score, sort the text block aggregate scores, and select the text blocks in the sorting results to form an evidence text block set; in this embodiment, the aggregation adopts the method of averaging.

[0096] S44. For each evidence text block in the evidence text block set, calculate the similarity score between the evidence text block and each retrieval query vector in the retrieval query vector sequence, and generate a similarity sequence according to the order of the retrieval query vector sequence.

[0097] S45. Perform a hit density calculation on the evidence text block set, count the number of times the same evidence text block appears in the candidate text block set, and calculate the hit density by comparing the number of occurrences with the length of the retrieval query vector sequence. The hit density is the ratio between the number of times the evidence text block is retrieved in a number of similarity searches and the total number of similarity searches.

[0098] S46. Read the start and end positions of the evidence text blocks in the text to be examined, calculate the positional distribution of the evidence text blocks in the text to be examined, and combine the similarity sequence, hit density, and positional distribution to generate evidence features. The start and end positions are recorded synchronously during the text segmentation stage and when the text block set is established.

[0099] In this embodiment, S5 specifically includes:

[0100] For each rule element, the one-dimensional function representation corresponding to the rule element is invoked, and the function calculation is performed on the similarity sequence, hit density and position distribution in the evidence features respectively, where the similarity sequence is the first input variable, the hit density is the second input variable and the position distribution is the third input variable;

[0101] Substitute each similarity value in the similarity sequence into the corresponding one-dimensional function to calculate the output sequence of the similarity function, and perform a summation calculation on the output sequence of the similarity function to obtain the similarity response value;

[0102] Substitute the hit density into the corresponding one-dimensional function to calculate the hit response value, and substitute the position distribution into the corresponding one-dimensional function to calculate the position response value;

[0103] The similarity response value, hit response value, and location response value are weighted and summed with weight coefficients of 0.5, 0.3, and 0.2, respectively, to generate the judgment response value corresponding to the rule element.

[0104] The decision response value is compared with the rejection decision threshold, wherein the rejection decision threshold is set to 0.65. When the decision response value is greater than or equal to 0.65, the rule element is marked as a rejection trigger state.

[0105] When multiple rule elements corresponding to the same evidence feature are marked as veto triggering state, the veto priority parameters bound to each rule element are read, and the veto decision is executed in the order of priority number corresponding to the veto priority parameters.

[0106] The rule element identifier that leads to the veto decision, the corresponding judgment response value, and the veto trigger status are written into the review event stream.

[0107] In this embodiment, the rejection trigger frequency statistics and trigger determination in S6 include the following steps:

[0108] The review event stream is divided into consecutive sliding windows in chronological order. Each sliding window contains a fixed number of review event records. Each sliding window contains 50 review event records, and there is an overlap of 25 review event records between adjacent sliding windows.

[0109] Within each sliding window, count the number of veto triggers corresponding to the rule elements marked as veto trigger states, and count the total number of review events within the sliding window;

[0110] The veto trigger frequency corresponding to the current sliding window is obtained by calculating the number of veto triggers and the total number of review events.

[0111] The rejection trigger frequency is compared with a preset trigger threshold of 0.4. When the rejection trigger frequency reaches or exceeds the trigger threshold, a search trigger flag is generated.

[0112] In this embodiment, the reverse search process based on the ABC algorithm in S6 specifically includes:

[0113] After generating the search trigger marker, the ABC algorithm is constructed to map the running parameter vector recorded in S5 to the individual positions of the bee colony, thus forming the initial bee colony state;

[0114] During the search process of the Artificial Bee Colony Algorithm (ABC), candidate parameter positions are generated for the position of each individual bee colony, and the candidate parameter positions are substituted into the KAN decision network to perform decision calculations to obtain the corresponding review event stream.

[0115] For each candidate parameter position, the corresponding review event stream calculates the rejection trigger frequency. When the rejection trigger frequency meets the search trigger condition in all sliding windows involving the candidate parameter position, the current candidate parameter position is determined to be an infeasible parameter position.

[0116] A backtracking analysis is performed on the review event stream corresponding to the infeasible parameter position to extract the set of rule element identifiers triggered during the veto triggering process. Based on the set of rule element identifiers, the failure contribution vector of the rule elements to the current parameter position is calculated. For candidate parameter positions determined to be infeasible, the review event stream corresponding to the candidate parameter position is read, and the rule element identifiers marked as veto triggering in the review event stream are counted. Using the rule element identifier as an index, the number of times each rule element enters the veto decision path in the review event stream is accumulated to form a rule element failure count sequence. The rule element failure count sequence is weighted and mapped with the veto priority parameter bound to the corresponding rule element to obtain the failure contribution value of each rule element at the current candidate parameter position. The failure contribution values ​​of each rule element are arranged in order of rule element identifier to form the failure contribution vector corresponding to the current candidate parameter position.

[0117] In the artificial bee colony algorithm, the reverse direction of the failure contribution vector is used as the search guidance direction, and the position of individual bees in the colony is updated by reverse diffusion.

[0118] The locations of infeasible parameters and their corresponding failure contribution vectors generated during the backdiffusion search are collected to generate a buffer parameter package.

[0119] In this embodiment, the formation of the review status quantity in S7 specifically includes:

[0120] Within a predetermined statistical period of 100 review event records, the review events marked as being in a veto trigger state in the review event stream are statistically analyzed, the number of veto triggers corresponding to each rule element is calculated, and combined with the total number of review events within the statistical period, a veto trigger frequency sequence is formed.

[0121] The review events that were manually or by rules reviewed after entering the rejection decision path in the review event flow are marked and statistically analyzed. The ratio between the number of events whose rejection decision results were corrected by review and the number of rejection decision events is calculated to obtain the review ratio.

[0122] The number of conflicts is obtained by statistically analyzing the cases where the same text block or the same evidence feature is simultaneously marked as a veto trigger state by more than one rule element in the review event flow, and calculating the number of times the veto judgment results between rule elements are inconsistent.

[0123] The rejection trigger frequency sequence, review ratio, and conflict number are aligned according to a unified time dimension and combined to form a review status metric.

[0124] In this embodiment, updating the function cluster parameters in S7 specifically includes:

[0125] For each rule element, the situation change vector corresponding to the rule element is calculated based on the difference results of the veto trigger frequency sequence between adjacent statistical periods, the difference results of the review ratio, and the difference results of the number of conflicts.

[0126] Merge the situation change vector to obtain the drift control quantity corresponding to the rule element, and map the drift control quantity to the value range of the one-dimensional function parameter associated with the rule element in the running parameter vector to determine the function parameter sub-interval; divide it continuously according to the parameter index;

[0127] In this embodiment, the merge calculation includes:

[0128] Perform sign consistency judgment on the veto trigger frequency difference result, review ratio difference result, and conflict number difference result in the situation change vector corresponding to the rule element, and classify the change results with the same sign into the same change direction;

[0129] The results of changes in the same direction are summed, and the results of changes with opposite signs are canceled out to obtain the net change value corresponding to the rule element.

[0130] The net change value is converted into a drift control quantity according to a preset mapping relationship. The drift control quantity is used to represent the adjustment level of the current rule element when the function parameter is adjusted in the running parameter vector. Drift control quantities of different magnitude levels are mapped to function parameter value ranges of different lengths. The mapping is based on the range index mapping table, which is established during the system initialization phase.

[0131] Perform directional update operations on the one-dimensional function parameters within the function parameter sub-interval, perform reduction update operations on the amplitude control parameters in the one-dimensional function parameters, and perform interval offset update operations on the input interval control parameters in the one-dimensional function parameters;

[0132] In this implementation, each rule element corresponds to a one-dimensional function representation with amplitude control parameters and input interval control parameters, where the amplitude control parameters serve as the output scaling factor of the one-dimensional function representation. During system initialization, the amplitude control parameters of all one-dimensional function representations are uniformly set to 1. During the function parameter update process, only parameters falling within the function parameter sub-interval are subject to directional updates: the amplitude control parameters are updated by reduction to 0.85, and the input interval control parameters are updated by interval offset of 0.05, with the direction determined by the sign of the offset control amount. The amplitude control parameters and input interval control parameters that do not fall within the function parameter sub-intervals retain their current values. All interval ranges are initialized to (0, 1).

[0133] For one-dimensional function parameters that do not fall within the function parameter subinterval, keep the current parameter value unchanged, and restrict parameter updates to only occur within the function parameter subinterval;

[0134] For the one-dimensional function representations used for similarity sequences, hit densities, and positional distributions under the same rule element, directional update operations are performed according to the function parameter sub-intervals, and the updated one-dimensional function parameter set is written into the function cluster parameters. Synchronous updates are then performed on the rule parameter package. Directional updates maintain the positive and negative directions during the update process.

[0135] In this embodiment, the buffer parameter packet update process in S7 specifically includes:

[0136] For the set of infeasible parameter locations recorded in the buffer parameter package, the running parameter vector components corresponding to the infeasible parameter locations are clustered according to the rule element identifier to form a rule element-level infeasible parameter subset;

[0137] For each rule element, the value distribution interval of each one-dimensional function parameter in the infeasible parameter subset is statistically analyzed, and the value distribution interval is marked as the infeasible parameter interval;

[0138] When updating the running parameter vector, an avoidance update operation is performed on the one-dimensional function parameters in the current running parameter vector that fall into the infeasible parameter interval. The avoidance update operation is completed by adjusting the current parameter value to outside the boundary of the infeasible parameter interval.

[0139] For one-dimensional function parameters in the current running parameter vector that do not fall into the infeasible parameter interval, keep their current values ​​unchanged;

[0140] The updated running parameter vector after the avoidance is completed will be used as the new running parameter vector for subsequent decision calculations.

[0141] Example 1: To verify the feasibility of this invention in practice, it was applied to a typical text compliance review scenario. In this scenario, the text to be reviewed has a complex source, and its content is characterized by semantic ambiguity, diverse expressions, and strong contextual dependencies. Traditional review methods based on keyword matching or a single deep learning model are prone to frequent rule mis-triggering, increased manual review pressure, and fluctuating judgment results during long-term operation. Especially when the number of review rules increases and there are cross-constraints between rules, the stability of the judgment results is difficult to guarantee, and the model parameters are difficult to control effectively through simple threshold adjustments.

[0142] In this application scenario, the deep learning-based text review method described in this invention is deployed. The system first segments the input text into sentence-level segments and vectorizes the text blocks. Simultaneously, existing review rules are transformed into functional one-dimensional representations to form a rule parameter package. Through similarity retrieval between example vectors and text block vectors, the system can locate evidence text blocks semantically related to the rules in the original text and further calculate evidence features such as similarity sequences, hit densities, and positional distributions. These evidence features are input into the KAN decision network, outputting a structured review event stream, thereby achieving rule element-level review judgment.

[0143] During continuous operation, the system performs statistical analysis on the review event flow, calculating the rejection trigger frequency, review ratio, and rule conflict count, and forming a review status metric reflecting the overall review status. When an abnormal concentration of rejection trigger frequency is detected, the system automatically triggers an artificial bee colony algorithm to perform a reverse search, directionally adjusting the operating parameter vector and recording infeasible parameter regions in a buffer parameter package. Subsequently, the system combines the review status metric to perform constrained updates to the function cluster parameters, making fine-grained adjustments to the decision function without disrupting the rule structure, thereby ensuring the continuity and stability of the subsequent text review process.

[0144] During continuous verification, the performance of traditional rule engine methods, conventional deep learning classification methods, and the method of this invention on the same text set was compared. Statistical results show that the method of this invention exhibits more stable characteristics in terms of rejection trigger stability, human review burden, and rule conflict control. As the scale of operation increases, the method of this invention can maintain low fluctuations in false triggers, avoid frequent human intervention, and demonstrate good sustainable operation capability.

[0145] Statistical analysis was performed on the performance of different methods on multiple evaluation indicators, and the relevant comparative data are shown in the table.

[0146] Table 1: Comparison of Review Performance of Different Methods

[0147] Rule Engine Method 0.48 0.21 32.6% 146 0.62 Deep learning classification methods 0.44 0.18 27.4% 121 0.68 Method of the present invention 0.41 0.09 14.8% 53 0.86

[0148] As shown in Table 1, different text review methods exhibit differences in operational stability and consistency of judgment. When using the traditional rule engine method, the average frequency of rejection triggering is relatively high, and the corresponding standard deviation is large, indicating significant fluctuations in rejection triggering behavior under different text input conditions, suggesting insufficient rule triggering stability. The proportion of manual review and the number of rule conflicts are both high, reflecting that rules are prone to cross-interference during actual judgment, requiring substantial manual intervention for correction.

[0149] Conventional deep learning classification methods show improvement over rule-based engines in the aforementioned metrics, but the frequency of rejection triggers still fluctuates, and rule conflicts remain unresolved, requiring external intervention for long-term operation. In contrast, the method of this invention maintains lower levels in both the mean and standard deviation of rejection trigger frequencies, indicating more concentrated and stable judgment behavior; the reduced proportion of manual review and fewer rule conflicts suggest improved coordination among rule elements. The overall judgment stability score further validates the invention's ability to operate continuously and stably in complex text review scenarios.

[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A text review method based on deep learning, characterized in that, Includes the following steps: S1. Obtain the text to be reviewed, divide it into blocks according to sentence boundaries to obtain a set of text blocks, encode the set of text blocks to obtain a set of text block vectors, and establish a text block vector index. S2. Obtain the review rules, input the review rules into the KAN rule function network, and generate a rule parameter package; S3. Read the retrieval example text from the rule parameter package, encode it to obtain an example vector set, and establish an example vector index; S4. For each rule element, perform similarity retrieval on the text block vector index using the example vector index, select text blocks to form an evidence text block set, and calculate the similarity sequence, hit density and position distribution of the evidence text block set to generate evidence features. S5. Input the evidence features and function cluster parameters into the KAN decision network, output the review event flow and record the running parameter vector; S6. Count the number of rejection triggers and the number of reviews for the review event flow using a sliding window and calculate the rejection trigger frequency. Compare the rejection trigger frequency with the trigger threshold. When the comparison result meets the trigger condition, construct the ABC algorithm, initialize the running parameter vector to the individual positions of the bee colony, perform a reverse diffusion search on the positions corresponding to high rejection trigger frequencies, and generate a buffer parameter package. S7. The review event flow statistics include the frequency of rejection triggers, the proportion of reviews, and the number of conflicts to form a review status quantity. The review status quantity is mapped to function drift parameters and the function cluster parameters are updated. The buffer parameter package is updated to the running parameter vector, and the KAN judgment network is called to output the text review result set. The formation of the review status quantity in S7 specifically includes: Within a predetermined statistical period, the review events marked as veto-triggered in the review event stream are statistically analyzed, the number of veto-triggers corresponding to each rule element is calculated, and combined with the total number of review events within the statistical period, a veto-trigger frequency sequence is formed. The review events that were manually or by rules reviewed after entering the rejection decision path in the review event flow are marked and statistically analyzed. The ratio between the number of events whose rejection decision results were corrected by review and the number of rejection decision events is calculated to obtain the review ratio. The number of conflicts is obtained by statistically analyzing the cases where the same text block or the same evidence feature is simultaneously marked as a veto trigger state by more than one rule element in the review event flow, and calculating the number of times the veto judgment results between rule elements are inconsistent. The rejection trigger frequency sequence, review ratio, and conflict number are aligned according to a unified time dimension and combined to form a review status metric.

2. The text review method based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Obtain the structured representation of the review rule text, input the review rule text into the KAN rule function network, perform sequence parsing and rule unit segmentation on the review rule text to obtain the rule element sequence; S22. Perform a functional mapping operation on the rule element sequence, mapping each rule element to a set of one-dimensional function representations, and assigning veto priority parameters to the one-dimensional function representations to generate function cluster parameters. S23. Aggregate and bind the parameters of the function cluster to form a regular parameter package.

3. The text review method based on deep learning according to claim 2, characterized in that, The allocation veto priority parameter specifically includes: When the ABC algorithm in S6 performs backdiffusion search, for each search iteration, the rule element identifier that causes the current running parameter vector to enter the veto trigger state is recorded. Accumulate the count of the rule element identifiers corresponding to the veto trigger state to obtain the failure count value of each rule element in the failed path of the back diffusion search. The failure count is mapped to a set of discrete priority numbers, and the failure counts of all rule elements are sorted, with the sorting number corresponding one-to-one with the priority number. The priority number is used as the veto priority parameter, written into the one-dimensional function representation of the corresponding rule element, and bound to the function cluster parameter.

4. The text review method based on deep learning according to claim 3, characterized in that, S4 specifically includes: S41. For each rule element in the rule parameter package, read the set of example vectors corresponding to the rule element in the example vector index, and generate a retrieval query vector sequence according to the storage order of the example vectors; S42. For each search query vector in the search query vector sequence, call the text block vector index to perform similarity search, obtain the candidate text block set corresponding to each search query vector, and record the text block identifier and similarity score of the candidate text blocks. S43. Perform a merging process on several candidate text block sets corresponding to the same rule element, aggregate the similarity scores according to the text block identifier, generate a text block aggregation score, sort the text block aggregation scores, and select the text blocks in the sorting results to form an evidence text block set. S44. For each evidence text block in the evidence text block set, calculate the similarity score between the evidence text block and each retrieval query vector in the retrieval query vector sequence, and generate a similarity sequence according to the order of the retrieval query vector sequence. S45. Perform a hit density calculation on the evidence text block set, count the number of times the same evidence text block appears in the candidate text block set, and calculate the hit density by combining the number of occurrences with the length of the retrieval query vector sequence. S46. Read the start and end positions of the evidence text block in the text to be examined, calculate the position distribution of the evidence text block in the text to be examined, and combine the similarity sequence, hit density and position distribution to generate evidence features.

5. A text review method based on deep learning according to claim 4, characterized in that, The rejection trigger frequency statistics and trigger determination in S6 include the following steps: The review event stream is divided into consecutive sliding windows in chronological order, and each sliding window contains a fixed number of review event records; Within each sliding window, count the number of veto triggers corresponding to the rule elements marked as veto trigger states, and count the total number of review events within the sliding window; The veto trigger frequency corresponding to the current sliding window is obtained by calculating the number of veto triggers and the total number of review events. The rejection trigger frequency is compared with a preset trigger threshold. When the rejection trigger frequency reaches or exceeds the trigger threshold, a search trigger flag is generated.

6. The text review method based on deep learning according to claim 5, characterized in that, The reverse search process based on the ABC algorithm in S6 specifically includes: After generating the search trigger marker, the ABC algorithm is constructed to map the running parameter vector recorded in S5 to the individual positions of the bee colony, thus forming the initial bee colony state; During the search process of the Artificial Bee Colony Algorithm (ABC), candidate parameter positions are generated for the position of each individual bee colony, and the candidate parameter positions are substituted into the KAN decision network to perform decision calculations to obtain the corresponding review event stream. For each candidate parameter position, the corresponding review event stream calculates the rejection trigger frequency. When the rejection trigger frequency meets the search trigger condition in all sliding windows involving the candidate parameter position, the current candidate parameter position is determined to be an infeasible parameter position. A retrospective analysis is performed on the review event flow corresponding to the infeasible parameter position to extract the set of rule element identifiers triggered during the rejection triggering process, and the failure contribution vector of the rule element to the current parameter position is calculated based on the set of rule element identifiers. In the artificial bee colony algorithm, the reverse direction of the failure contribution vector is used as the search guidance direction, and the position of individual bees in the colony is updated by reverse diffusion. The locations of infeasible parameters and their corresponding failure contribution vectors generated during the backdiffusion search are collected to generate a buffer parameter package.

7. A text review method based on deep learning according to claim 6, characterized in that, The update function cluster parameters in S7 specifically include: For each rule element, the situation change vector corresponding to the rule element is calculated based on the difference results of the veto trigger frequency sequence between adjacent statistical periods, the difference results of the review ratio, and the difference results of the number of conflicts. The situation change vector is merged to obtain the drift control quantity corresponding to the rule element, and the drift control quantity is mapped to the value range of a one-dimensional function parameter associated with the rule element in the operating parameter vector to determine the function parameter sub-range. Perform directional update operations on the one-dimensional function parameters within the function parameter sub-interval, perform reduction update operations on the amplitude control parameters in the one-dimensional function parameters, and perform interval offset update operations on the input interval control parameters in the one-dimensional function parameters; For one-dimensional function parameters that do not fall within the function parameter subinterval, keep the current parameter value unchanged, and restrict parameter updates to only occur within the function parameter subinterval; For the one-dimensional function representations used for similarity sequences, hit densities, and positional distributions under the same rule element, targeted update operations are performed according to the function parameter sub-intervals, and the updated one-dimensional function parameter set is written into the function cluster parameters, and the rule parameter package is updated synchronously.

8. A text review method based on deep learning according to claim 7, characterized in that, The buffer parameter packet update process in S7 specifically includes: For the set of infeasible parameter locations recorded in the buffer parameter package, the running parameter vector components corresponding to the infeasible parameter locations are clustered according to the rule element identifier to form a rule element-level infeasible parameter subset; For each rule element, the value distribution interval of each one-dimensional function parameter in the infeasible parameter subset is statistically analyzed, and the value distribution interval is marked as the infeasible parameter interval; When updating the running parameter vector, an avoidance update operation is performed on the one-dimensional function parameters in the current running parameter vector that fall into the infeasible parameter interval. The avoidance update operation is completed by adjusting the current parameter value to outside the boundary of the infeasible parameter interval. For one-dimensional function parameters in the current running parameter vector that do not fall into the infeasible parameter interval, keep their current values ​​unchanged; The updated running parameter vector after the avoidance is completed will be used as the new running parameter vector for subsequent decision calculations.

Images