A method, server, product, and medium for monitoring and identifying health network information.

Abstract

A method, server, product, and medium for monitoring and identifying health and hygiene network information are disclosed, relating to the field of natural language processing. The method includes: acquiring long text information from a health and hygiene network and segmenting it into logical block sequences; extracting the actual semantic feature vector of each block; inputting the sequence into an autoregressive prediction network to predict the predicted semantic feature vector of each block based on prior features; calculating a sequence of deviation metrics between the actual and predicted semantic feature vectors; determining attention weight coefficients based on the deviation metric sequence; weighting and summing the actual semantic feature vectors of each block to obtain a focal semantic feature vector representing local logical mutation information; classifying the long text information from the health and hygiene network based on the focal semantic feature vectors; and outputting compliance identification results. This application enables logical mutation blocks to dominate the synthesis of focal features, avoiding the dilution of non-compliant features by compliant content, and effectively improving the accuracy of identifying logically misleading non-compliant long text information from the health and hygiene network.

Description

Technical Field

[0001] This application relates to the field of natural language processing, and particularly to a method for monitoring and identifying health network information, a server, a product, and a medium. Background Art

[0002] With the development of the health industry, the information about hospitals, doctors, nurses, patients, and medical policies on the Internet is increasing day by day. At present, the health network information monitoring system mainly focuses on the collection of obvious medical disputes, ward environment, illegal charges and other information and keyword alerts. However, in health network information, there are still a large number of hidden "nine truths and one false" type illegal marketing soft texts, such as implanting illegal advertisements at the end of normal sharing of medical experience or doctor popular science. The traditional keyword frequency-based technology is easily overwhelmed by a large number of correct medical terms in the text, resulting in the classifier misjudging it as high-quality compliant popular science, and there is a defect of low recognition accuracy for hidden illegal texts.

[0003] To solve the above defects, related technologies usually adopt a global semantic vector classification method based on a pre-trained language model. This technology truncates the long text according to the model input length limit or slices it with a sliding window and then inputs it into the pre-trained model, uses the self-attention mechanism to capture the deep semantic associations of the context, and extracts the global semantic features representing the whole article (such as the global feature vector after pooling), and then comprehensively classifies and evaluates the core intention and overall tendency of the whole article through a fully connected layer to identify whether the text contains illegal marketing intentions.

[0004] However, in actual application scenarios such as long health popular science articles or long posts in medical forums, since most of the text is an objective and correct discussion of medical common sense, and the core intention only deviates at local positions, when related technical solutions process such long texts, their attention mechanism will mix and aggregate the features of long-length correct medical knowledge with the features of a small number of illegal product recommendations, making the global semantic features dominated by the features of a large amount of compliant content, making it difficult for related technical solutions to accurately capture the logical breaks or causal jumps that occur locally, thus reducing the recognition accuracy of logical misleading illegal health network long text information. Summary of the Invention

[0005] This application provides a method for monitoring and identifying health network information, a server, a product, and a medium, which is used to improve the recognition accuracy of logical misleading illegal health network long text information.

[0006] Firstly, this application provides a method for monitoring and identifying health and wellness network information, applied to an audit server. The method includes: acquiring long text information of the health and wellness network to be detected; dividing the long text information of the health and wellness network into multiple logical blocks according to preset semantic integrity rules to obtain a sequence of logical blocks arranged in text order; the preset semantic integrity rules include at least one of paragraph boundary recognition, sentence semantic dependency analysis, or topic conversion detection; extracting the actual semantic feature vectors of each logical block in the logical block sequence to obtain an actual semantic feature sequence arranged in text order; inputting the actual semantic feature sequence into a preset autoregressive prediction network, and, for each target logical block in the logical block sequence, performing a prediction based on the target logical block... The previous actual semantic feature vectors predict the semantic evolution direction of the target logical block, generating the predicted semantic feature vector corresponding to the target logical block; the deviation metric between the actual semantic feature vector and the predicted semantic feature vector of the target logical block is calculated, and the deviation metric sequence corresponding to each logical block is obtained by traversing the logical block sequence; based on the deviation metric sequence, the attention weight coefficient corresponding to each logical block is determined, and the actual semantic feature vectors corresponding to each logical block are weighted and summed to obtain the focal semantic feature vector, which is used to represent local logical mutation information; based on the focal semantic feature vector, the long text information of the health network is classified, and the compliance recognition result of the long text information of the health network is output.

[0007] By employing the above technical solution, the review server first divides the long text information of the health and wellness network into logical block sequences and generates predicted semantic feature vectors corresponding to each actual semantic feature vector. When a logical block experiences a sudden change in intent to violate regulations, the deviation metric formed by the actual semantic feature vector deviating from the predicted semantic feature vector will surge, creating an identifiable anomalous signal. Next, the review server determines the attention weight coefficients corresponding to each logical block based on the deviation metric sequence, thereby generating a focal semantic feature vector for each logical block. This ensures that logical blocks with larger deviation metric values ​​have higher weights in the final representation, allowing the focal semantic feature vectors to concentrate on representing local logical changes rather than being diluted by a large amount of compliant content features. Finally, the review server completes compliance classification based on the focal semantic feature vectors, solving the problem of global semantic features being overwhelmed by a large amount of compliant content and making it difficult to identify hidden local changes in intent to violate regulations, thus improving the accuracy of identifying logically misleading violations in long text information of the health and wellness network.

[0008] In conjunction with some embodiments of the first aspect, in some embodiments, after calculating the deviation metric between the actual semantic feature vector and the predicted semantic feature vector of the target logical block and traversing the logical block sequence to obtain the deviation metric sequence corresponding to each logical block, the method further includes: performing medical logic rationality verification on candidate mutation blocks in the deviation metric sequence whose deviation metric is greater than or equal to a preset mutation threshold; adjusting the deviation metric corresponding to the candidate mutation blocks according to the result of the medical logic rationality verification to obtain a corrected deviation metric sequence; and correspondingly, determining the attention weight coefficient corresponding to each logical block according to the deviation metric sequence, including: determining the attention weight coefficient corresponding to each logical block according to the corrected deviation metric sequence.

[0009] By adopting the above technical solution, since medical texts contain normal compliant medical stage transitions, their semantic topic shifts will also generate a certain deviation metric. If attention weight coefficients are directly assigned based on the original deviation metric, compliant semantic transition blocks will receive unduely high weights, interfering with the focused semantic feature vector's concentrated representation of real non-compliant mutation information. The review server reduces the interference of compliant semantic transitions on the synthesis of focused semantic feature vectors by introducing medical logic rationality checks and correcting the deviation metric accordingly, thereby improving the targeting of attention weight coefficients to the location of real non-compliant mutations.

[0010] In conjunction with some embodiments of the first aspect, in some embodiments, medical logic rationality verification is performed on candidate mutation blocks in the deviation metric sequence whose deviation metric value is greater than or equal to a preset mutation threshold. Specifically, this includes: extracting the current set of medical entities within the candidate mutation block, and the set of preceding medical entities within the preceding logical block adjacent to the candidate mutation block in the logical block sequence; mapping the current set of medical entities and the set of preceding medical entities to a pre-built medical association knowledge graph, and obtaining the graph connection path between the current set of medical entities and the set of preceding medical entities in the medical association knowledge graph; and calculating the logical coherence confidence of the transfer from the preceding logical block to the candidate mutation block based on the shortest hop count of the graph connection path and the medical association probability of each node on the path.

[0011] By adopting the above technical solution, the review server compares the semantic deviation phenomenon of candidate mutation blocks with structured medical knowledge, and uses logical coherence confidence as a quantitative indicator to distinguish between compliant medical conversions and intentional mutations lacking medical basis. This avoids misjudgments caused by purely relying on semantic vector distance to judge the nature of mutations, and provides an objective basis based on medical knowledge for adjusting the deviation metric, thereby improving the rationality and accuracy of the deviation metric adjustment.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, based on the results of medical logic rationality verification, the deviation metric value corresponding to the candidate mutation block is adjusted to obtain a corrected deviation metric value sequence. Specifically, this includes: if the logical coherence confidence is less than a reasonable transition threshold, the candidate mutation block is determined not to belong to a compliant medical stage transition, and a false marketing risk value is calculated; the deviation metric value of the candidate mutation block is increased based on the false marketing risk value to obtain a corrected deviation metric value sequence; if the logical coherence confidence is greater than or equal to a preset reasonable transition threshold, the candidate mutation block is determined to belong to a compliant medical stage transition, and a decay smoothing coefficient is calculated based on the logical coherence confidence; the deviation metric value of the candidate mutation block is decreased using the decay smoothing coefficient to obtain a corrected deviation metric value sequence.

[0013] By employing the above technical solution, the review server performs bidirectional adjustments to the bias metric of candidate mutation blocks based on logical coherence confidence. When a candidate mutation block does not belong to a compliant medical stage transition, the review server increases the bias metric of the candidate mutation block, further highlighting the abnormality of the illegal mutation location in the corrected bias metric sequence. When a candidate mutation block belongs to a compliant medical stage transition, the review server decreases the bias metric of the candidate mutation block, suppressing the proportion of compliant transition blocks in subsequent attention weight calculations. Through bidirectional adjustments, the review server simultaneously reduces the probability of compliant medical transitions being misjudged as illegal mutations and the probability of real illegal mutations being missed due to the dilution of attention weights, thereby improving the effectiveness of the focus semantic feature vector in representing the concentrated illegal content.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, if the logical coherence confidence level is less than a reasonable transition threshold, the candidate mutation block is determined not to belong to a compliant medical stage transition, and a false marketing risk value is calculated. Specifically, this includes: if the logical coherence confidence level is less than a reasonable transition threshold, identifying unregistered entities in the current medical entity set within the medical association knowledge graph; extracting the affix features and contextual modification features of the unregistered entities, and calculating the false marketing risk value of the unregistered entities based on a preset medical marketing risk classification model.

[0015] By adopting the above technical solution, after determining that the logical coherence confidence of a candidate mutation block is less than a reasonable transition threshold, the review server further identifies unregistered entities in the current medical entity set whose corresponding nodes cannot be found in the medical association knowledge graph. Since illegal medical articles often circumvent knowledge base retrieval by fabricating or exaggerating medical terminology, unregistered entities are potential carriers of violation signals that cannot be directly perceived by knowledge graph connectivity analysis. The review server extracts the affix features and contextual modification features of unregistered entities, inputs them into a preset medical marketing risk classification model, and calculates the false marketing risk value of the unregistered entities. By combining semantic-level logical incoherence signals with lexical-level suspicious terminology signals, the false marketing risk value simultaneously incorporates evidence of both medical logical incoherence and fabricated terminology, enhancing the accuracy of identifying candidate mutation blocks that use unregistered terms to commit covert violations.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, the deviation metric value of the candidate mutation block is adjusted upward based on the false marketing risk value to obtain a corrected deviation metric value sequence. Specifically, this includes: if the false marketing risk value is greater than a preset risk threshold, determining that the candidate mutation block belongs to an illegal mutation guided by forged terms; calculating a risk amplification coefficient based on the false marketing risk value, and adjusting the deviation metric value of the candidate mutation block upward based on the risk amplification coefficient to obtain a corrected deviation metric value sequence.

[0017] By adopting the above technical solution, since the risk amplification coefficient monotonically increases with the risk value of false marketing, the upward adjustment of the deviation metric value is positively correlated with the degree of violation of the forged terms. Candidate mutation blocks with higher violation degrees become more prominent in the corrected deviation metric value sequence, thus gaining higher weight in subsequent attention weight calculations. This allows the focus semantic feature vector to form a more concentrated representation of the illegal mutation content guided by forged terms. In summary, when the review server performs compliance classification based on the focus semantic feature vector, it can improve the accuracy of compliance identification of long-text health and hygiene network information with concealed illegal intentions.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, the actual semantic feature vectors of each logical block in the logical block sequence are extracted to obtain the actual semantic feature sequence arranged in text order. Specifically, this includes: inputting each logical block into a pre-trained semantic encoder to obtain the contextual semantic representation of each logical block; and performing pooling processing on the contextual semantic representation of each logical block to obtain the actual semantic feature vector corresponding to each logical block.

[0019] By adopting the above technical solution, the review server performs pooling processing on the contextual semantic representation, compressing the variable-length word-level vector sequence into a fixed-dimensional actual semantic feature vector, so that the entire actual semantic feature sequence is consistent in feature dimension, thereby improving the accuracy of the deviation metric in quantifying logical mutations.

[0020] Secondly, this application provides an audit server, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the audit server to perform the method described in the first aspect and any possible implementation thereof.

[0021] Thirdly, this application provides a computer-readable storage medium storing computer instructions that, when executed on an audit server, cause the audit server to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, this application provides a computer program product, including a computer program or instructions that, when run on an audit server, cause the audit server to perform the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the review server provided in the second aspect, the computer-readable storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the methods provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0025] 1. By employing a technical solution that divides long text information from the health and wellness network into logical block sequences according to semantic integrity rules, and generates predicted semantic feature vectors for each target logical block based on the preceding actual semantic feature vectors through an autoregressive prediction network, and uses a bias metric sequence to drive the allocation of attention weight coefficients and weighted summation to generate focal semantic feature vectors, logically aberrant blocks with larger bias metric values ​​will dominate the synthesis of focal semantic feature vectors with higher weights. This allows them to centrally represent local logically aberrant information rather than being diluted by a large number of compliant content features. This effectively solves the defects in related technologies where global semantic features are dominated by correct medical knowledge content and it is difficult to perceive local illegal intent aberrations. Thus, it achieves compliance identification of long text information from the health and wellness network that violates hidden logical misleading rules.

[0026] 2. By employing a technical solution that performs medical logic rationality verification on candidate mutation blocks whose deviation metric values ​​are greater than or equal to a preset mutation threshold, and adjusts the deviation metric values ​​of the candidate mutation blocks according to the verification results to obtain a corrected deviation metric value sequence, and then determines the attention weight coefficient of each logic block based on the corrected deviation metric value sequence, the corrected deviation metric value sequence can distinguish between semantic deviations generated by compliant medical stage transitions and mutations with real illegal intent, suppress the interference of compliant transition blocks on attention weight allocation, effectively solve the problem in related technologies where normal medical topic transitions are misjudged as illegal mutations, thereby reducing the accuracy of the focus semantic feature vector representation, and thus realize the rationality identification and attention focus of candidate mutation blocks.

[0027] 3. By employing a bidirectional adjustment of the deviation metric value of candidate mutation blocks based on the comparison results of logical coherence confidence and reasonable transition threshold, the deviation metric value is adjusted upward based on the false marketing risk value when the confidence is below the threshold, and downward based on the decay smoothing coefficient when the confidence is above the threshold. Therefore, the corrected deviation metric value sequence can simultaneously highlight the abnormality of the real illegal mutation blocks and suppress the weight ratio of the compliant medical conversion blocks. This effectively solves the problem of compliant medical conversion being misjudged and real illegal mutations being missed due to weight dilution in related technologies, thereby achieving a concentrated and accurate representation of illegal content by the focus semantic feature vector. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of a system architecture for a method for identifying violations of long text information in a health network, as described in this application.

[0029] Figure 2 This is a flowchart illustrating a method for monitoring and identifying health network information in an embodiment of this application;

[0030] Figure 3 This is another flowchart illustrating a method for monitoring and identifying health network information in this application embodiment;

[0031] Figure 4 This is a schematic diagram of the physical device structure of an audit server in an embodiment of this application. Detailed Implementation

[0032] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0033] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0034] This application provides a method for monitoring and identifying health and wellness network information, which will be described below in conjunction with... Figure 1 This section introduces application scenarios for embodiments of this application. Please refer to [link / reference]. Figure 1 This is a schematic diagram of a system architecture for a method of identifying violations of long text information on a health and hygiene network, as described in this application. It includes the interaction between a data terminal and an audit server, as well as the functional modules within the audit server. The audit server includes:

[0035] The text segmentation module is used to segment long text information from the health network into logical block sequences;

[0036] The feature extraction module is used to extract the actual semantic feature vectors of each logical block;

[0037] The autoregressive prediction module includes a preset autoregressive prediction network used to generate predictive semantic feature vectors.

[0038] The deviation calculation module is used to calculate the deviation metric between the actual and predicted semantic feature vectors;

[0039] The weighted summation and aggregation module is used to determine attention weights based on the deviation metric and aggregate the focus semantic feature vectors.

[0040] The classification module is used for compliance classification based on the focal semantic feature vector.

[0041] like Figure 1As shown, the data terminal sends the long text information of the health and hygiene network to be detected to the review server. After receiving the text, the review server first segments it into logical block sequences based on semantic integrity rules using the text segmentation module, and then uses the feature extraction module to extract the actual semantic feature vector sequence of each logical block. Next, this sequence is input into the autoregressive prediction network to generate the corresponding predicted semantic feature vector sequence. The bias calculation module compares the two sequences and calculates the bias metric sequence. Based on this bias metric sequence, attention weights are calculated, and the weighted summation aggregation module performs a weighted summation of the actual semantic feature vectors to generate a focal semantic feature vector representing local logical mutation information. Finally, the classification module outputs the compliance identification result based on the focal semantic feature vector.

[0042] In related technologies, a comprehensive classification and evaluation of the overall tendency of long text information on health networks can be achieved by inputting the entire article into a pre-trained model to extract global semantic features.

[0043] The autoregressive prediction and bias measurement attention weighting mechanism in this application's embodiments calculates the bias by comparing the actual semantic features with the expected evolution features, and uses this to guide the aggregation of focal semantic features, thereby capturing local illegal mutation signals. This not only avoids excessive interference from lengthy, compliant, and correct medical knowledge on features, but also preserves abnormal features with broken contextual logic.

[0044] As can be seen, the focus semantic feature vector construction scheme in this application embodiment can not only realize the automated compliance review of long text information on the health and wellness network, but also effectively solve the defect in related technologies where global features are easily dominated by a large amount of compliant popular science content and it is difficult to capture local intention shifts. Thus, it can achieve a high accuracy rate in identifying long text information on the health and wellness network that is illegal due to hidden logic misleading, such as "nine truths and one falsehood".

[0045] The following describes the process of the method provided in this implementation. Please refer to [link / reference]. Figure 2 This is a flowchart illustrating a method for monitoring and identifying health network information in an embodiment of this application.

[0046] S201. Obtain the long text information of the health network to be detected, and divide the long text information of the health network into multiple logical blocks according to the preset semantic integrity rules to obtain a logical block sequence arranged in text order.

[0047] Among them, the long-form health and wellness texts to be detected are text records involving medical behavior or medical science popularization collected from health and wellness-related internet platforms (such as medical forums, online consultation platforms, social media, etc.), such as records of hospital, doctor, nurse, and patient medical experiences, health science popularization articles, or long online consultation replies interspersed in descriptions of medical disputes; the preset semantic integrity rules refer to the pre-configured segmentation criteria used to ensure semantic coherence and self-consistency within each logical block after segmentation, aiming to prevent abrupt truncation that would damage key causal chains or thematic coherence, including at least one of paragraph boundary recognition, sentence semantic dependency analysis, or topic transformation detection; a logical block refers to the smallest semantic processing unit obtained after the text is segmented according to the above rules, corresponding to a semantically complete section of content in the original text.

[0048] The review server executes this step upon receiving an externally submitted medical text compliance review request. Specifically, the review server first parses and obtains the long text information of the health and wellness network to be reviewed from the request. Then, it performs hierarchical segmentation based on preset semantic integrity rules: paragraph boundary recognition is used first for coarse-grained initial segmentation; then, semantic dependency analysis is performed on sentences within each coarse-grained segment, merging sentences with strong causal relationships or referential relationships into the same logical segment to avoid cutting off the core logical chain; finally, topic conversion detection is used to verify the boundaries of the logical segments, correcting the segmentation points where there is a significant shift in semantic topic. The review server organizes all segmentation results according to the original text's textual order, outputting a logical segmentation sequence consisting of n logical segments. It fully preserves the semantic flow information of the text.

[0049] It should be noted that for the long text information collected from the health and wellness network, the health and wellness network information monitoring system first monitors and diverts explicit information such as medical disputes, ward environment complaints, and reports of illegal charges through conventional classification models or keyword matching. The review server, as the deep intelligent review engine of the health and wellness network information monitoring system, is used to further review complex long text information that contains medical terminology, appears to be compliant (such as under the guise of doctors' popular science or disease analysis), but actually undergoes local logical mutations to carry out covert illegal marketing.

[0050] S202. Extract the actual semantic feature vectors of each logical block in the logical block sequence to obtain the actual semantic feature sequence arranged in the text order.

[0051] The actual semantic feature vector refers to the fixed-dimensional dense vector output after the individual logical blocks are independently encoded by the pre-trained language model. It is used to represent the deep semantic information carried by the logical block, including the topic orientation, expression intention and content polarity of the medical knowledge statement.

[0052] After constructing the logical block sequence, the review server independently extracts semantic features from each logical block. Specifically, the server feeds each logical block into a pre-trained language model (such as a BERT or RoBERTa model fine-tuned for the medical field). Each logical block is treated as a complete encoding input, and the model's self-attention mechanism is used to model the semantic relationships within the logical block's context. The actual semantic feature vector of that logical block is then extracted from the model's output layer. The review server performs the above extraction operation sequentially on all n logical blocks in the logical block sequence, and then summarizes and arranges the obtained actual semantic feature vectors strictly according to the original text's order, ultimately obtaining the actual semantic feature sequence. .

[0053] S203. Input the actual semantic feature sequence into the preset autoregressive prediction network. For each target logical block in the logical block sequence, predict the semantic evolution direction of the target logical block based on the actual semantic feature vectors before the target logical block, and generate the predicted semantic feature vector corresponding to the target logical block.

[0054] Here, the pre-defined autoregressive prediction network refers to a sequence neural network that strictly adheres to unidirectional causal constraints. When making semantic predictions for the i-th target logical block, it relies only on the actual semantic feature vectors of the first to (i-1)th logical blocks, without using information from the i-th step and subsequent steps, to ensure the unidirectional causality of the prediction process. It can be implemented using structures such as unidirectional LSTM, GRU, or causal mask Transformer decoders. The semantic evolution direction refers to the contextual semantic trajectory established based on the preceding logical blocks of the article, reflecting the thematic direction, knowledge statement logic, and overall intent trend formed before the target logical block. The predicted semantic feature vector refers to the expected vector representation of the semantic content of the target logical block under the assumption of "normal semantic evolution," based on the actual semantic features of all blocks preceding the target logical block, denoted as . .

[0055] After obtaining the actual semantic feature sequence, the review server inputs the actual semantic feature sequence into a preset autoregressive prediction network, and sequentially performs semantic evolution prediction on each target logical block in the logical block sequence. Specifically, the review server inputs the actual semantic feature sequence... Input the pre-defined autoregressive prediction network step by step according to the time step sequence, and divide the i-th target logic block. The network is based on the data received in the previous time step. The accumulated hidden states (i.e., the network's compressed memory representation of the semantic evolution pattern of the first half of the article) are used to output the predicted semantic feature vector of the target logical block at step i. This vector represents the model's reasonable expectation of the semantic content that the i-th logical block should normally have, under the prior semantic path of continuously providing compliant and objective medical science popularization in the preceding content of the article.

[0056] The training process of the pre-defined autoregressive prediction network includes:

[0057] A massive dataset of long-text information from both compliant and non-compliant health and hygiene networks is acquired as a training sample library. Each long-text information is segmented into training logical block sequences according to the same semantic integrity rules as in the inference phase. A trained semantic encoder is used to extract the actual semantic feature vectors of each training logical block sequence, constructing a training actual semantic feature sequence. This training actual semantic feature sequence is then progressively input into a pre-defined autoregressive prediction network. At each time step, the pre-defined autoregressive prediction network outputs the predicted semantic feature vector corresponding to the current time step based on all actual semantic feature vectors prior to the current time step. The mean squared error loss or cosine similarity loss between the predicted semantic feature vector at the current time step and the actual semantic feature vector at that time step is calculated. The loss function values ​​from all time steps are summed to obtain the overall loss of the prediction network. The network parameters of the pre-defined autoregressive prediction network are updated using a backpropagation algorithm until the overall loss converges.

[0058] The training sample library should contain long-text information from health networks across different medical specialties to ensure that the autoregressive prediction network can learn general semantic evolution patterns across domains. During training, dedicated autoregressive prediction networks can be trained separately for different medical specialties, or a multi-task learning framework can be used to introduce departmental labels as auxiliary inputs into a single network.

[0059] Through the above training process, the pre-set autoregressive prediction network can learn the normal evolution rules of contextual semantics in long text information of compliant health and hygiene networks, thereby accurately predicting the expected semantic direction of the target logical block in the compliant context in practical applications.

[0060] S204. Calculate the deviation metric between the actual semantic feature vector and the predicted semantic feature vector of the target logical block, and traverse the logical block sequence to obtain the deviation metric sequence corresponding to each logical block.

[0061] Among them, the deviation metric refers to the actual semantic feature vector of the target logical block. With predicted semantic feature vector The quantitative assessment value of the degree of semantic difference between them is denoted as: This reflects the degree of deviation between the actual semantic content of the target logical block and the reasonable semantic expectation based on the logical inference from the preceding text.

[0062] After generating the predicted semantic feature vectors for each target logical block, the review server proceeds to this step, traversing the logical block sequence and calculating the deviation metric for each logical block. Specifically, the review server performs a process for each logical block in the logical block sequence. Take its actual semantic feature vector With the corresponding predicted semantic feature vector Calculate the deviation measure between the two. Taking cosine distance as an example, ,Right now The larger the value, the more significant the semantic difference between the two vectors.

[0063] For the logical blocks in the beginning of the article that continuously present correct medical knowledge, their actual semantics highly match the semantic expectations of the pre-set autoregressive prediction network. The difference is smaller; however, for logical blocks that suddenly shift to recommending illegal products or making false claims about therapeutic effects, the intention changes significantly. A significant jump will occur, forming a recognizable numerical anomaly peak.

[0064] After the audit server traverses all logical blocks, it summarizes the deviation metric values ​​of each block in text order to obtain a sequence of deviation metric values. .

[0065] S205. Based on the deviation metric sequence, determine the attention weight coefficients corresponding to each logical block, and perform a weighted summation of the actual semantic feature vectors corresponding to each logical block to obtain the focal semantic feature vector. The focal semantic feature vector is used to characterize local logical mutation information.

[0066] The attention weight coefficient refers to the normalized weight value calculated based on the deviation metric sequence, used to measure the contribution of each logical block in the synthesis of the focal semantic feature vector, and is denoted as . The focal semantic feature vector, denoted as F, is an aggregated vector obtained by weighting and summing the actual semantic feature vectors of each logical block using attention weight coefficients. It is used to characterize the local logical mutation information in the entire long text of health and wellness network information. Compared with the global semantic feature vector, which is obtained by directly averaging the features of all blocks, the focal semantic feature vector can more prominently reflect the semantic features of key positions in the text where intent shifts. Local logical mutation information refers to the semantic mutation phenomenon that occurs in a certain local position of the long text of health and wellness network information and is inconsistent with the logical context of the preceding text. It is typically manifested as the article continuously presenting compliant medical science popularization in the beginning, while suddenly inserting content such as recommendations for unqualified health products or false claims of special effects treatments in a certain local position, which deviates from the intended meaning.

[0067] After obtaining the bias metric sequence, the review server calculates the attention weight coefficients and aggregates the focus semantic feature vectors. Specifically, the review server first applies a Softmax function or normalization function to the bias metric sequence, transforming the bias metric values ​​of each block into an attention weight coefficient sequence that satisfies the probability distribution constraint (the sum of all attention weight coefficients is 1). This results in higher attention weight coefficients for logical blocks with larger deviation metrics. Subsequently, the review server uses the attention weight coefficient sequence to perform a weighted summation of the actual semantic feature vectors of each logical block in the actual semantic feature sequence, i.e. We obtain the focal semantic feature vector F.

[0068] This aggregation process allows the focal semantic feature vector to retain the overall semantic background of the text while incorporating the actual semantic feature vectors of logically abrupt regions into the final representation with greater weight, thereby enabling subsequent classifiers to focus on the illegal intent offset signal in the text.

[0069] Optionally, in some embodiments, the review server can also extract layout mutation feature vectors for each logical block in the original text, such as the frequency of bolding, underline density, and sudden increases in the frequency of external links or contact information; and fuse the layout mutation feature vectors with semantic feature vectors obtained by weighted summation based on attention weight coefficients to jointly construct the focus semantic feature vector. Since long text information on violating health and hygiene networks is often accompanied by strong visual guidance layout changes at local intent mutation points, cross-fusion of layout mutation signals and semantic mutation signals can further enhance the three-dimensionality and accuracy of the focus semantic feature vector in representing the violation intent.

[0070] S206. Classify long text information of health and wellness networks based on focus semantic feature vectors, and output the compliance identification results of long text information of health and wellness networks.

[0071] Among them, the compliance identification result refers to the final judgment of the audit server on whether the long text information of health and health network to be tested violates the compliance requirements of health and health network information. It is usually output in the form of a combination of classification labels and corresponding confidence scores for reference by the upper-level audit management system or human auditors.

[0072] After obtaining the focal semantic feature vector, the review server performs a final compliance classification on the long-text health and wellness network information to be detected based on the focal semantic feature vector. Specifically, the review server inputs the focal semantic feature vector into a preset classifier, which typically consists of one or more fully connected layers. The output layer is then connected to a Sigmoid activation function (for binary classification scenarios) or a Softmax activation function (for multi-class classification scenarios), mapping the focal semantic feature vector to a predicted probability distribution for each compliance category. The review server determines the category with the highest probability based on the probability distribution and uses it as the compliance identification result for the long-text health and wellness network information. This result, along with the confidence score, is output to the upper-level review management platform or returned to the caller.

[0073] The training process of the classifier model includes:

[0074] A training set of long-text health and hygiene network information with genuine compliance labels is obtained. This training set includes compliant medical science popularization texts and long-text health and hygiene network information containing localized violations. Following the aforementioned process of segmenting, extracting features, predicting, and calculating deviation metrics for long-text health and hygiene network information, the actual semantic feature vectors and corresponding deviation metric sequences for each logical block corresponding to each training text are obtained. The attention weight coefficients for each logical block are calculated based on the deviation metric sequence, and the actual semantic feature vectors are weighted and summed to obtain the focal semantic feature vectors used to represent localized logical mutation information. The focal semantic feature vectors are input into the classifier to be trained, and the classifier outputs the predicted compliance probability distribution for each training text. The cross-entropy loss function value between the predicted compliance probability distribution and the genuine compliance label is calculated. Based on the cross-entropy loss function value, the network connection weights and bias parameters of the classifier are updated using the gradient descent optimization algorithm through backpropagation. The above iterative optimization process is repeated until the cross-entropy loss function value converges to a stable state, thereby enabling the classifier to accurately identify the overall compliance of long-text health and hygiene network information based on the focal semantic feature vectors.

[0075] The classifier parameters are jointly optimized end-to-end with the pre-set autoregressive prediction network and semantic feature extraction model. The training objective is to minimize the cross-entropy loss of compliance classification, thereby ensuring the consistency of each module in the optimization objective and improving the overall accuracy of violation identification.

[0076] In this embodiment, a technical solution is adopted that divides long text information of health and wellness networks into logical block sequences according to preset semantic integrity rules, generates predicted semantic feature vectors based on the preceding actual semantic feature vectors through a preset autoregressive prediction network, and calculates the deviation metric sequence between the actual and predicted semantic feature vectors. This determines the attention weight coefficients and sums them to obtain the focal semantic feature vectors representing local logical mutation information for classification. Therefore, logical mutation blocks with larger deviation metric values ​​can dominate the synthesis of focal semantic feature vectors with higher weights, avoiding the dilution of mutation features. This effectively solves the problem in related technologies where global semantic features are easily dominated by long and correct medical knowledge, making it difficult to capture local logical breaks. Thus, accurate identification of long text information of health and wellness networks with local intent shifts is achieved.

[0077] In the above embodiment, attention weights are assigned by calculating the deviation metric between actual and predicted semantic features, effectively extracting focal features of local logical mutations for compliance identification. In practical applications, normal cross-paragraph transitions in medical knowledge are often accompanied by significant semantic deviations, and relying solely on semantic deviations may lead to misjudgments.

[0078] Based on the above embodiments, the method provided in this embodiment will be described in further detail below. Please refer to... Figure 3 This is another flowchart illustrating a method for monitoring and identifying health network information in this application.

[0079] S301. Obtain the long text information of the health and wellness network to be detected, and divide the long text information of the health and wellness network into multiple logical blocks according to the preset semantic integrity rules, to obtain a sequence of logical blocks arranged in text order. The preset semantic integrity rules include at least one of paragraph boundary recognition, sentence semantic dependency analysis, or topic transformation detection.

[0080] Step S301 is similar to step S201 in the above embodiments, and will not be repeated here.

[0081] S302. Input each logical block into the pre-trained semantic encoder to obtain the contextual semantic representation of each logical block.

[0082] Among them, the pre-trained semantic encoder refers to a deep neural network model pre-trained on a large-scale health and wellness text corpus, used to encode the input text sequence into a vector representation containing deep semantic information; the context semantic representation refers to the hidden state vector sequence output by the pre-trained semantic encoder after encoding all words in the logical block, containing the position information of each word and its contextual interaction information with other words in the block. It is usually a two-dimensional tensor, whose first dimension corresponds to the position index of each word in the logical block, and the second dimension corresponds to the hidden state vector dimension of each word.

[0083] After constructing the logical block sequence, the review server performs deep semantic encoding on each logical block. Specifically, the review server inputs each logical block in the logical block sequence independently into a pre-trained semantic encoder. Before input, each logical block is first segmented into a word sequence, and special markers (such as [CLS] and [SEP] markers) are added to the beginning and end of the sequence. Then, it is fed into the multi-layer Transformer encoding structure of the pre-trained semantic encoder. The pre-trained semantic encoder fully models the semantic dependencies between words within the logical block using a multi-head self-attention mechanism, enabling the hidden state vector of each word to incorporate the semantic role information of that word in the context of the current logical block. The review server extracts the contextual semantic representation corresponding to each logical block from the last layer output of the pre-trained semantic encoder.

[0084] The training process of the pre-trained semantic encoder includes:

[0085] This study collects large-scale unlabeled professional text data in the medical field, including medical textbooks, clinical practice guidelines, authoritative medical journal articles, and compliant health science articles, to construct a medical corpus. The text in the medical corpus is segmented, and some words are randomly masked in the word sequence to construct a masked language model training task. The processed word sequence is input into a pre-trained semantic encoder, and a multi-layer multi-head self-attention mechanism is used to extract the contextual hidden state representation of each word in the sequence. Based on the contextual hidden state representation of the masked position, the prediction layer outputs the lexical probability distribution of the masked words. The cross-entropy loss function value between the predicted lexical probability distribution and the real word label is calculated. Based on the cross-entropy loss function value, the network parameters of the pre-trained semantic encoder are updated by backpropagation using an optimization algorithm. After iterative training with a large number of batches of data, the pre-trained semantic encoder can fully capture the professional terminology dependencies and deep contextual semantic interaction information in medical texts, and output accurate contextual semantic representations.

[0086] Optionally, in some embodiments, the review server can also perform entity annotation on medical terms within the logical blocks before inputting the logical blocks into the pre-trained semantic encoder, and superimpose preset entity type embedding vectors on medical entity tokens during the token embedding stage, so that the pre-trained semantic encoder can explicitly perceive the type information of medical entities (such as diseases, drugs, treatment methods, etc.) during the encoding process, thereby improving the accuracy of contextual semantic representation in capturing key information in the medical field.

[0087] S303. Perform pooling processing on the context semantic representation of each logical block to obtain the actual semantic feature vector corresponding to each logical block.

[0088] Pooling refers to the operation of aggregating and compressing multiple hidden state vectors of lexical units contained in the context semantic representation, aiming to convert a variable-length sequence of lexical-level vectors into a single vector representation of fixed dimensions.

[0089] After obtaining the contextual semantic representation of each logical block, the review server performs pooling processing on the contextual semantic representation to generate a fixed-dimensional actual semantic feature vector. Specifically, the review server applies a preset pooling strategy to aggregate the vectors based on the contextual semantic representation corresponding to each logical block in the logical block sequence.

[0090] Taking average pooling as an example, the review server calculates the arithmetic mean of the hidden state vectors of all words in the context semantic representation according to the feature dimension, and obtains a dense vector of fixed dimension, which serves as the actual semantic feature vector of the logical block.

[0091] The review server performs pooling processing on all logical blocks in the logical block sequence in sequence, and arranges the actual semantic feature vectors of each logical block in text order to obtain the actual semantic feature sequence.

[0092] S304. Input the actual semantic feature sequence into the preset autoregressive prediction network. For each target logical block in the logical block sequence, predict the semantic evolution direction of the target logical block based on the actual semantic feature vectors before the target logical block, and generate the predicted semantic feature vector corresponding to the target logical block.

[0093] S305. Calculate the deviation metric between the actual semantic feature vector and the predicted semantic feature vector of the target logical block, and traverse the logical block sequence to obtain the deviation metric sequence corresponding to each logical block.

[0094] Steps S304 and S305 are similar to those described in steps S203 and S204 in the above embodiments, and will not be repeated here.

[0095] The review server performs medical logic rationality checks on candidate mutation blocks in the deviation metric sequence whose deviation metric values ​​are greater than or equal to a preset mutation threshold, including:

[0096] S306. Extract the current medical entity set within the candidate mutation block, and the preceding medical entity set within the logical block adjacent to the candidate mutation block in the logical block sequence.

[0097] Among them, candidate mutation blocks indicate that the actual semantics of the logical block deviate significantly from the expected semantics inferred from the preceding text; the preset mutation threshold refers to the deviation metric used to screen candidate mutation blocks. This threshold can be determined based on the distribution statistics of deviation metrics between normal medical transition blocks and illegal mutation blocks in historical labeled samples (such as taking the quantile at the intersection of the two distributions), and can be adjusted according to the business's tolerance for false positive and false negative rates; the current medical entity set refers to the set of all medical domain named entities identified and extracted from the text content of the candidate mutation blocks, including entity types such as disease names, drug names, treatment methods, medical devices, and health product names; the preceding medical entity set refers to the set of all medical domain named entities identified and extracted from the text content of the logical block adjacent to the candidate mutation block in the logical block sequence, used to characterize the medical topic focus before the candidate mutation block.

[0098] After obtaining the deviation metric value sequence, the review server performs medical entity extraction on candidate mutation blocks with significant deviation metric values. Specifically, the review server first traverses the deviation metric value sequence, filtering out logical blocks with deviation metric values ​​greater than or equal to a preset mutation threshold, and marking them as candidate mutation blocks. Subsequently, for each candidate mutation block, the review server applies a preset medical named entity recognition model (such as an entity recognizer based on BiLSTM-CRF or BERT-CRF) to perform entity recognition on the text content of the candidate mutation block, extracting all medical domain entities to form the current medical entity set. Simultaneously, the review server locates the preceding logical block in the logical block sequence of the candidate mutation block, and also performs medical named entity recognition on the text content of the preceding logical block, extracting all medical entities to form the preceding medical entity set.

[0099] Specifically, when the candidate mutation block is the first block in the sequence, since there is no preceding logical block, the review server directly extracts the title or summary content of the long text information of the health network, and applies a preset medical named entity recognition model to extract medical entities as the preceding medical entity set. If the long text information of the health network does not have a title or summary, the review server directly skips the step of obtaining the graph connection path, sets its logical coherence confidence to a preset confidence level. If it is equal to zero, it directly determines that it does not belong to the compliant medical stage transition, and triggers the subsequent calculation of the false marketing risk value of the unregistered entities in the block, so as to ensure that the abnormal situation of illegal mutation appearing at the beginning can be effectively intercepted.

[0100] The training process for the medical named entity recognition model includes:

[0101] A large dataset of manually annotated medical entities is obtained, containing a massive amount of long text samples from the health and wellness network. Each medical named entity, such as disease names, drug names, and treatment methods, is assigned a corresponding entity type boundary label. This long text information with entity type boundary labels is input into a pre-defined medical named entity recognition model. First, the text sequence is encoded using a low-level word embedding layer and a bidirectional long short-term memory network layer, extracting the contextual feature representation of each character or word in the text sequence. The extracted contextual feature representation is then input into a conditional random field decoding layer to calculate the transition probability and emission probability of each possible label sequence corresponding to the entire text sequence, outputting the predicted entity label sequence. The predicted entity label sequence is compared with the manually annotated true entity type boundary labels, and the negative log-likelihood loss function value of the sequence label is calculated. The parameters of each layer of the pre-defined medical named entity recognition model are adjusted based on the negative log-likelihood loss function value using a backpropagation algorithm until the loss value is minimized, enabling the model to accurately extract the current set of medical entities and the previous set of medical entities.

[0102] Optionally, in some embodiments, the review server can also extract global medical department tags from long text information on the health network, for example, by mapping high-frequency entities in the full text to their respective departments, and dynamically generate corresponding preset mutation thresholds based on the global medical department tags; wherein, higher preset mutation thresholds are assigned to departments with large logical spans (such as general internal medicine or general practice), and lower preset mutation thresholds are assigned to logically rigorous specialties (such as cardiovascular surgery). By introducing a dynamic threshold mechanism driven by department themes, misjudgments caused by inherent differences in expression habits among different medical professional fields are further reduced, and the robustness of mutation detection is improved.

[0103] S307. Map the current set of medical entities and the previous set of medical entities to the pre-built medical association knowledge graph, and obtain the graph connection path between the current set of medical entities and the previous set of medical entities in the medical association knowledge graph.

[0104] Among them, the pre-built medical association knowledge graph refers to a pre-constructed structured knowledge network with medical entities as nodes and medical relationships as edges. The nodes include medical entities such as diseases, symptoms, drugs, treatment methods, and examination items, and the edges represent the medical relationships between entities (such as disease-symptom, disease-treatment drug, drug-side effects, etc.). This knowledge graph can be constructed based on authoritative medical databases (such as ICD coding database, drug instruction database, clinical guideline literature, etc.).

[0105] A graph connection path refers to a complete path in a pre-built medical knowledge graph that starts from a certain entity node in the previous set of medical entities, goes through several related edges, and reaches a certain entity node in the current set of medical entities. For example, a path from the hypertension disease node to the antihypertensive drug node through the treatment drug relationship edge, and then to the calcium channel blocker node through the drug classification relationship edge.

[0106] After extracting the current and previous sets of medical entities, the review server maps these two sets to a pre-built medical association knowledge graph and calculates the graph connection paths between entities. Specifically, the review server first matches the names of each entity in the current and previous sets with nodes in the pre-built medical association knowledge graph. For entities that successfully match, they are mapped to corresponding nodes in the knowledge graph. Subsequently, the review server executes path search algorithms in the knowledge graph, such as breadth-first search or bidirectional search, to calculate all reachable paths or the shortest path between each node in the previous and current sets of medical entities. The review server then summarizes the obtained graph connection paths, with each path containing information such as the sequence of nodes traversed, the sequence of relation edge types, and the total number of hops.

[0107] If there is no connection path between the previous set of medical entities and the current set of medical entities in the knowledge graph, or if there are entities in the two sets that cannot be mapped to the knowledge graph (i.e., unregistered entities), it indicates that the medical topics involved in the two blocks lack direct connection at the level of medical knowledge, which is a strong signal of incoherent medical logic.

[0108] At this point, the review server directly sets the logical coherence confidence level to the preset minimum confidence level value and marks the candidate mutation block as awaiting manual review. This preset minimum confidence level is 0 or a very small value close to 0 (e.g., 0.01), ensuring that the logical coherence confidence level is necessarily less than the reasonable transition threshold in subsequent steps. This identifies the candidate mutation block as not belonging to a compliant medical stage transition, triggering an upward adjustment process based on the deviation metric value for false marketing risk.

[0109] S308. Based on the shortest number of hops in the graph connection path and the medical association probability of each node on the path, calculate the logical coherence confidence of the transfer from the pre-logic block to the candidate mutation block.

[0110] Among them, the shortest hop count refers to the minimum number of relation edges traversed in the graph connection path from a node in the preceding medical entity set to a node in the current medical entity set; the medical association probability refers to the medical co-occurrence probability or causal strength corresponding to the relation edges between nodes in the graph connection path, reflecting the credibility of the relationship in real medical literature or clinical practice; the logical coherence confidence score is a confidence score that quantitatively evaluates the medical logical rationality of the transfer from the preceding logical block to the candidate mutation block after comprehensively considering the shortest hop count of the graph connection path and the medical association probability of each node in the path.

[0111] After obtaining the knowledge graph connection paths, the review server first extracts the shortest hop count from all knowledge graph connection paths. If multiple paths exist, the minimum hop count is selected. Then, the review server extracts the medical association probability corresponding to each relation edge on the shortest path. This probability can be pre-calculated based on entity co-occurrence statistics in large-scale medical literature or relation strength annotations in clinical knowledge bases and stored in the edge attributes of the knowledge graph. The medical association probabilities of each edge on the path are multiplied or weighted averaged to obtain the overall medical association strength of the path. The review server combines the shortest hop count and the path's medical association strength to calculate the logical coherence confidence score using a preset confidence score calculation formula. The confidence score calculation formula can use... Where hops is the shortest number of hops. Let α represent the path-medical association probability, and β represent hyperparameters. α and β are hyperparameters used to balance the weights of the shortest hop count and the medical association probability in the confidence calculation. The recommended value range for α is 0.3-0.5, and the recommended value range for β is 0.5-0.7. The specific values ​​can be determined through grid search based on the performance on the validation set.

[0112] It should be noted that the above confidence calculation formula only applies to the case where there is at least one connection path between the previous set of medical entities and the current set of medical entities in the knowledge graph.

[0113] The audit server employs special handling strategies when the following abnormal situations occur:

[0114] (1) If all entities in the previous medical entity set or the current medical entity set cannot be mapped to the knowledge graph, that is, all are unregistered entities, then the graph path search cannot be performed, and the review server directly sets the logical coherence confidence to 0, indicating that there is a complete lack of medical logic support.

[0115] (2) If both entity sets can be partially mapped to the knowledge graph, but there is no connection path between the mapped entities, it indicates that the medical topics involved in the two blocks are completely isolated at the knowledge level, and the review server will also set the logical coherence confidence to 0.

[0116] (3) If only some entities in two entity sets can be mapped to the knowledge graph, and there is a connection path between the mapped entities, the audit server will only calculate the logical coherence confidence based on the mapped entities, but will conduct a special evaluation of the unlogged entities in the subsequent calculation of the false marketing risk value.

[0117] Through the aforementioned layered processing strategy, this application is able to maintain the ability to effectively identify illegal content even when the knowledge graph coverage is incomplete.

[0118] At this point, step S308 falls into the following two categories:

[0119] If there is no connection path between the current set of medical entities and the previous set of medical entities in the medical association knowledge graph, or if there are entities in the current set of medical entities and the previous set of medical entities that cannot be mapped to the medical association knowledge graph, then the logical coherence confidence is set to the preset minimum confidence value.

[0120] If there is at least one connection path between the current set of medical entities and the previous set of medical entities in the medical association knowledge graph, then the logical coherence confidence of the transfer from the previous logical block to the candidate mutation block is calculated based on the shortest hop count of the graph connection path and the medical association probability of each node on the path.

[0121] Based on the results of medical logic rationality verification, the deviation metric values ​​corresponding to the candidate mutation blocks are adjusted to obtain a corrected deviation metric value sequence, including:

[0122] S309. If the logical coherence confidence level is less than the reasonable transition threshold, the candidate mutation block is determined not to be a compliant medical stage transition, and the false marketing risk value is calculated.

[0123] This step specifically includes:

[0124] If the logical coherence confidence is less than a reasonable transition threshold, then identify the unregistered entities in the medical association knowledge graph within the current medical entity set.

[0125] Extract the affix features and contextual modification features of unregistered entities, and calculate the false marketing risk value of unregistered entities based on a pre-defined medical marketing risk classification model.

[0126] Among them, the reasonable transition threshold refers to the logical coherence confidence criterion used to determine whether a candidate mutation block belongs to the compliant medical stage transition. This threshold can be determined based on the logical coherence confidence distribution of compliant medical transition blocks and non-compliant mutation blocks in historical labeled samples (such as taking the optimal split point of the two distributions); unregistered entities refer to entities in the current medical entity set that cannot find a matching node in the pre-built medical association knowledge graph, indicating that the entity has not been included in the authoritative medical knowledge base and may belong to non-compliant content such as forged medical terms, fictitious health product names, or exaggerated special effects treatments; false marketing risk value refers to the risk probability score output by the preset medical marketing risk classification model after evaluating unregistered entities.

[0127] After calculating the logical coherence confidence score, the review server first determines whether the logical coherence confidence score is less than a preset reasonable transition threshold. If this condition is met, the candidate mutation block is determined to be an compliant medical stage transition and may contain mutations with illegal intent. Subsequently, the review server performs a knowledge graph matching check on each entity in the current medical entity set to identify unregistered entities whose corresponding nodes cannot be found in the pre-built medical association knowledge graph.

[0128] For each unregistered entity, the review server extracts its affix features (identifying marketing affixes in the entity name through regular expressions or affix dictionary matching) and contextual modification features (extracting adjectives, verbs, and their syntactic relationships that modify the entity through dependency parsing). The review server inputs the extracted feature vectors into a pre-defined medical marketing risk classification model, and the model outputs a false marketing risk value for the unregistered entity.

[0129] If there are multiple unregistered entities in the current medical entity set, the review server takes the maximum or weighted average of the false marketing risk values ​​of each unregistered entity as the overall false marketing risk value of the candidate mutation block.

[0130] The preset medical marketing risk classification model outputs a probability score for unregistered entities belonging to false marketing content. This can be achieved using traditional classifiers based on feature engineering (such as logistic regression and random forest) or text classification models based on deep learning (such as TextCNN and BERT classifiers).

[0131] The training process for the medical marketing risk classification model includes:

[0132] This process involves collecting samples of various unregistered entities from historical medical audit data. Each unregistered entity is assigned a risk label through manual review and labeling or business rule matching. Fake terminology and fraudulent treatments are labeled as high-risk marketing, while normal emerging medical terms are labeled as low-risk. A risk classification training set is constructed. For each unregistered entity sample in the training set, natural language processing tools are used to extract its affix features and contextual modification features from the original text. These features are then concatenated and fused into a comprehensive feature vector. This comprehensive feature vector is input into a pre-defined medical marketing risk classification model to be trained. After processing through a multi-layer neural network and non-linear mapping, the predicted false marketing risk value for the unregistered entity is output. The binary cross-entropy loss function between the predicted false marketing risk value and the true risk label is calculated. Gradient descent is used to iteratively update the weight parameters of the pre-defined medical marketing risk classification model based on the cross-entropy loss function value until the network performance reaches its optimal level on the validation set, ensuring that the model can accurately calculate the false marketing risk value when facing unknown entities.

[0133] S310. Adjust the deviation metric value of the candidate mutation block upwards based on the false marketing risk value to obtain a corrected deviation metric value sequence, including:

[0134] If the risk value of false marketing is greater than the preset risk threshold, the candidate mutation block is determined to be an illegal mutation guided by fake terminology;

[0135] The risk amplification factor is calculated based on the risk value of false marketing, and the deviation metric value of the candidate mutation block is adjusted upward according to the risk amplification factor to obtain the corrected deviation metric value sequence.

[0136] Among them, the preset risk threshold is a benchmark for determining whether an unregistered entity belongs to the category of illegal mutations guided by forged terminology. Its function is to define the boundary between high-risk illegal marketing entities and normal unregistered entities (such as emerging medical terms, rare disease names, etc.). This threshold can be determined by statistically analyzing the distribution of illegal marketing risk values ​​of illegal marketing entities and normal unregistered entities in historical labeled samples. The risk amplification coefficient is a multiplicative coefficient used to adjust the deviation metric value of candidate mutation blocks based on the illegal marketing risk value.

[0137] After calculating the false marketing risk value of candidate mutation blocks, the audit server adjusts the deviation metric value of the candidate mutation blocks upwards based on the false marketing risk value. Specifically, the audit server first determines whether the false marketing risk value of the candidate mutation block is greater than a preset risk threshold. If this condition is met, the candidate mutation block is determined to be a violation mutation guided by fabricated terminology, and its deviation metric value needs to be amplified. The audit server calculates a risk amplification coefficient based on the false marketing risk value. The calculation method can use linear or non-linear mapping to ensure that the risk amplification coefficient monotonically increases with the false marketing risk value. Subsequently, the audit server multiplies the original deviation metric value of the candidate mutation block by the risk amplification coefficient to obtain the adjusted deviation metric value, and replaces the corresponding value in the original deviation metric value sequence with this corrected value. After the audit server performs the upward adjustment process on all candidate mutation blocks that meet the conditions, it obtains a deviation metric value sequence containing the upward adjustment.

[0138] S311. If the logical coherence confidence is greater than or equal to the preset reasonable transition threshold, the candidate mutation block is determined to be a compliant medical stage transition, and the decay smoothing coefficient is calculated based on the logical coherence confidence.

[0139] Among them, compliant medical stage transition refers to the topic transition between the candidate mutation block and the preceding logical block having sufficient medical connection path support in the pre-built medical association knowledge graph, and a reasonable logical jump that conforms to the normal medical knowledge evolution law. For example, if an article shifts from a discussion of "the pathogenesis of diabetes" to an introduction of "insulin treatment plan", although the topic transition will cause a certain deviation measurement value, the transition is a reasonable progression of diagnosis and treatment stage in medical logic and should not be regarded as an illegal intentional mutation.

[0140] The attenuation smoothing coefficient is a multiplicative coefficient calculated based on logical coherence confidence and used to downgrade the deviation metric of compliant medical phase transition blocks. Its value range is limited to between zero and one (excluding zero, to avoid completely erasing the deviation metric of the block).

[0141] After determining that the logical coherence confidence of a candidate mutation block is greater than or equal to a preset reasonable transition threshold, the review server calculates a decay smoothing coefficient for that candidate mutation block. Specifically, for candidate mutation blocks with a logical coherence confidence greater than or equal to the preset reasonable transition threshold, the review server determines that the block belongs to a compliant medical stage transition. The review server calculates the decay smoothing coefficient based on the logical coherence confidence, which can be achieved using an inverse proportional mapping method or through the inverse mapping of the Sigmoid function. This ensures that the decay smoothing coefficient monotonically decreases as the logical coherence confidence increases. Compliant transition blocks with high confidence will receive a smaller decay smoothing coefficient, corresponding to a larger subsequent reduction in the deviation metric.

[0142] S312. The deviation metric of the candidate mutation block is down-adjusted using the decay smoothing coefficient to obtain the corrected deviation metric sequence.

[0143] Specifically, the review server multiplies the original deviation metric of the candidate mutation block by the decay smoothing coefficient calculated in step S311 to obtain the adjusted deviation metric of the candidate mutation block. Since the value range of the decay smoothing coefficient is limited to between zero and one, the adjusted deviation metric is necessarily less than the original deviation metric, which meets the expected effect of deviation suppression for compliant conversion blocks.

[0144] The review server sequentially adjusts the deviation metric value of all candidate mutation blocks in the logical block sequence according to the processing logic of steps S309 to S312, while keeping the deviation metric value of non-candidate mutation blocks unchanged, i.e., logical blocks whose deviation metric value is less than the preset mutation threshold. The deviation metric values ​​of all logical blocks are summarized in the original text order to obtain the corrected deviation metric value sequence, which is then passed to step S313 for the calculation of attention weight coefficients.

[0145] S313. Determine the attention weight coefficients corresponding to each logical block based on the corrected deviation metric sequence.

[0146] S314. The actual semantic feature vectors corresponding to each logical block are weighted and summed according to the attention weight coefficient to obtain the focus semantic feature vector, which is used to represent local logical mutation information.

[0147] S315. Classify long text information of health and wellness networks based on focus semantic feature vectors, and output the compliance identification results of long text information of health and wellness networks.

[0148] Steps S313 to S315 are similar to steps S205 and S206 in the above embodiments, and will not be repeated here.

[0149] In this embodiment, the technical solution employs a method that, after calculating the deviation metric sequence, extracts the medical entity set of candidate mutation blocks and preceding logic blocks, maps it to a medical knowledge graph for medical logic rationality verification, and adjusts the deviation metric in both directions based on logical coherence confidence combined with false marketing risk values ​​or decay smoothing coefficients. Finally, it aggregates the focal semantic feature vector based on the corrected deviation metric sequence. Therefore, this method can distinguish between compliant medical stage transitions and illegal mutations lacking medical basis. While highlighting genuine violation signals, it suppresses the weight ratio of compliant transition blocks, effectively solving the problem that relying solely on semantic distance can easily misjudge normal medical knowledge jumps as illegal intent mutations. This achieves high-precision, low-false-report compliance identification of long text information on health and hygiene networks that is concealed and logically misleading.

[0150] The following description of the review server in the embodiments of this invention application is from the perspective of hardware processing. Please refer to [link / reference needed]. Figure 4 This is a schematic diagram of the physical device structure of the review server in the embodiments of this application.

[0151] It should be noted that, Figure 4 The structure of the audit server shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0152] like Figure 4 As shown, the audit server includes a CPU 401, which can perform various appropriate actions and processes based on a program stored in the read-only memory ROM 402 or a program loaded from the storage section 408 into the random access memory RAM 403, such as performing the methods described in the above embodiments. The RAM 403 also stores various programs and data required for system operation. The CPU 401, ROM 402, and RAM 403 are interconnected via a bus 404. An I / O interface 405 is also connected to the bus 404.

[0153] The following components are connected to I / O interface 405: input section 406 including audio input devices, push-button switches, etc.; output section 407 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 408 including a hard disk, etc.; and communication section 409 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from them can be installed into storage section 408 as needed.

[0154] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program / instructions carried on a computer-readable medium, the computer program / instructions containing computer program / instructions for performing the methods shown in the flowcharts. In such embodiments, the computer program / instructions can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by CPU 401, it performs the various functions defined in the present invention.

[0155] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0156] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0157] Specifically, the review server in this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the monitoring and identification method for health network information provided in the above embodiment.

[0158] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the audit server described in the above embodiments; or it may exist independently and not assembled into the audit server. The storage medium carries one or more computer programs, which, when executed by a processor of the audit server, cause the audit server to implement the health network information monitoring and identification method provided in the above embodiments.

[0159] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0160] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0161] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for monitoring and identifying health and hygiene network information, applied to an audit server, characterized in that, The method includes: The system acquires long text information of the health network to be detected, and divides the long text information of the health network into multiple logical blocks according to preset semantic integrity rules, resulting in a logical block sequence arranged in text order; the semantic integrity rules include at least one of paragraph boundary recognition, sentence semantic dependency analysis, or topic conversion detection. Extract the actual semantic feature vectors of each logical block in the logical block sequence to obtain the actual semantic feature sequence arranged in the text arrangement order; The actual semantic feature sequence is input into a preset autoregressive prediction network. For each target logical block in the logical block sequence, the semantic evolution direction of the target logical block is predicted based on the actual semantic feature vectors before the target logical block, and the predicted semantic feature vector corresponding to the target logical block is generated. Calculate the deviation metric between the actual semantic feature vector and the predicted semantic feature vector of the target logical block, and traverse the logical block sequence to obtain the deviation metric sequence corresponding to each logical block; Based on the deviation metric sequence, the attention weight coefficients corresponding to each logical block are determined, and the actual semantic feature vectors corresponding to each logical block are weighted and summed to obtain the focus semantic feature vector, which is used to characterize local logical mutation information. Based on the focal semantic feature vector, the long text information of the health network is classified, and the compliance identification result of the long text information of the health network is output.

2. The method according to claim 1, characterized in that, After calculating the deviation metric between the actual semantic feature vector and the predicted semantic feature vector of the target logical block, and traversing the logical block sequence to obtain the deviation metric sequence corresponding to each logical block, the method further includes: Perform medical logic rationality verification on candidate mutation blocks in the deviation metric value sequence that have a deviation metric value greater than or equal to a preset mutation threshold. Based on the results of the medical logic rationality verification, the deviation metric values ​​corresponding to the candidate mutation blocks are adjusted to obtain a corrected deviation metric value sequence. Accordingly, determining the attention weight coefficients corresponding to each logical block based on the deviation metric sequence includes: Based on the corrected sequence of deviation metrics, the attention weight coefficients corresponding to each logical block are determined.

3. The method according to claim 2, characterized in that, The step of performing medical logic rationality verification on candidate mutation blocks in the deviation metric value sequence whose deviation metric value is greater than or equal to a preset mutation threshold specifically includes: Extract the current set of medical entities within the candidate mutation block, and the set of preceding medical entities within the preceding logical block adjacent to the candidate mutation block in the logical block sequence; Map the current set of medical entities and the previous set of medical entities to a pre-built medical association knowledge graph, and obtain the graph connection path between the current set of medical entities and the previous set of medical entities in the medical association knowledge graph; Based on the shortest hop count of the graph connection path and the medical association probability of each node on the path, the logical coherence confidence of the transfer from the pre-logic block to the candidate mutation block is calculated.

4. The method according to claim 3, characterized in that, The step of adjusting the deviation metric value corresponding to the candidate mutation block based on the result of the medical logic rationality verification to obtain a corrected deviation metric value sequence specifically includes: If the logical coherence confidence is less than the reasonable transition threshold, the candidate mutation block is determined not to be a compliant medical stage transition, and a false marketing risk value is calculated. The deviation metric value of the candidate mutation block is adjusted upward based on the false marketing risk value to obtain a corrected deviation metric value sequence. If the logical coherence confidence is greater than or equal to a preset reasonable transition threshold, the candidate mutation block is determined to be a compliant medical stage transition, and the decay smoothing coefficient is calculated based on the logical coherence confidence. The deviation metric value of the candidate mutation block is down-adjusted using the decay smoothing coefficient to obtain a corrected deviation metric value sequence.

5. The method according to claim 4, characterized in that, If the logical coherence confidence level is less than the reasonable transition threshold, then the candidate mutation block is determined not to belong to a compliant medical stage transition, and a false marketing risk value is calculated, specifically including: If the logical coherence confidence is less than the reasonable transition threshold, then identify the unregistered entities in the medical association knowledge graph within the current medical entity set; Extract the affix features and contextual modification features of the unregistered entities, and calculate the false marketing risk value of the unregistered entities based on a preset medical marketing risk classification model.

6. The method according to claim 4 or 5, characterized in that, The deviation metric value of the candidate mutation block is adjusted upward based on the false marketing risk value to obtain a corrected deviation metric value sequence, specifically including: If the risk value of false marketing is greater than the preset risk threshold, the candidate mutation block is determined to be an illegal mutation guided by fake terminology; Based on the risk value of false marketing, a risk amplification factor is calculated, and the deviation metric value of the candidate mutation block is adjusted upward according to the risk amplification factor to obtain a corrected deviation metric value sequence.

7. The method according to claim 1, characterized in that, The step of extracting the actual semantic feature vectors of each logical block in the logical block sequence to obtain the actual semantic feature sequence arranged in the text order specifically includes: Each logical block is input into a pre-trained semantic encoder to obtain the contextual semantic representation of each logical block. Pooling is performed on the contextual semantic representation of each logical block to obtain the actual semantic feature vector corresponding to each logical block.

8. An audit server, characterized in that, The audit server includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the audit server to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed on the audit server, the audit server performs the method as described in any one of claims 1-7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are run on the audit server, the audit server performs the method as described in any one of claims 1-7.

Images