Output content detection method based on semantic analysis detection model

Abstract

The invention discloses an output content detection method based on a semantic analysis detection model, and the method specifically comprises the following steps: S1, collecting multi-modal data, and carrying out the unified coding of the data, and generating a structured semantic vector; s2, analyzing the structured semantic vector, extracting a subject-predicate-object triple, and generating a dynamic semantic vector matrix; s3, constructing a two-channel semantic detection model, and outputting syntactic structure scores and semantic feature scores according to data in the dynamic semantic vector matrix; s4, calculating a comprehensive risk value according to the syntactic structure score and the semantic feature score, and performing dynamic joint decision making; s5, generating a three-dimensional detection report according to a decision result; and S6, dynamically adjusting and optimizing the model according to the content of the detection report. According to the method, multi-dimensional quantification of complex semantics is achieved, misjudgment and missed judgment of traditional single threshold judgment are reduced, a user is assisted in efficiently processing problems, and the long-term application requirement in a complex scene is met.

Description

Technical Field

[0001] This invention relates to the field of semantic analysis technology, specifically to a method for detecting the output content of a semantic analysis detection model. Background Technology

[0002] Traditional solutions often employ a single-modal data processing approach, analyzing text, images, and audio data independently. For text data, they only extract surface features such as keyword frequency and part-of-speech tagging; for images, they rely solely on pixel values ​​or simple visual features, lacking the ability to mine cross-modal semantic relationships. This approach leads to the fragmentation of complementary information in multi-source data, easily resulting in incomplete semantic understanding, especially when processing mixed-modal content, where detection accuracy drops significantly. Traditional feature extraction methods often rely on static rules or fixed templates, such as sentiment polarity judgment based on predefined dictionaries or fixed structure matching based on syntactic trees. These methods struggle to handle dynamic semantic changes, and static feature extraction in complex sentences often loses deep semantic relationships, resulting in insufficient accuracy of semantic feature scores. Traditional solutions often employ a single model structure, lacking multi-dimensional semantic detection capabilities. These models typically focus only on local semantic features, making it difficult to simultaneously consider syntactic structural integrity and semantic coherence. When detecting semantic contradictions, a single model may only identify keyword conflicts while ignoring contextual logical contradictions; when processing long texts, the model is prone to gradient vanishing, leading to the failure to capture long-distance semantic dependencies and resulting in biased scores for syntactic structure and semantic features. Traditional model optimization relies on manual experience and lacks a dynamic feedback mechanism. When errors occur during detection, manual analysis of erroneous cases and modification of feature engineering processes or model parameters are required, a process that is time-consuming and costly. Traditional detection reports are mostly in two-dimensional table or text format, presenting only a single-dimensional risk value and failing to intuitively demonstrate the correlation between syntactic and semantic risks. This makes it difficult for decision-makers to quickly pinpoint the root cause of problems, resulting in low efficiency in model iteration and optimization. Summary of the Invention

[0003] In view of this, the present invention aims to propose an output content detection method based on a semantic analysis detection model. The method is characterized by comprising a data acquisition end, a feature extraction end, a model construction end, a comprehensive analysis end, a report generation end, and a dynamic optimization end. It collects multimodal data, uniformly encodes the data to generate structured semantic vectors, analyzes the structured semantic vectors, extracts subject-verb-object triples, generates a dynamic semantic vector matrix, constructs a dual-channel semantic detection model, outputs syntactic structure scores and semantic feature scores based on the data within the dynamic semantic vector matrix, calculates a comprehensive risk value, performs dynamic joint decision-making, generates a three-dimensional detection report, and dynamically adjusts and optimizes the model, effectively solving the problems mentioned in the background technology.

[0004] The objective of this invention can be achieved through the following technical solution: a method for detecting output content based on a semantic analysis detection model, characterized by comprising a data acquisition end, a feature extraction end, a model construction end, a comprehensive analysis end, a report generation end, and a dynamic optimization end, specifically including the following steps: S1. Collect multimodal data and encode the data in a unified manner to generate structured semantic vectors; S2. Analyze the structured semantic vectors, extract subject-verb-object triples, and generate a dynamic semantic vector matrix; S3. Construct a dual-channel semantic detection model and output syntactic structure score and semantic feature score based on the data in the dynamic semantic vector matrix; S4. Calculate the comprehensive risk value based on the syntactic structure score and semantic feature score, and make dynamic joint decisions. S5. Generate a 3D inspection report based on the decision results; S6. Dynamically adjust and optimize the model based on the content of the test report.

[0005] The multimodal data includes: structured data such as e-government forms and financial transaction records; unstructured data such as social media texts, surveillance videos, and audio recordings; and spatiotemporal data such as GPS trajectories, geographic information data, and sensor time-series data.

[0006] The method for analyzing structured semantic vectors is as follows: break down compound sentences into independent semantic units and identify the boundaries of main behaviors in the text; extract relationships, record the core structure of each semantic unit, and mark the logical relationships between various clauses; establish a dynamic matrix with the horizontal axis representing time or version sequence and the vertical axis representing clause attributes. When a new agreement is signed, the matrix columns are automatically expanded, and when clauses are revised, version differences are marked.

[0007] The syntactic structure score is calculated by extracting the average dependency path length of the subject-verb-object core components, the percentage of subject-verb-object structural integrity, and different path attenuation coefficients set according to different language types from the dynamic semantic vector matrix. It can be done through the expression: Where S1 is the syntactic structure score, L is the average dependency path length of the subject-verb-object core components, C is the percentage of subject-verb-object structure completeness, and λ is the path decay coefficient.

[0008] The semantic feature score is calculated by extracting the attention entropy value and the semantic coherence of adjacent sentences from the dynamic semantic vector matrix. It can be done through the expression: Where S2 is the semantic feature score, H is the attention entropy value, and D is the semantic coherence between adjacent sentences.

[0009] The comprehensive risk value is calculated using the following expression: Where R is the overall risk value, S1 is the syntactic structure score, S2 is the semantic feature score, and a is the domain adjustment factor.

[0010] The dynamic joint decision-making method is as follows: the comprehensive risk value is judged. When the comprehensive risk value R is less than x, it is judged as low risk and can be directly released; when the comprehensive risk value R is greater than x and less than y, it is judged as medium risk, and manual review is carried out and risk points are marked; when the comprehensive risk value R is greater than y, it is judged as high risk, the output content is intercepted and correction suggestions are generated.

[0011] The three-dimensional inspection report specifically includes a risk tracing dimension, a correction suggestion dimension, and a three-dimensional visualization dimension. The risk tracing dimension includes highlighting the non-compliant statements and associating them with the timestamps of the original audio recordings or scanned documents. The correction suggestion dimension includes generating compliant statements based on semantic recombination technology. The three-dimensional visualization dimension includes using heatmaps to display the associated risks between clauses.

[0012] The method for dynamically adjusting and optimizing the model is as follows: extracting high-frequency misjudgment cases and new risk patterns from the 3D report; automatically reducing the weight of misjudgment features to improve the sensitivity of new risk detection; converting confirmed violation samples into detection rules and updating the knowledge base in real time; using multiple semantic variants to strengthen the model; and testing the new and old models in parallel, automatically replacing the old version after it meets the standards.

[0013] Combining all the above technical solutions, the positive effects of this invention are as follows: 1. This invention significantly improves the completeness and accuracy of content detection through multimodal data fusion and deep semantic mining. Unified encoding of multimodal data generates structured semantic vectors, breaking down the information barriers of traditional single-modal processing and capturing cross-modal semantic relationships between text, images, and audio. Dynamic semantic vector matrix extraction of subject-verb-object triples deeply mines deep semantic relationships, solving the problem of insufficient capture of contextual dependence and semantic changes in traditional static feature extraction, thus laying the foundation for accurate detection.

[0014] 2. This invention enhances the adaptability and reliability of semantic analysis through dynamic feature extraction and a precise scoring mechanism. Syntactic structure scoring incorporates dynamic parameters including average dependency path length, while semantic feature scoring includes indicators such as attention entropy, achieving multi-dimensional quantification of complex semantics. The comprehensive risk value calculation introduces a domain adjustment factor, allowing the scoring to adapt to different scenarios. The dynamic joint decision-making mechanism processes data according to risk level, balancing detection rigor with processing efficiency and reducing false positives and false negatives associated with traditional single-threshold judgments.

[0015] 3. This invention enhances the practicality and operability of testing results through 3D testing reports and intelligent decision support. The risk tracing dimension links to the original data timestamps for quick problem location; the correction suggestion dimension generates compliant statements, reducing manual modification costs; and the heatmap visually presents the risks associated with each clause, making decision-making more intuitive. Compared to traditional 2D reports, 3D reports achieve a closed loop from risk identification to resolution, helping users efficiently handle problems.

[0016] 4. This invention ensures the model's continuous evolution and environmental adaptability through a dynamic optimization mechanism. It automatically adjusts feature weights and updates the knowledge base based on high-frequency misclassification cases, and strengthens training through semantic variants, enabling the model to cope with new risk patterns. The parallel testing and automatic replacement mechanism between old and new models ensures the stability of the optimization process, solving the problems of long cycles and high costs associated with traditional manual optimization. This achieves dynamic improvement in detection capabilities, meeting the long-term application needs in complex scenarios. Attached Figure Description

[0017] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the implementation steps of the method of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 As shown, this invention proposes an output content detection method based on a semantic analysis detection model, including a data acquisition end, a feature extraction end, a model construction end, a comprehensive analysis end, a report generation end, and a dynamic optimization end.

[0021] In a more specific application of this invention, the comprehensive collection and preprocessing of multimodal data is as follows: For structured data, including e-government forms and financial transaction records, it is necessary to establish standardized interfaces or ETL processes for extraction to ensure accurate mapping of data fields, handling of missing values, and uniform format, thus laying the foundation for subsequent coding.

[0022] For unstructured data, social media texts need to undergo cleaning, including noise removal, advertising and irrelevant symbol removal, word segmentation and standardization; surveillance videos need keyframe extraction, target detection and tracking, and frame extraction to reduce computational load; audio recordings need speech recognition to convert into text, and voiceprint separation or sentiment analysis feature extraction.

[0023] For spatiotemporal data, including GPS trajectory and geographic information data, coordinate system 1, trajectory point denoising, stop point or movement pattern recognition, and map matching are required; for sensor time series data, feature engineering is required, including filtering and denoising, outlier detection, timestamp alignment, and extraction of statistical features and frequency domain features.

[0024] After preprocessing multimodal data, a model is built to uniformly encode it, generating a structured semantic vector. A Transformer-based model can be used, employing a multimodal encoder for different modalities. The structured data branch can use embedding layers or simple fully connected networks to process tabular fields; the text branch uses a pre-trained language model to obtain deep semantic representations; the visual and auditory branches utilize convolutional neural networks or audio-specific models to extract high-level features, respectively; the spatiotemporal data branch requires models capable of capturing spatiotemporal dependencies, such as LSTM, GRU, Transformer, or 3DCNN, to handle trajectory and temporal information. A shared semantic space is designed, and through cross-modal attention mechanisms, contrastive learning, or multi-task learning techniques, these heterogeneous feature representations are forced to align and interact within the same high-dimensional vector space, ultimately fusing to output a unified structured semantic vector.

[0025] The structured semantic vectors are analyzed using natural language processing techniques, including dependency parsing and semantic role labeling, to break down complex sentences into independent semantic units. For example, in contract clauses, sentence boundaries are defined by identifying conjunctions, punctuation, or logical connectives, stripping away nested structures so that each unit expresses only a single core semantic meaning. For the decomposed semantic units, the boundaries of the subject's actions need to be further identified. Entity recognition is used to determine the subject, verb analysis extracts the core predicate, and contextual information is used to pinpoint the object, ultimately generating standardized subject-verb-object triples. During this process, domain knowledge bases are used to disambiguate polysemous words; for example, "termination" in a technical agreement might refer to service termination, while in financial clauses it might correspond to contract termination.

[0026] Each triple serves as the basic unit of the matrix, with its attributes such as subject type, predicate semantic strength, and object state encoded as vector features. The horizontal axis of the matrix is ​​arranged according to time or version sequence, such as agreement signing date and revision batch, while the vertical axis relates to the hierarchical levels of clause attributes, such as responsible party, obligation content, and effective conditions. When a new agreement is signed, the system automatically parses the triple structure of its clauses. If new attribute categories are involved, the vertical axis dimension is expanded; if it is a refinement of existing attributes, it is marked as a child node under the same dimension. When clauses are revised, by comparing the changes in predicates or object expansions in the triples of the new and old versions, difference markers are generated at the corresponding time points, and historical version snapshots are retained for retrospective purposes.

[0027] Each cell in the matrix stores a triple vector, and the logical relationships between clauses are modeled using a graph neural network to form weighted directed edges. When an external event is triggered, the system can quickly locate the relevant matrix area and use vector similarity calculations to recommend related clauses or simulate the cascading impact of clause changes on the overall agreement structure. This matrix can be integrated into a smart contract platform to achieve automatic tracking of clause execution status, while providing a structured traceability framework for risk prediction and compliance auditing.

[0028] A dual-channel semantic detection model is constructed, which outputs syntactic structure scores and semantic feature scores based on the data in the dynamic semantic vector matrix.

[0029] The syntactic structure score is calculated as follows: The average dependency path length of the subject-verb-object core components, the percentage of subject-verb-object structural completeness, and different path decay coefficients are extracted from the dynamic semantic vector matrix. The average dependency path length of the subject-verb-object core components can be calculated by using directed edges between nodes in the dependency tree to determine the path steps between the subject and verb, verb and object, and subject and object, and then averaging these three values ​​to obtain the dependency path length. The percentage of subject-verb-object structural completeness requires statistical analysis of the proportion of complete triples in the sentence that simultaneously contain a clear subject, verb, and object. If the sentence contains omitted or implicit components, these are filled in using contextual information or by marking missing items. The final completeness is quantified as (the actual number of subject-verb-object elements / 3) × 100%.

[0030] It can be done through the expression: Where S1 is the syntactic structure score, L is the average dependency path length of the subject-verb-object core components, C is the percentage of subject-verb-object structure completeness, and λ is the path decay coefficient.

[0031] The semantic feature score is calculated as follows: attention entropy and semantic coherence of adjacent sentences are extracted from the dynamic semantic vector matrix. Attention entropy is calculated for different versions of the target clause vector on the timeline, using a cross-version self-attention mechanism to calculate the interaction weight distribution between the clause vector and other related clause vectors. A high concentration of weights results in a lower entropy, while a uniform distribution across multiple clauses results in a higher entropy. The uncertainty of the weight distribution is calculated using the information entropy formula; a higher entropy indicates a more dispersed semantic dependency of the clause. Semantic coherence of adjacent sentences is calculated by taking the cosine similarity or Euclidean distance of adjacent semantic units at the same time point in the latent space, and then applying a dynamic weighted average using a contextual sliding window. High similarity with smooth fluctuations between adjacent unit vectors indicates strong coherence; conversely, a sudden drop in similarity requires verification of the semantic jump's rationality using logical relationship markers.

[0032] It can be done through the expression: Where S2 is the semantic feature score, H is the attention entropy value, and D is the semantic coherence between adjacent sentences.

[0033] The comprehensive risk score is calculated based on the syntactic structure score and the semantic feature score, and can be expressed as: Where R is the overall risk value, S1 is the syntactic structure score, S2 is the semantic feature score, and a is the domain adjustment factor.

[0034] The system assesses the overall risk value. If the overall risk value R is less than x, it is considered low risk and can be released directly. If R is greater than x but less than y, it is considered medium risk, requiring manual review and risk point marking. If R is greater than y, it is considered high risk, the output is intercepted, and correction suggestions are generated. The calculation of the overall risk value incorporates a domain adjustment factor, allowing the scoring to adapt to different scenarios. The dynamic joint decision-making mechanism processes data according to risk level, balancing detection rigor with processing efficiency and reducing false positives and false negatives associated with traditional single-threshold judgments.

[0035] In one specific embodiment, the average dependency path length of the subject-verb-object core components is 2, the subject-verb-object structural integrity percentage is 100%, the path decay coefficient is 0.1, the syntactic structure score is 0.83, the attention entropy value is 0.8, the semantic coherence is 0.9, the semantic feature score is 0.50, the domain factor is 0.8, and the comprehensive risk value is 0.236.

[0036] The preset thresholds are x = 0.3 and y = 0.5. If the overall risk value is less than 0.3, it is considered low risk and can be released directly. If the overall risk value is greater than 0.3 but less than 0.5, it is considered medium risk, and manual review and risk point marking are required. If the overall risk value is greater than 0.5, it is considered high risk, the output content is intercepted, and correction suggestions are generated.

[0037] Based on the decision-making results, a 3D inspection report is generated. This report includes three dimensions: risk tracing, remedial recommendations, and 3D visualization. The risk tracing dimension highlights non-compliant statements and links them to the timestamps of the original audio recordings or scanned documents. The remedial recommendations dimension generates compliant statements based on semantic recombination technology. The 3D visualization dimension uses heatmaps to illustrate the risks associated with different clauses. This 3D report creates a closed loop from risk identification to resolution, helping users efficiently handle problems.

[0038] The model is dynamically adjusted and optimized based on the content of the detection report. This can be achieved through: real-time parsing and feature extraction of the 3D detection report; identification of common features of high-frequency misjudgment cases using natural language processing and clustering algorithms; and analysis of anomalous signals in text representation and behavioral trajectory dimensions of new risk patterns. The system generates dynamic weight decay instructions based on misjudgment features, reducing the decision weight of corresponding features during model inference, while automatically enhancing the detection sensitivity of new risk features through incremental learning.

[0039] For confirmed violation samples, the system triggers a rule generation engine to transform typical cases into executable detection rules and store them in a knowledge graph. The system employs a dual verification mechanism: first, the generalization of the rules is tested using adversarial examples, and then, after manual review and confirmation, the rules are synchronized to the online detection system in real time. The system automatically generates augmented datasets containing synonym replacements, sentence reconstructions, and multilingual translation variations, and trains the model through comparative learning to identify core risk factors.

[0040] The new and old models are iterated in parallel using a shadow mode, where the same input data is fed into both models for prediction result comparison. Key metrics are monitored, including the rate of decrease in the false positive rate, the timeliness of new risk capture, and inference latency. When the false positive rate of the new model decreases to a set threshold, the system automatically triggers a seamless switching mechanism. Containerized deployment ensures service continuity, and the old model version is gradually replaced through canary releases. This parallel testing and automatic replacement mechanism ensures the stability of the optimization process, solves the problems of long cycles and high costs associated with traditional manual optimization, and enables dynamic improvement in detection capabilities, meeting the long-term application needs of complex scenarios.

[0041] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting output content based on a semantic analysis detection model, characterized in that, It includes a data acquisition end, a feature extraction end, a model building end, a comprehensive analysis end, a report generation end, and a dynamic optimization end, specifically including the following steps: S1. Collect multimodal data and encode the data in a unified manner to generate structured semantic vectors; S2. Analyze the structured semantic vectors, extract subject-verb-object triples, and generate a dynamic semantic vector matrix; S3. Construct a dual-channel semantic detection model and output syntactic structure score and semantic feature score based on the data in the dynamic semantic vector matrix; S4. Calculate the comprehensive risk value based on the syntactic structure score and semantic feature score, and make dynamic joint decisions. S5. Generate a 3D inspection report based on the decision results; S6. Dynamically adjust and optimize the model based on the content of the test report.

2. The output content detection method based on the semantic analysis detection model as described in claim 1, characterized in that: The multimodal data includes: structured data such as e-government forms and financial transaction records; unstructured data such as social media texts, surveillance videos, and audio recordings; and spatiotemporal data such as GPS trajectories, geographic information data, and sensor time-series data.

3. The output content detection method based on the semantic analysis detection model as described in claim 1, characterized in that: The method for analyzing structured semantic vectors is as follows: break down compound sentences into independent semantic units, identify the boundaries of main behaviors in the text; extract relationships, record the core structure of each semantic unit, and annotate the logical relationships between various clauses; Create a dynamic matrix with the horizontal axis representing time or version sequence and the vertical axis representing clause attributes. The matrix columns are automatically expanded when a new agreement is signed, and version differences are marked when clauses are revised.

4. The output content detection method based on the semantic analysis detection model as described in claim 1, characterized in that: The syntactic structure score is calculated by extracting the average dependency path length of the subject-verb-object core components, the percentage of subject-verb-object structural integrity, and different path attenuation coefficients set according to different language types from the dynamic semantic vector matrix. It can be done through the expression: Where S1 is the syntactic structure score, L is the average dependency path length of the subject-verb-object core components, C is the percentage of subject-verb-object structure completeness, and λ is the path decay coefficient.

5. The output content detection method based on the semantic analysis detection model as described in claim 1, characterized in that: The semantic feature score is calculated by extracting the attention entropy value and the semantic coherence of adjacent sentences from the dynamic semantic vector matrix. It can be done through the expression: Where S2 is the semantic feature score, H is the attention entropy value, and D is the semantic coherence between adjacent sentences.

6. The output content detection method based on the semantic analysis detection model as described in claim 1, characterized in that: The comprehensive risk value is calculated using the following expression: Where R is the overall risk value, S1 is the syntactic structure score, S2 is the semantic feature score, and a is the domain adjustment factor.

7. The output content detection method based on the semantic analysis detection model as described in claim 1, characterized in that: The dynamic joint decision-making method is as follows: the comprehensive risk value is judged. When the comprehensive risk value R is less than x, it is judged as low risk and can be released directly; when the comprehensive risk value R is greater than x and less than y, it is judged as medium risk, and manual review is carried out and risk points are marked. When the overall risk value R is greater than y, it is judged as high risk, the output content is intercepted and a correction suggestion is generated.

8. The output content detection method based on the semantic analysis detection model as described in claim 1, characterized in that: The three-dimensional inspection report specifically includes a risk tracing dimension, a correction suggestion dimension, and a three-dimensional visualization dimension. The risk tracing dimension includes highlighting the non-compliant statements and associating them with the timestamps of the original audio recordings or scanned documents. The correction suggestion dimension includes generating compliant statements based on semantic recombination technology. The three-dimensional visualization dimension includes using heatmaps to display the associated risks between clauses.

9. The output content detection method based on the semantic analysis detection model as described in claim 1, characterized in that: The method for dynamically adjusting and optimizing the model is as follows: extracting high-frequency misjudgment cases and new risk patterns from the 3D report; automatically reducing the weight of misjudgment features and improving the sensitivity of new risk detection; converting confirmed violation samples into detection rules and updating the knowledge base in real time; The model is enhanced using multiple semantic variants; the new and old models are tested in parallel, and the old version is automatically replaced once the model meets the requirements.

Images