Semantic processing method for 5g message parsing

Abstract

This invention relates to the field of generative artificial intelligence and discloses a semantic processing method for 5G message parsing, comprising: inputting a 5G message text sequence into a generative language model; calculating the information entropy of the intermediate network layer; terminating the calculation and extracting the hidden layer state tensor when the information entropy is lower than a threshold; determining the set of legitimate intentions based on the 5G protocol state and transition topology; modulating the initial logarithmic probability tensor using a mask tensor to mask semantic outputs outside the protocol range; calculating the normalized probability distribution sequence and determining the target parsing intention. This invention maps protocol rules to the boundary of the logarithmic probability space, blocks signaling routing failures caused by probability divergence, resolves the conflict between semantic generation and protocol rigor, and ensures the determinism of service parsing under telecommunications conditions.

Description

Technical Field

[0001] This invention relates to a semantic processing method for 5G message parsing, belonging to the field of generative artificial intelligence technology. Background Technology

[0002] Currently, using pre-trained language models to extract semantic features from text and generate business intents has become the mainstream technical approach for processing multi-round, multimodal interactive messages. In the generative text decoding process, achieving precise alignment between the semantic representation tensor and the logic of specific downstream tasks is a core challenge in improving parsing accuracy in the semantic processing field. Such models capture long-distance semantic dependencies through deep self-attention mechanisms, demonstrating strong semantic representation capabilities when dealing with the parsing of 5G rich media messages with low structure levels. However, in the real-world business evolution of telecommunications core networks, gateway systems face tens of millions of concurrent accesses and stringent... The stringent real-time response constraints and the requirement of the underlying communication protocol that the result of service intent parsing conforms to the legally defined session state machine topology, meaning that each parsed intent instruction must satisfy the logical determinism of the current signaling state, and since the underlying mathematical mechanism of generative artificial intelligence is based on the autoregressive probability distribution of an open vocabulary, its decoding process exhibits nondeterministic characteristics. This mathematical essence has a deep-seated fundamental conflict with the physical conditions that require 100% determinism in telecom-grade services, causing the model to produce semantic illusions that deviate from the legally defined communication state when processing 5G signaling interactions with strict protocol boundaries.

[0003] Besides the limitations of gateway physical throughput and latency control, semantic recognition logic control methods also face technical bottlenecks. For example, Chinese invention patent application CN116521822A discloses a user intent recognition method and device based on a 5G message multi-turn session mechanism. This method constructs an initial profile based on historical interaction behavior and iteratively converges the profile intent by receiving progressive or negative correction text from the user. Analysis of the underlying logic reveals that this mechanism relies on the premise that the user actively provides feedback during multi-turn interactions. However, in a telecom core network production environment, gateway devices require instantaneous determinism in signaling parsing, needing to provide accurate routing instructions upon the arrival of the first message. Most automated business scenarios do not allow redundant multi-turn interaction costs. Relying on post-hoc correction rather than prior control for iterative identification paths leads to logical lag in high-concurrency, single-triggered telecom-grade service flows, failing to eradicate the generative model's first-turn decoding principle from a mathematical perspective. Sexual illusions; to suppress the aforementioned semantic illusions, the industry's conventional improvement approach usually uses front-end instruction fine-tuning or post-regularization rule interception. However, this is a logical correction at the model output end and does not touch the generation mechanism of the decoding layer. If the model parameter scale is increased or an enhancement path of external large-scale knowledge graph is introduced, although the parsing accuracy can be improved in a statistical sense, it will cause the system to generate high computing power overhead in the inference stage, resulting in non-linear growth of processing latency, which violates the low latency physical characteristics of 5G core network gateway equipment. If illegal intents are blocked simply by hard-coding interception rules, it will increase the maintenance complexity of the gateway system and cannot eliminate the ineffective dissipation of computing power inside the model on the path of generating illegal intents. This contradiction between the deep-seated bottleneck of intent parsing reliability and the efficiency of inference resource utilization means that generative artificial intelligence always faces the potential risk of routing failure due to logical divergence in strongly constrained telecommunications-grade communication network applications.

[0004] Therefore, how to construct a method that can transform telecommunications-grade service specifications into a probabilistic boundary intervention mechanism for the model decoding layer, thereby ensuring that intent generation absolutely obeys the deterministic constraints of the protocol state machine while maintaining deep semantic understanding capabilities, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A semantic processing method for 5G message parsing, comprising the following steps: Step S1: Obtain the 5G message text sequence; Step S2: Input the 5G message text sequence into the generative language model, extract semantic features through the forward propagation network of the generative language model, and obtain the hidden layer state tensor representing the current dialogue context. Step S3: At the preset depth level of the generative language model, calculate the information entropy of the probability distribution of the current generation. The information entropy is determined by the distribution dispersion after the current hidden layer output vector is mapped to the intent word space. The information entropy is used to quantify the semantic uncertainty of the generative language model in parsing the current text. Step S4: When the information entropy is determined to be lower than the preset information threshold, stop the calculation of subsequent layers of the forward propagation network and extract the hidden layer state tensor at the current depth. Step S5: Obtain the 5G protocol state identifier associated with the current session, and determine the set of legal intents that are allowed to jump in the current protocol state by retrieving the preset protocol state transition topology. Step S6: Map the extracted hidden layer state tensor to the intent vocabulary space through a linear layer, and calculate the initial log probability tensor of each candidate intent. Step S7: Construct a mask tensor with the same dimension as the initial log probability tensor, use the mask tensor to perform log probability space modulation on the initial log probability tensor, and calculate the normalized target intent probability distribution sequence. Step S8: Extract the intent index corresponding to the item with the highest probability value from the target intent probability distribution sequence to determine the target parsing intent.

[0006] Preferably, step S2 specifically includes the following sub-steps: step S21, tokenizing the 5G message text sequence to generate an input embedding vector; step S22, inputting the input embedding vector into the self-attention network of the generative language model; step S23, collecting neuron activation state data at a preset depth level to construct the hidden layer state tensor.

[0007] Preferably, the process of calculating information entropy in step S3 includes: step S31, calculating the negative log-likelihood expectation value after mapping the hidden layer state tensor of the current depth to the intent vocabulary space; step S32, determining the negative log-likelihood expectation value as information entropy.

[0008] Preferably, the generative language model adopts an attention bias mechanism that introduces a time decay coefficient, including the following steps: extracting feature vector data of associated attachments in the 5G message text sequence; when calculating attention weights, adjusting the weight ratio between the feature vector data and the current message node according to the time difference between the current processing time and the time when the associated attachments are generated; and using the adjusted weight ratio to enhance the hidden layer state tensor to complete feature association in asynchronous message scenarios.

[0009] Preferably, step S5, which determines the set of legitimate intents, includes: step S51, parsing the signaling status field in the 5G protocol status identifier; step S52, retrieving downstream candidate nodes that are logically related to the signaling status field in the protocol state transition topology; and step S53, defining the service instructions corresponding to the downstream candidate nodes as the set of legitimate intents.

[0010] Preferably, the method further includes a logical degradation step based on the protocol synchronization state: step S61, monitoring the update cycle of the 5G protocol state identifier; step S62, when the update cycle exceeds a preset duration threshold, switching the mask tensor from a hard constraint mask to a soft penalty tensor; step S63, using the soft penalty tensor to reduce the calculation weight of illegal intent in the initial logarithmic probability tensor, and retaining a non-zero proportion of redundant generation probability to handle signaling state lag conditions.

[0011] Preferably, step S8 specifically includes the following sub-steps: step S81, extracting the intent index corresponding to the item with the highest probability value from the target intent probability distribution sequence; step S82, retrieving the intent map table according to the intent index, and outputting the routing instruction representing the target parsing intent.

[0012] Preferably, the preset depth level of the forward propagation network is set at 2 / 3 of the total depth of the model.

[0013] Preferably, the method further includes an intent routing feedback correction step: real-time collection of routing response data from the service gateway to the target's parsed intent; when it is determined that the routing response data indicates an incorrect intent direction, the preset information threshold in step S4 is increased to increase the proportion of layers involved in the calculation by the forward propagation network.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. In the semantic processing of 5G message parsing, the processor introduces a dynamic mask tensor associated with the communication protocol state topology at the end of the autoregressive decoding process. This reduces the macroscopic telecommunications service rules to microscopic logarithmic probability space hard constraints within the model. Essentially, it constructs a controlled natural language generation and decoding mechanism, achieving mathematical consistency between the semantic parsing results and the logical topology. This mechanism ensures that the generated service intent is mathematically locked within a legal state transition path, blocking signaling routing failures caused by nondeterministic probability divergence at the source. Without changing the model parameters, it resolves the logical conflict between deep semantic understanding capabilities and protocol flow rigor based on the deterministic telecommunications-grade state machine properties of the generative architecture.

[0015] 2. The processor collaborative application uses a dynamic truncation mechanism based on semantic entropy and a state space reshaping method to achieve adaptive matching between computing resources and business complexity. By monitoring the probability distribution dispersion in real time in the hidden layer of the model, the short instructions that appear frequently trigger early exit during the rapid feature convergence stage. The mask reset logic is connected to ensure output compliance. This cross-level linkage design not only compresses the inference latency of large models, but also ensures that shallow features can still be forcibly guided by physical state boundaries even without subsequent network layer refinement, thereby improving the system's processing efficiency under high-concurrency telecommunications conditions.

[0016] 3. The processor utilizes attention bias intervention with the introduction of a time decay coefficient, combined with end probability distribution reconstruction, to solve the problem of referential resolution in cross-modal asynchronous delivery scenarios. By artificially adjusting the association weight between multimedia feature nodes and the current timestamp during the self-attention weight calculation process, the hidden state vector has completed a strong association with physical attachment features before generating the initial probability distribution. This feature enhancement process provides a high-quality prior probability foundation for subsequent log probability modulation, making the controlled probability distribution sequence converge more naturally to the real business intent, and avoiding semantic parsing deviations caused by temporal shifts in multi-turn dialogues. Attached Figure Description

[0017] Figure 1 This is a flowchart of the 5G message semantic parsing and intent recognition method of the present invention; Figure 2 This is a diagram of the semantic processing logic architecture for the state machine constraints of the protocol of this invention.

[0018] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A semantic processing method for 5G message parsing includes the following steps: Step S1: Obtain the 5G message text sequence; Step S2: Input the 5G message text sequence into the generative language model, extract semantic features through the forward propagation network of the generative language model, and obtain the hidden layer state tensor representing the current dialogue context. Step S3: At the preset depth level of the generative language model, calculate the information entropy of the probability distribution of the current generation. The information entropy is determined by the distribution dispersion after the current hidden layer output vector is mapped to the intent word space. The information entropy is used to quantify the semantic uncertainty of the generative language model in parsing the current text. Step S4: When the information entropy is determined to be lower than the preset information threshold, stop the calculation of subsequent layers of the forward propagation network and extract the hidden layer state tensor at the current depth. Step S5: Obtain the 5G protocol state identifier associated with the current session, and determine the set of legal intents that are allowed to jump in the current protocol state by retrieving the preset protocol state transition topology. Step S6: Map the extracted hidden layer state tensor to the intent vocabulary space through a linear layer, and calculate the initial log probability tensor of each candidate intent. Step S7: Construct a mask tensor with the same dimension as the initial log probability tensor, use the mask tensor to perform log probability space modulation on the initial log probability tensor, and calculate the normalized target intent probability distribution sequence. Step S8: Extract the intent index corresponding to the item with the highest probability value from the target intent probability distribution sequence to determine the target parsing intent.

[0021] Preferably, step S2 specifically includes the following sub-steps: step S21, tokenizing the 5G message text sequence to generate an input embedding vector; step S22, inputting the input embedding vector into the self-attention network of the generative language model; step S23, collecting neuron activation state data at a preset depth level to construct the hidden layer state tensor.

[0022] Preferably, the process of calculating information entropy in step S3 includes: step S31, calculating the negative log-likelihood expectation value after mapping the hidden layer state tensor of the current depth to the intent vocabulary space; step S32, determining the negative log-likelihood expectation value as information entropy.

[0023] Preferably, the generative language model adopts an attention bias mechanism that introduces a time decay coefficient, including the following steps: extracting feature vector data of associated attachments in the 5G message text sequence; when calculating attention weights, adjusting the weight ratio between the feature vector data and the current message node according to the time difference between the current processing time and the time when the associated attachments are generated; and using the adjusted weight ratio to enhance the hidden layer state tensor to complete feature association in asynchronous message scenarios.

[0024] Preferably, step S5, which determines the set of legitimate intents, includes: step S51, parsing the signaling status field in the 5G protocol status identifier; step S52, retrieving downstream candidate nodes that are logically related to the signaling status field in the protocol state transition topology; and step S53, defining the service instructions corresponding to the downstream candidate nodes as the set of legitimate intents.

[0025] Preferably, the method further includes a logical degradation step based on the protocol synchronization state: step S61, monitoring the update cycle of the 5G protocol state identifier; step S62, when the update cycle exceeds a preset duration threshold, switching the mask tensor from a hard constraint mask to a soft penalty tensor; step S63, using the soft penalty tensor to reduce the calculation weight of illegal intent in the initial logarithmic probability tensor, and retaining a non-zero proportion of redundant generation probability to handle signaling state lag conditions.

[0026] Preferably, step S8 specifically includes the following sub-steps: step S81, extracting the intent index corresponding to the item with the highest probability value from the target intent probability distribution sequence; step S82, retrieving the intent map table according to the intent index, and outputting the routing instruction representing the target parsing intent.

[0027] Preferably, the preset depth level of the forward propagation network is set at 2 / 3 of the total depth of the model.

[0028] Preferably, the method further includes an intent routing feedback correction step: real-time collection of routing response data from the service gateway to the target's parsed intent; when it is determined that the routing response data indicates an incorrect intent direction, the preset information threshold in step S4 is increased to increase the proportion of layers involved in the calculation by the forward propagation network.

[0029] Example 1: In a 5G rich media message parsing scenario deployed in a financial industry gateway, the system faces real-time service flows with 10,000,000 concurrent requests and multi-turn interaction characteristics. Because the 5G session protocol requires the parsed service intent to absolutely conform to the statutory state transition topology constructed by the SIP signaling state machine, and when using a large language model based on an autoregressive mechanism to perform semantic parsing, the probability distribution of its decoding layer in the open vocabulary space exhibits uncertain divergence. This non-deterministic generation characteristic conflicts with the rigid constraints of deterministic routing in telecom-grade services. Consequently, when processing sessions in an unregistered state, the large language model, influenced by the historical long text context, may generate intent indexes that are illegal in the current protocol state, such as balance inquiries or transfer requests, leading to logical interruptions or routing failures in the signaling gateway. To address the signaling conflict challenges caused by the uncertainty of semantic generation, the semantic processing method provided by this invention runs in a processor. The processor receives the 5G message text sequence to be parsed and inputs it into a pre-trained generative language model. A feedforward neural network extracts semantic features. During the forward propagation of the generative language model, to meet the requirements of low latency at the telecom level, the processor performs information entropy monitoring at a preset depth level in the model. The preset depth level is set at 2 / 3 of the total number of layers in the model. The processor calculates the negative log-likelihood expectation value after mapping the current hidden layer state tensor to the intent vocabulary space, and uses it as the information entropy to measure semantic uncertainty. When the calculated information entropy is lower than the preset confidence threshold, it indicates that the current features are sufficient to support intent determination. The processor then interrupts the operation of subsequent neuron layers to achieve dynamic truncation of inference resources. The above physical operation process essentially maps the semantic processing logic architecture under the constraint of the protocol state machine. By monitoring the semantic information entropy to characterize the degree of feature convergence, dynamic truncation of inference resources is performed to ensure low latency. In subsequent processing of illegal intents and abnormal conditions, the system will sequentially complete the operation of determining the set of legal intents and performing log probability space modulation in the log probability space, and perform logical degradation design in case of anomalies. Finally, the confidence threshold is corrected based on the actual routing response feedback from the gateway.

[0030] After obtaining the hidden layer state tensor representing the current dialogue context, the processor uses a linear layer to map it to a preset intent vocabulary space, generating an unnormalized initial log probability tensor. Simultaneously, the processor reads the 5G protocol status identifier of the current user terminal maintained in the communication gateway device, retrieves the preset protocol state transition topology, determines the set of legal intents allowed to jump under the current protocol state, and constructs the dimension and initial log-probability tensor. Consistent mask tensor In this mask tensor, the element at the index corresponding to the set of legitimate intentions is assigned a value of 0, while the element at the index corresponding to the set of illegitimate intentions is assigned a preset negative penalty scalar. By performing element-wise addition, the initial logarithmic probability tensor is... With mask tensor Perform log-probability space modulation to generate a modulated log-probability tensor. Its mathematical expression is ,in, This is the modulated logarithmic probability tensor. The original log-probability tensor output by the linear layer. For the mask tensor generated based on the protocol state, by introducing hard constraints based on a physical state machine at the decoding end, the probability distribution that diverges across the full vocabulary is converged to a subset of legitimate intentions. The processor then modulates the logarithmic probability tensor. Normalization is performed to generate a reshaped target intent probability distribution sequence, and the target business intent index is determined according to the maximum probability search strategy, thereby outputting a routing instruction representing the target parsing intent. In this way, even if the generative language model tends to generate illegal intents in terms of statistical probability, the illegal index position is given a large negative bias, making the normalized generation probability approach 0, thus shielding the semantic output that deviates from the 5G protocol specification at the mathematical source.

[0031] The synergistic operation of the aforementioned technical features enables the system to maintain deep semantic understanding capabilities to handle complex cross-modal referential relationships. Based on a generative language model and determined carrier-grade signaling state machine attributes, experimental data shows that after introducing the mask reshaping mechanism of this scheme, the gateway system achieves a 0% intent violation rate in 1,000,000 consecutive parsing tasks. Furthermore, combined with a dynamic truncation mechanism based on information entropy, the average inference latency for a single message parsing is reduced from 150ms to 45ms, mitigating the risk of core network routing failures caused by semantic biases, improving system throughput under high concurrency conditions, and enhancing the logical reliability of generative artificial intelligence in strongly constrained telecommunications scenarios. Simultaneously, the processor acquires the 5G message text sequence and extracts the 5G protocol state identifier of the current session by parsing the signaling field in the SIP protocol packet header. 5G protocol status identifier The logical position of the user terminal in the communication core network is represented by a preset 8-bit binary code. The processor performs a hash search in the memory database based on the binary code to determine that the current terminal is in an initial state that is registered but has not been authenticated for services. Then, it retrieves the list of legal jump intentions corresponding to the initial state from the preset protocol state transition topology.

[0032] Example 2: To verify the accuracy and response latency of the semantic processing method provided by this invention in 5G message parsing tasks, the processor invoked a simulation test environment, including a processor cluster and a memory database storing 5G protocol status identifiers. The data used in the experiment originated from 1,000,000 anonymized 5G rich media message logs from the communication gateway. A traffic generator simulated a concurrent operation of tens of millions of messages, continuously inputting message text sequences into the processor. To simulate electromagnetic interference and network jitter in a real communication environment, additive white Gaussian noise with a signal-to-noise ratio of 25dB was actively superimposed onto the input data stream, and a preset depth level was selected. With preset information threshold To control parameters, a preset depth level is set for a generative language model with a total of 12 layers. The design principle is based on the balance between semantic feature extraction depth and computational load. When the number of layers is too shallow, the forward propagation network cannot capture cross-modal referential relationships, resulting in insufficient semantic representation ability of the hidden layer state tensors. When the number of layers is too deep, the truncation mechanism tends to saturate in reducing latency. Therefore, the design principle is chosen based on the following: 8 layers; preset signal threshold The selection criteria are based on a trade-off between judgment confidence and inference speed. The criteria are determined by statistical analysis of the semantic entropy distribution of 50,000 samples. It is 0.12.

[0033] The experimental design established three control groups to verify the synergistic effect between features. The present invention's control group employed a complete scheme including a dynamic truncation mechanism and a mask tensor modulation mechanism. Control group A removed the mask tensor modulation mechanism and used only the raw decoding output of the generative language model. Control group B removed the dynamic truncation mechanism and performed full-depth inference operations on all messages. Control group C used a pre-set threshold. The parameter was set to 0.9 to verify the performance degradation after the parameter exceeded the optimal working window. During the baseline setting phase, the system input a message sequence containing a business processing request: "I want to process a cross-border large-amount transfer." At this time, the current user terminal 5G protocol status recorded by the communication gateway was: non-real-name status. During the core derivation phase, the prototype of this invention extracted the hidden layer state tensor at the 8th hidden layer and mapped it to the preset intent vocabulary space to generate an initial log-probability tensor. The original score for the intent index [cross-border transfer] is 15.62. At this point, the processor searches the protocol state transition topology and determines that the set of legal intents allowed to jump in the non-real-name state does not include transfer-type operations. The processor then constructs a mask tensor. Assign the value -10000.00 to the element at the corresponding [cross-border transfer] index position, and calculate the modulation logarithm probability tensor. .

[0034] In the value closed-loop verification stage, the target parsing intent output by the sample group of this invention was [account inquiry], with a parsing accuracy of 99.51% and an average inference latency of 42.83ms per message parsing. In contrast, the parsing result generated by control group A was [cross-border transfer], and the intent recognition result conflicted with the protocol state, causing the recognition accuracy to drop to 82.34%. The parsing accuracy of control group B was 99.62%, but the average inference latency increased to 148.56ms. Due to the excessively high preset threshold, control group C triggered dynamic truncation before the semantic features converged, and the parsing accuracy dropped to 75.24%. This indicates that when the parameters deviate from the range defined by this invention, the system cannot balance parsing determinism and real-time performance. The experimental data proves that the synergistic effect of the mask tensor modulation mechanism and the dynamic truncation mechanism maintains low latency at the telecommunications level while ensuring that the semantic parsing result conforms to the constraint logic of the physical state machine.

[0035] Example 3: In a financial RCS message gateway system supporting 10,000,000 active users, the processor needs to perform semantic parsing on business intentions with high security requirements, such as credit card activation, large installment applications, and overseas credit limit adjustments. When the system receives a message sequence containing sensitive financial terms, the autoregressive decoding mechanism of the generative language model is essentially based on statistical fitting of massive general corpora. In the absence of physical constraints, it is very easy to generate intention illusions that deviate from the current user account state. For example, when a user is in the process of closing an account, the model may still parse an installment application intention due to the strong correlation of historical dialogues. This non-deterministic generative logic conflicts with the deterministic admission criteria based on protocol state machines in financial business, causing the gateway to face intention misjudgment or transaction routing failures. To address the need for transparency in the information entropy calculation process, the processor obtains the state tensor of the intermediate hidden layer of the generative language model. The input is then fed into a linear transformation layer, where matrix-vector product operations map the feature dimensions to a dimension consistent with the intent vocabulary space. The system obtains the original scores for each intent index. The linear transformation layer uses a pre-set auxiliary mapping parameter matrix. During the offline model computation phase, the system calculates the cross-entropy loss between the output distribution of the last neuron level and the current preset depth level, takes the weight values ​​when the loss is minimized, and independently calibrates the auxiliary mapping parameter matrix to ensure that the semantic representation dimension of the intermediate hidden layer is physically aligned with the underlying intent space. To prevent feature dimension collapse when directly mapping high-dimensional hidden states to low-dimensional space, the system uses a two-layer network containing a non-linear activation function layer and a fully connected dimensionality reduction layer as the physical mapping structure. It extracts massive amounts of desensitized real 5G signaling interaction data as basic training data. Through this structure, it learns the key dimensionality reduction weight logic for a smooth transition from high-dimensional dense vectors to discrete intent spaces, thus ensuring that the feature mapping of each dimension has a deterministic physical derivation path. Before inputting into the linear transformation layer, based on the principle of natural exponential decay in dynamic systems, the processor extracts the associated attachment feature vector data from the 5G message text sequence according to the formula... Determine the time decay coefficient, where, Characterized by a dimensionless decay scalar used for weighting. Characterizes the preset time decay constant, taking the value of a real number in the range (0,1]. Representing the physical time difference between the current processing moment and the moment the associated attachments were generated, before performing the addition operation, considering the dimensionality difference between the multimedia attachments and the text tensor at the structural scale, the processor performs a linear projection transformation (dimensionality reduction or dimensionality increase) on the extracted associated attachment feature vector data through a preset multilayer perceptron structure, uniformly mapping them to the state tensor. The same 512-dimensional continuous feature space ensures that both have an aligned basis for performing feature enhancement. The processor will extract the feature vector data and multiply it by a decay scalar. and the state tensor Element-wise addition is performed, and the state tensor after feature enhancement is input into the linear transformation layer. The processor then performs normalization on the original scores to obtain the probability distribution value corresponding to each intent index at the current depth. ,in Given the intent index number and i∈[1,V], the processor calculates the distribution dispersion of the current hidden layer output vector projected onto the preset intent vocabulary space, i.e., the information entropy. Information entropy The calculation method is through formula Expression, among which, For information entropy, For the intention vocabulary space dimension, For the first The probability distribution value of an intent index, in this mathematical unit, the combination of the negative sign and the natural logarithm operation makes the probability distribution flatter. The larger the value, the greater the uncertainty of the semantic parsing result. When the processor detects... When the signal drops below the preset threshold of 0.12, the semantic features are determined to have reached extreme convergence, and the processor interrupts the operation of the subsequent four layers of the feedforward neural network.

[0036] The processor reads the 5G protocol status identifier associated with the current session from the memory database. Construct a mapping matrix based on the pre-defined statutory state transition topology. In this matrix, elements corresponding to valid state transition paths are assigned a value of 1, and invalid paths are assigned a value of 0. The processor then compares the current intent index with... Combined, the set of legal intents allowed to jump under the current protocol state is selected, targeting the mask tensor. Punishing negative scalar For calibration, the processor uses an offline fitting method to analyze the probability drift curves under 10,000 sets of extreme noise samples to determine when... When set to -10000.00, the normalized output probability of an illegal intent, after modulation in the logarithmic probability space, is suppressed to... Below the order of magnitude, this physically blocks the possible output of illegal semantic intent, aligns the generation probability space with the telecommunications protocol space, and allows the processor to monitor 5G protocol status indicators in parallel during continuous system processing. The update cycle, based on a signal transmission degradation model, determines that a communication node is in a signaling lag abnormal state when the update cycle exceeds a preset duration threshold of 200 milliseconds. The processor then processes this abnormal state according to the formula... Constructing a soft penalty tensor ,in, The log-probability bias tensor characterizes abnormal communication states. The dimensionless penalty scaling factor is defined, and its legal value range is limited to 1. The processor will use the mask tensor Replace with soft penalty tensor To address asynchronous operating conditions, the system retains non-zero proportional redundancy in the logarithmic probability space outside the valid baseline path for generating probabilities. It also collects real-time routing response data from the business gateway regarding target parsing intent. Based on the principle of an error negative feedback control system, when the routing response data contains an intent-pointing error feedback code, the processor uses a formula... Update the preset information threshold, where, The newly pre-set information threshold obtained from the characterization calculation has a physical upper limit constraint of 0.5. A threshold is preset before characterization updates. The gain step size is characterized by a dimensionless constant of 1 and a fixed value of 0.05. The system represents the number of intent-directed error feedbacks accumulated within the continuous statistical algorithm window. To ensure that the increased confidence threshold parameter truly achieves the design goal of increasing the number of computational layers, the system internally uses a negative correlation mapping mechanism to inversely bind the absolute value of this threshold to its underlying physical triggering conditions. The increase in the numerical value is internally translated by the physical controller into a proportional decrease in the actual logarithmic boundary value used for truncation decisions. This causes the information entropy of the model features to converge to a more stringent low level before triggering the truncation condition. Consequently, the forward propagation network is forced to satisfy the decision at a deeper level. This logically and consistently implements the function of increasing the proportion of layers involved in the calculation. The processor changes the preset information threshold to... To increase the number of neural network layers involved in subsequent computations via the forward propagation network, this mask modulation mechanism based on protocol state machine mapping reduces the dimensionality of the language model output into a controlled, carrier-grade instruction stream. According to experimental replay data, the introduction of the aforementioned offline calibration-based... After parameterization, the financial gateway maintains a routing accuracy of 100.00% in intent conflict scenarios, and further improves this accuracy by adjusting information entropy. The monitoring and inference resources are truncated, and the processor's computing throughput is increased by 2.85 times. This method solves the threat of semantic illusion to the signaling security of 5G core network by embedding a mask interference quantizer based on quantization calibration at the decoding end without adjusting the underlying weights of the language model. It also improves the engineering reliability of generative artificial intelligence in strongly constrained financial scenarios. The processor performs tokenization processing on the 5G message text sequence, and uses byte-pair encoding to split the original character stream into discrete sub-word index sequences. By looking up the preset word vector space matrix, each sub-word index is converted into a 512-dimensional dense input embedding vector. The processor inputs the input embedding vector into a generative language model containing 12 self-attention blocks. Multi-head attention operation and residual connection are performed inside each self-attention block to generate a hidden layer state tensor that represents the semantic features of the current text.

[0037] Example 4: In a semantic parsing node deployment scenario with cross-industry multi-gateway access, the processor invokes the parameter calibration procedure, reads historical message text sequences within a preset time window as calibration samples, and uses a feedforward neural network of a generative language model to calculate the probability distribution sequence of each calibration sample projected onto the intent vocabulary space. For each probability distribution sequence, the processor uses the formula... Calculate the entropy point, where, For information entropy, For the intention vocabulary space dimension, For the first The processor uses the probability distribution values ​​of each intent index to apply a mean-shift clustering algorithm to the resulting set of entropy points, identifying the entropy cluster centers corresponding to the definite intent determination state, and using these cluster centers as a pre-set confidence threshold. Based on the benchmark value, the processor synchronously analyzes the energy distribution of the illegal intent index in the logarithmic probability space. By calculating the difference between the mean of the legal intent and the maximum value of the illegal intent, a negative scalar penalty is determined by selecting a value on the order of 10 of this difference. The minimum value.

[0038] When the system faces a telecommunications protocol version change, the processor invokes the state machine mapping procedure to update the mask generation logic. The processor parses the state transition topology in the 5G signaling protocol definition file, maps the logical state identifiers defined in the state transition topology to the corresponding memory address indices, and generates a Boolean matrix containing legal jump intentions based on the state transition relationships. During the system initialization phase, the processor loads the Boolean matrix into the memory database and establishes an index association with the output vocabulary of the generative language model. During system operation, the processor uses the received current user terminal 5G protocol state identifiers... The corresponding Boolean mask row vector is retrieved from the memory database, and the zero elements in the Boolean mask row vector are converted into penalty negative scalars. Generate mask tensor .

[0039] Example 5: In a 5G rich media message semantic processing system for cross-regional distributed deployment, the processor invokes the environment consistency calibration procedure. The processor reads 50,000 service parsing logs within a preset time window as a calibration benchmark data stream. This benchmark data stream is input into the feedforward neural network of the generative language model. The processor collects the distribution dispersion data of the output vector of the 8th hidden layer projected onto the intent vocabulary space. The processor extracts the information entropy distribution curve of each calibration sample when the intent judgment is at the convergence critical state. The preset information threshold is determined by calculating the extreme point of the first derivative of this information entropy distribution curve. During the calibration process, the processor detected that when When the value is 0.12, the convergence rate of the semantic features is above 99.90%, accompanied by the inflection point of the decrease in inference latency.

[0040] To determine the value of the negative scalar penalty in the mask tensor, the processor establishes a penalty function fitting model. The processor obtains the minimum value of the initial logarithmic probability distribution in the set of legitimate intentions and simultaneously obtains the maximum value of the initial logarithmic probability distribution in the set of illegitimate intentions. The processor then applies the formula... Calculate the negative scalar of the penalty ,in, To punish negative scalars, This is the preset resolution accuracy tolerance term. This represents the initial logarithmic probability maximum of an illegal intent. The processor will calculate the initial logarithmic probability minimum of the legitimate intent. The configuration fields written to the in-memory database are based on the received current user terminal 5G protocol status identifier. Retrieve the Boolean mask matrix, Injected into mask tensor The non-zero bits suppress the probability of generating illegitimate intent after log-space modulation. Within this range, the alignment of the generated probability distribution with the protocol state machine logic is completed, and this analytical precision tolerance term... The specific value is determined based on the physical throughput and fault tolerance boundary tests of the telecommunications core network. The system conducts extreme loading experiments on routing fault boundaries under tens of millions of concurrent operations, determining that when the probability of generating an illegal intent exceeds one in one hundred million, there is a high probability of triggering a gateway routing congestion avalanche. Therefore, [the value is determined]. The upper limit of the project is defined as This provides a benchmark for quantifying the reliability of reconstruction.

[0041] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A semantic processing method for 5G message parsing, characterized in that, Includes the following steps: Step S1: Obtain the 5G message text sequence; Step S2: Input the 5G message text sequence into the generative language model, extract semantic features through the forward propagation network of the generative language model, and obtain the hidden layer state tensor representing the current dialogue context. Step S3: At the preset depth level of the generative language model, calculate the information entropy of the probability distribution of the current generation. The information entropy is determined by the distribution dispersion after the current hidden layer output vector is mapped to the intent word space. The information entropy is used to quantify the semantic uncertainty of the generative language model in parsing the current text. Step S4: When the information entropy is determined to be lower than the preset information threshold, stop the calculation of subsequent layers of the forward propagation network and extract the hidden layer state tensor at the current depth. Step S5: Obtain the 5G protocol state identifier associated with the current session, and determine the set of legal intents that are allowed to jump in the current protocol state by retrieving the preset protocol state transition topology. Step S6: Map the extracted hidden layer state tensor to the intent vocabulary space through a linear layer, and calculate the initial log probability tensor of each candidate intent. Step S7: Construct a mask tensor with the same dimension as the initial log probability tensor, use the mask tensor to perform log probability space modulation on the initial log probability tensor, and calculate the normalized target intent probability distribution sequence. Step S8: Extract the intent index corresponding to the item with the highest probability value from the target intent probability distribution sequence to determine the target parsing intent. 2.The semantic processing method for 5G message parsing of claim 1, wherein, Step S2 specifically includes the following sub-steps: Step S21, tokenize the 5G message text sequence to generate an input embedding vector; Step S22, input the input embedding vector into the self-attention network of the generative language model; Step S23: Collect neuron activation state data at a preset depth level to construct the hidden layer state tensor.

3. The semantic processing method for 5G message parsing according to claim 1, characterized in that, The process of calculating information entropy in step S3 includes: step S31, calculating the negative log-likelihood expectation after mapping the hidden layer state tensor of the current depth to the intent word space; step S32, determining the negative log-likelihood expectation as the information entropy.

4. The semantic processing method for 5G message parsing according to claim 1, characterized in that, The generative language model employs an attention bias mechanism that introduces a time decay coefficient, including the following steps: extracting feature vector data of associated attachments in the 5G message text sequence; when calculating attention weights, adjusting the weight ratio between the feature vector data and the current message node based on the time difference between the current processing time and the time when the associated attachments were generated; and using the adjusted weight ratio to enhance the hidden layer state tensor to complete feature association in asynchronous message scenarios.

5. A semantic processing method for 5G message parsing according to claim 1, characterized in that, The process of determining the set of legitimate intents in step S5 includes: step S51, parsing the signaling status field in the 5G protocol status identifier; step S52, retrieving downstream candidate nodes that are logically related to the signaling status field in the protocol state transition topology; and step S53, defining the service instructions corresponding to the downstream candidate nodes as the set of legitimate intents.

6. A semantic processing method for 5G message parsing according to claim 1, characterized in that, The method also includes a logical degradation step based on the protocol synchronization state: step S61, monitor the update cycle of the 5G protocol state identifier; step S62, when the update cycle exceeds the preset duration threshold, switch the mask tensor from a hard constraint mask to a soft penalty tensor; step S63, use the soft penalty tensor to reduce the calculation weight of illegal intent in the initial log probability tensor, and retain a non-zero proportion of redundant generation probability to handle signaling state lag conditions.

7. A semantic processing method for 5G message parsing according to claim 1, characterized in that, Step S8 specifically includes the following sub-steps: Step S81, extract the intent index corresponding to the item with the highest probability value in the target intent probability distribution sequence; Step S82, retrieve the intent map table according to the intent index, and output the routing instruction representing the target parsing intent.

8. A semantic processing method for 5G message parsing according to claim 1, characterized in that, The preset depth level of the forward propagation network is set at 2 / 3 of the total depth of the model.

9. A semantic processing method for 5G message parsing according to claim 1, characterized in that, The method also includes an intent routing feedback correction step: real-time collection of routing response data from the business gateway in response to the target's parsed intent; If the routing response data indicates an incorrect intended direction, the preset information threshold in step S4 is increased to increase the proportion of layers involved in the forward propagation network calculation.

Images