Data processing method, communication device and computer program product

Abstract

Embodiments of the present application provide a data processing method, a communication device and a computer program product, wherein the method comprises: obtaining multi-modal data based on a session initiated by a terminal, and extracting multi-modal features based on the multi-modal data; analyzing SIP signaling to obtain a structured signaling context, and generating an intent anchor point vector based on the structured signaling context; determining a unified intent vector based on the multi-modal features and the intent anchor point vector; and processing the multi-modal data based on the unified intent vector. Through the embodiments of the present application, the problem of multi-modal semantic fragmentation and alignment deviation in related technologies, which leads to instruction coverage, repeated response or intent understanding error, can be solved. The intent anchor point vector is generated based on the structured signaling context obtained from the SIP signaling, and the output unified intent vector is a fusion representation of multi-modal data and signaling context, which realizes precise semantic convergence of heterogeneous multi-modal features and improves the accuracy of intent understanding.

Description

Technical Field

[0001] This application relates to the field of communications, and more specifically, to a data processing method, a communication device, and a computer program product. Background Technology

[0002] Existing technologies suffer from architectural limitations in areas such as memory collaboration, signaling semantic utilization, intent alignment arbitration, and state-aware routing. These limitations stem from the traditional design paradigms of "decoupling signaling from services," "independent multimodal parsing," and "static rule scheduling," which cannot guarantee low-latency real-time interaction while achieving cross-domain knowledge consistency, accurate disambiguation of complex intents, and carrier-grade service reliability. Multimodal data such as voice, video, and text are typically parsed separately by independent perception models, lacking a unified semantic convergence benchmark. When users express complex needs through "voice + gesture + text," the parsing results from each channel deviate in temporal sequence and semantic space, leading to instruction overriding, repetitive responses, or errors in intent understanding. Summary of the Invention

[0003] This application provides a data processing method, communication device, and computer program product to at least solve the problems of multimodal semantic fragmentation and alignment deviation in related technologies, which lead to instruction overriding, duplicate responses, or incorrect intent understanding.

[0004] According to one embodiment of this application, a data processing method is provided, including:

[0005] Multimodal data is acquired based on the session initiated by the terminal, and multimodal features are extracted based on the multimodal data;

[0006] The Session Initiation Protocol (SIP) signaling is parsed to obtain a structured signaling context, and an intent anchor vector is generated based on the structured signaling context.

[0007] A unified intent vector is determined based on the multimodal features and the intent anchor vector;

[0008] The multimodal data is processed based on the unified intent vector.

[0009] According to yet another embodiment of this application, a network device is also provided, the network device including a receiver, a transmitter, and a processor, the network device being configured to perform the steps of any of the above method embodiments via at least one of the receiver, the transmitter, and the processor.

[0010] According to yet another embodiment of this application, a computer-readable storage medium is also provided, wherein a computer program is stored therein, and the computer program is configured to perform the steps in any of the above method embodiments when it is run.

[0011] According to yet another embodiment of this application, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0012] According to yet another embodiment of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0013] The above embodiments of this application can solve the problems of multimodal semantic fragmentation and alignment deviation in related technologies, which lead to instruction overriding, repeated responses, or incorrect intent understanding. Based on the structured signaling context obtained from SIP signaling, an intent anchor vector is generated and used as a positive sample anchor in multimodal contrastive learning. The output unified intent vector is a fusion representation of multimodal data and signaling context, realizing accurate semantic convergence of heterogeneous multimodal features and improving the accuracy of intent understanding. Attached Figure Description

[0014] Figure 1 This is a flowchart of a data processing method according to an embodiment of this application;

[0015] Figure 2 This is a flow chart of a data processing method according to an optional embodiment of this application. Figure 1 ;

[0016] Figure 3 This is a flow chart of a data processing method according to an optional embodiment of this application. Figure 2 ;

[0017] Figure 4 This is a system framework diagram of 5G new call multi-agent hierarchical memory collaboration and intent alignment according to an embodiment of this application;

[0018] Figure 5 This is a flowchart of 5G new call multi-agent hierarchical memory collaboration and intent alignment according to an embodiment of this application;

[0019] Figure 6 This is a flowchart of dynamic arbitration according to an embodiment of this application;

[0020] Figure 7 This is a flowchart of hierarchical memory synchronization and arbitration based on IMS signaling event-driven implementation according to an embodiment of this application;

[0021] Figure 8 This is a flowchart illustrating multimodal intent alignment based on a UMF media pipeline according to an embodiment of this application;

[0022] Figure 9 This is a structural block diagram of a network device according to an embodiment of this application. Detailed Implementation

[0023] The embodiments of this application will be described in detail below with reference to the accompanying drawings and examples.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0025] This embodiment provides a data processing method. Figure 1 This is a flowchart of a data processing method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps:

[0026] Step S102: Obtain multimodal data based on the session initiated by the terminal, and extract multimodal features based on the multimodal data;

[0027] In step S102 above, media streams and data channels can be obtained based on the session initiated by the terminal. Multimodal feature extraction is to extract features from multimodal data such as voice, video, and text transmitted from the terminal.

[0028] In an exemplary embodiment, step S102 may include: extracting speech features, video features, and text features from the multimodal data to obtain multimodal features.

[0029] Step S104: Parse the Session Initiation Protocol (SIP) signaling to obtain the structured signaling context, and generate an intent anchor vector based on the structured signaling context;

[0030] Step S104 above involves SIP signaling parsing and anchor point generation. The SIP signaling is parsed to obtain the structured context, and an intent anchor point vector is generated from it. This transforms control plane signaling in related technologies into semantic vectors that AI can understand.

[0031] Step S106: Determine a unified intent vector based on multimodal features and intent anchor vector;

[0032] Step S108: Process the multimodal data based on the unified intent vector.

[0033] Through the above steps S102 to S108, the problems of multimodal semantic fragmentation and alignment deviation in related technologies, which lead to instruction overriding, repeated responses, or incorrect intent understanding, can be solved. Based on the structured signaling context obtained from SIP signaling, an intent anchor vector is generated and used as a positive sample anchor in multimodal contrastive learning. The output unified intent vector is a fusion representation of multimodal data and signaling context, realizing accurate semantic convergence of heterogeneous multimodal features and improving the accuracy of intent understanding.

[0034] Figure 2 This is a flow chart of a data processing method according to an optional embodiment of this application. Figure 1 ,like Figure 2 As shown, step S106 above may include:

[0035] Step S202: Using the intent anchor vector as the positive sample benchmark, a structured signaling context adaptive contrastive loss function is constructed based on the intent anchor vector, multimodal features, and signaling context confidence. The total loss value of the contrastive loss function is minimized iteratively through backpropagation, and a unified intent vector is obtained after convergence. (The signaling anchor vector is used as the basis for this step.) As a positive sample benchmark, a Signaling Context Adaptive Contrastive Alignment Loss (SIP-CACL) was constructed. The loss was minimized through backpropagation, the model parameters were iteratively optimized, and finally a unified intent vector was obtained by aggregation.

[0036] Figure 3 This is a flow chart of a data processing method according to an optional embodiment of this application. Figure 2 ,like Figure 3 As shown, in step S202 above, based on the intent anchor vector Multimodal features and signaling context confidence Constructing a signaling context adaptive contrastive loss function can include one of the following:

[0037] S301, constructing a signaling context adaptive contrastive loss function based on intent anchor vector, multimodal features, modality adaptive weights, and signaling context confidence; or

[0038] S302 determines the adaptive temperature coefficient based on the signaling context confidence and constructs the signaling context adaptive contrastive loss function based on the intent anchor vector, multimodal features and the adaptive temperature coefficient.

[0039] In step S302 above, the contrastive loss function can be calculated using the following formula:

[0040] ;

[0041] in, This represents the predicted modal uncertainty variance (positive constraint) output by the network in parallel. , The speech, video, and text feature vectors are obtained by mapping the multimodal features using a projection matrix. The regularization balance coefficient balances the weights of the negative contrastive example (NCE) loss term and the consistency regularization term. It is the loss term for a single mode.

[0042] In an exemplary embodiment, step S301 may further include one of the following:

[0043] S3011, calculate modality adaptive weights based on multimodal features and intent anchor vectors, and base them on signaling context confidence. The modal adaptive weights and the adaptive temperature coefficient are used to construct a signaling context adaptive contrastive loss function; or

[0044] S3012, Obtain the base temperature coefficient Based on intent anchor vector The multimodal features and the modality adaptive weights The basic temperature coefficient and signaling context confidence Construct a contrastive loss function.

[0045] In S3012 above, the contrastive loss function can be calculated using the following formula:

[0046] ;

[0047] Replace InfoNCE with Dynamic Marginal Triple Ranking Loss, where positive sample distance is the Euclidean distance from the modal feature to the signaling anchor point, and negative sample distance is the distance to the interference pool; change the temperature coefficient τ from a linear mapping to an exponential smoothing mapping. The exponentially smoothed dynamic margin (or adaptive temperature / boundary coefficient) for signaling confidence modulation has the following specific meaning and technical function:

[0048] It is the signaling context confidence score, a normalized index (within the range [0,1]), which quantifies the reliability of the current IP Multimedia Subsystem (IMS) signaling in real time.

[0049] It is an exponential smoothing mapping function that transforms the linear decay of confidence into a nonlinear accelerated expansion of the boundary.

[0050] The base temperature coefficient is ρ, the fault tolerance adjustment factor is ρ, and the consistency regularization term is changed from L2 distance to cosine similarity. , for one of the, These are negative sample values.

[0051] In an exemplary embodiment, in step S3011 above, confidence is based on signaling context. The modal adaptive weights and adaptive temperature coefficients used to construct the comparative loss function can include:

[0052] S111, Confidence based on signaling context The multimodal features and the intent anchor vector are used to calculate the single-modal NCE loss term. ;

[0053] S112, Determine meaning Figure 1 Consistency regularization term ;

[0054] S113, based on the single-mode NCE loss term Modal adaptive weights and Figure 1 Consistency regularization term Construct a contrastive loss function.

[0055] The contrastive loss function can be calculated using the following formula:

[0056] .

[0057] In one exemplary embodiment, step S111 described above may include one of the following:

[0058] Method 1: Confidence based on signaling context Constructing an adaptive temperature coefficient Based on the cosine similarity between the multimodal features and the intention anchor vector, and the adaptive temperature coefficient Calculate the single-mode NCE loss term ;

[0059] For example, it can be calculated using the following formula:

[0060] ;

[0061] in, It is the j-th negative sample vector, derived from a preset set of negative sample values. This is used as a deflection mechanism in contrastive learning. In InfoNCE loss calculation, the model calculates the current modality features. The similarity with the negative sample vector is calculated, and its similarity to the positive sample anchor point is forced to increase through gradient backpropagation. distance, Let be the j-th negative sample vector.

[0062] Method 2: Confidence based on signaling context Determine the dynamic margins of signaling confidence modulation, for example, , The dynamic margin of signaling confidence modulation. It is the initial interval constant of the basic margin, and the dynamic margin of signaling confidence modulation. The baseline value. When the signaling quality is optimal ( When =1), the dynamic marginal degenerates into Its function is to set the minimum fundamental Euclidean distance that positive and negative samples must maintain in the feature space, ensuring that the model can sharpen the convergence boundary when the signaling is stable, and avoiding premature convergence or gradient vanishing. The single-modal NCE loss term is calculated based on the dynamic margin of the signaling confidence modulation, multimodal features, and the Euclidean distance of the intention anchor vector. For example, it can be calculated using the following formula: .

[0063] In one exemplary embodiment, step S112 may include one of the following:

[0064] Calculate the sum of squared Euclidean distances between each modal feature in the multimodal features, as the intention. Figure 1 Consistency regularization term;

[0065] For example, ;or

[0066] , The Kullback-Leibler divergence is calculated to measure the degree of difference between two probability distributions.

[0067] Based on the intent anchor vector Calculate the cosine similarity or KL divergence between modal features in a multimodal feature set, as the meaning... Figure 1 Consistency regularization term.

[0068] For example, .

[0069] In an exemplary embodiment, step S11 above, calculating the modality adaptive weights based on the multimodal features and the intent anchor vector, includes:

[0070] The cosine similarity between the multimodal features and the intention anchor vector is determined respectively, and the modality adaptive weights are calculated based on the cosine similarity.

[0071] For example, .

[0072] The element-wise product of the multimodal features and the intention anchor vector is determined based on element-wise multiplication, and the modal adaptive weights are calculated based on the element-wise product.

[0073] For example, , , These are trainable parameters for lightweight gated networks used to replace cosine similarity in calculating modality weights. Their specific meanings are as follows:

[0074] This is the gate weight matrix / vector, which represents the linear projection parameters that the model can learn. It is used to compare the signaling intent anchor vector Vsig with the current modal features. The element-wise interaction results (⊙) are weighted and mapped to determine the contribution weight of each feature dimension to the "modal-signaling semantic matching degree", thereby capturing a more complex nonlinear association than simple cosine similarity.

[0075] This is the gated bias term, a learnable bias parameter of the model. It is used to perform baseline translation adjustment on the weighted projection results, so that the output range of the activation function (such as Sigmoid or Softmax) can be dynamically calibrated independently of the input features, avoiding gradient saturation caused by extreme feature values, and improving the stability and flexibility of weight allocation.

[0076] In one exemplary embodiment, confidence based on signaling context Constructing an adaptive temperature coefficient includes one of the following:

[0077] Obtain the basic temperature coefficient Based on the confidence level of the signaling context With basic temperature coefficient Construct an adaptive temperature coefficient;

[0078] For example, .

[0079] Adaptive temperature modulation based on feature variance and information entropy of multimodal features, and confidence based on signaling context. Determine the adaptive temperature coefficient.

[0080] For example, or , It is information entropy calculation. It is a variance calculation.

[0081] In an exemplary embodiment, step S104 above, parsing the SIP signaling to obtain a structured signaling context may include: parsing the standard header fields and 5G extended fields of the SIP signaling; by concatenating fields, verifying and filling missing values ​​according to a predefined schema, hashing sensitive fields according to a privacy policy, and outputting the structured signaling context.

[0082] In one exemplary embodiment, the standard header field includes P-Asserted-Identity (user identity), P-Access-Network-Info (access location), and Privacy (privacy policy), wherein the P-Asserted-Identity is used to construct the intent anchor vector. The static base dimension, the P-Access-Network-Info access location, is used to construct the intent anchor vector. Spatial context dimension; 5G extended fields include P-5GNC-Mem-Sync-ID memory synchronization identifier and P-Intent-Anchor-Profile intent anchor metadata.

[0083] In an exemplary embodiment, step S104 above, generating an intent anchor vector based on the structured signaling context, may include: performing discrete field encoding and continuous field normalization on the structured signaling context to obtain the processed user identity and the processed access location; and generating an intent anchor vector based on the processed user identity. The static base dimension, and the spatial context dimension that generates the intent anchor vector based on the processed access location.

[0084] In an exemplary embodiment, step S108 may include: determining a target business agent and its corresponding processing strategy based on a unified intent vector; processing the multimodal data using the target business agent and its corresponding processing strategy to obtain a processing result. The target agent and processing strategy are determined based on the unified intent vector, and the target agent executes specific business logic. This achieves intelligent routing based on unified semantic understanding, avoiding erroneous routing based on a single modality or rule.

[0085] In an exemplary embodiment, the method further includes: intercepting an update request for the processing result and parsing metadata from the update request, wherein the metadata includes at least a target business agent identifier and a knowledge generation timestamp; obtaining the current call session state, network slice Quality of Service (QoS) metrics, user registration status priority, and the historical credibility of the target business agent; determining the arbitration comprehensive score corresponding to the update request submitted by the target business agent at the current moment based on the target business agent identifier, knowledge generation timestamp, current call session state, network slice QoS metrics, user registration status priority, and the historical credibility of the target business agent; and processing the update request based on the arbitration comprehensive score. Incorporating network state (QoS, session stage) and agent historical performance into the arbitration makes memory updates more consistent with the current network environment and agent reliability, improving service continuity and credibility.

[0086] In an exemplary embodiment, processing an update request based on an arbitration comprehensive score may include: updating the processing result to a shared space in response to the difference between the arbitration comprehensive score corresponding to the currently submitted update request and the arbitration comprehensive score corresponding to the previously submitted update request being less than a preset conflict threshold; sending a conflict summary request confirmation to the terminal in response to the difference between the arbitration comprehensive score corresponding to the currently submitted update request and the arbitration comprehensive score corresponding to the previously submitted update request being greater than or equal to the preset conflict threshold and less than or equal to a soft arbitration trigger threshold; and updating the processing result to a shared space based on the confirmation message if the terminal confirms the conflict summary.

[0087] No significant conflict: If the score difference is less than the preset conflict threshold, update directly.

[0088] Soft arbitration: When the score difference is between [preset conflict threshold, soft arbitration trigger threshold], a conflict summary is sent to the user, and the system updates after confirmation.

[0089] Hard arbitration (implicit): If the score difference exceeds the soft arbitration trigger threshold, the high-scoring version will be locked directly (although not explicitly stated, logically it corresponds to direct overwriting or archiving).

[0090] Introducing a soft arbitration mechanism allows users to participate in decision-making when minor conflicts exist, improving user satisfaction and transparency. Automated handling ensures low latency in cases of no conflict or high-confidence conflicts; manual intervention is introduced in ambiguous areas to ensure accuracy.

[0091] This application addresses the shortcomings of 5G new calls, such as "signaling being used only for session control and multimodal independent parsing leading to intent fragmentation." The core improvement lies in proposing a signaling-driven adaptive intent alignment mechanism. This involves mapping SIP standard and extended signaling fields to a 256-dimensional signaling intent anchor vector via an MLP projection network. The signaling is transformed into a stable benchmark for multimodal semantic understanding; then, the signaling context adaptive contrastive loss function LSIP-CACL is designed, using the intent anchor vector. As positive sample anchor points, modal weights are dynamically allocated by calculating the cosine similarity between each modal feature and the anchor point, and intention is introduced. Figure 1 Consistency regularization ensures that heterogeneous features such as voice, video, and text converge accurately in a unified semantic space, outputting a unified intent vector. This anchoring alignment result, along with the 1+N hierarchical memory topology and the dynamic arbitration engine based on SIP state modulation, achieves closed-loop collaboration between intent, memory, and routing. Applied to the 5G New Calling personal AI Agent accompanying service scenario, it is deployed on the operator's network-side application server (AS) or multi-access edge computing (MEC) nodes, deeply coupled with the IMS core network.

[0092] Figure 4 This is a system framework diagram of 5G new call multi-agent hierarchical memory collaboration and intent alignment according to an embodiment of this application, such as... Figure 4 As shown, it includes a service execution and routing distribution module, a signaling and intent processing module, and a hierarchical memory arbitration module.

[0093] The service execution and routing distribution module includes a dynamic routing distribution module and a service agent cluster. The dynamic routing distribution module is the central hub for service scheduling in the entire system, responsible for receiving unified intent vectors from the signaling and intent processing module. The system uses the arbitration results from the hierarchical memory arbitration module to execute intelligent task distribution. This dynamic routing and distribution module comprises three core components: a state-aware engine, a capability matching evaluator, and a confidence arbitrator. The business agent cluster consists of multiple containerized microservices, including specialized business units such as call agents, scheduling agents, meeting agents, and translation agents. Each agent is an independent, lightweight container instance, achieving hard isolation between CPU (Central Processing Unit) and memory resources through Cgroups (Control Groups), NUMA (Non-Uniform Memory Access), or container technology, ensuring that business policies do not interfere with each other.

[0094] The signaling and intent processing module includes a SIP signaling parsing module, a multimodal feature extraction module, a contrastive learning intent alignment module, and a signaling intent anchor generation module. The SIP signaling parsing module primarily parses the SIP signaling content and extended field content. The multimodal feature extraction module is responsible for extracting features from multimodal data such as text, audio, and video. This module includes three parallel channels: a speech feature extractor, a video feature extractor, and a text / gesture feature extractor. The signaling intent anchor generation module is the component that transforms context into a vectorizable semantic benchmark. This module includes a field concatenation unit, an MLP projection network, and a vector output unit. The contrastive learning intent alignment module is the core engine for eliminating multimodal semantic fragmentation, responsible for converging heterogeneous features into a unified semantic space. This module includes a projection matrix optimizer, a loss calculator, and a vector aggregator.

[0095] Figure 5 This is a flowchart of 5G new call multi-agent hierarchical memory collaboration and intent alignment according to an embodiment of this application, such as... Figure 5 As shown, it includes:

[0096] S501, Signaling Parsing and Anchor Generation, parses the identifier header field and 5G extended fields of SIP signaling initiated by the terminal to generate an intent anchor vector. Specifically, after the user initiates a call, the original SIP message stream is captured as input. The standard header fields (P-Asserted-Identity, P-Access-Network-Info, Privacy policy) and 5G extended fields (P-5GNC-Mem-Sync-ID, P-Intent-Anchor-Profile metadata) are parsed. A field splicer validates the data according to a predefined schema and fills in missing values. Sensitive fields are hashed according to the privacy policy. The resulting structured signaling context is then saved in JSON format, enabling structured extraction of signaling semantics. The signaling intent anchor generation module receives the structured signaling context, performs One-Hot encoding of discrete fields and normalizes continuous fields using a field encoder, and maps it to a 256-dimensional signaling intent anchor vector through an MLP projection network. The telecom-grade context is transformed into a semantic benchmark that can be vectorized and participated in. At the same time, the multimodal feature extraction module processes the input speech, video and text media streams in parallel, and outputs the original multimodal tensor {Fa, Fv, Ft} ​​through the speech feature extractor, video feature extractor and text / gesture feature extractor respectively.

[0097] S502, multimodal feature extraction, injects multimodal data into media streams and data channels, and extracts speech, video and text features in parallel.

[0098] S503, Contrastive learning intent alignment, calculates the modal features and intent anchor vectors. The cosine similarity is used to dynamically calculate modality adaptive weights and adaptive temperature coefficients; an adaptive contrastive loss function for signaling context is constructed, the projection matrix is ​​optimized through backpropagation, and a unified intent vector is output.

[0099] Comparison of learning intent alignment modules For positive sample anchor points, the features {Fa, Fv, Ft} ​​after being mapped by a learnable projection matrix { Using} as candidate samples, construct a signaling context adaptive contrastive loss function. First, calculate the cosine similarity between each mode and the signaling anchor point. Dynamic weights are generated using Softmax. To achieve adaptive modality allocation, and then combine it with the signaling context confidence normalization factor. Calculate the single-modal NCE loss term, and then calculate the meaning. Figure 1 Consistency regularization term By iteratively optimizing the projection matrix through backpropagation, a unified intent vector with fixed dimensions is finally aggregated and output, achieving accurate convergence of heterogeneous multimodal features in a unified semantic space, and providing standardized input for downstream memory arbitration and routing distribution.

[0100] The loss function embeds the registration status confidence score parsed from SIP signaling and the dynamic arbitration weights into a comparative learning space to achieve adaptive computation. The following formula is used:

[0101] ;

[0102] in, The modal adaptive weights are calculated using the following formula:

[0103] ;

[0104] Where sim is the cosine similarity calculation function, and η is the weight adjustment coefficient. The formula for calculating the single-mode NCE loss term is as follows:

[0105] ;

[0106] in For the purpose Figure 1 The regularization term is calculated using the following formula:

[0107] ;

[0108] in It is the signaling intent anchor vector. These are the speech, video, and text feature vectors mapped by a learnable projection matrix. To unify the intent vector.

[0109] For negative sample pool, This is the signaling context confidence score, calculated using the normalization method shown below:

[0110] ;

[0111] `sim` is the cosine similarity calculation function, and `K` is the number of negative samples. For the j-th negative sample vector, ρ is the base temperature coefficient, ρ is the fault tolerance adjustment factor, and λ is the regularization equilibrium coefficient.

[0112] The signaling context adaptive contrastive loss function takes as input a 256-dimensional signaling intent anchor vector output by the signaling intent anchor generation module. The multimodal feature extraction module outputs in parallel the original speech / video / text feature tensor {Fa, Fv, Ft} ​​and its feature vector {Fa, Fv, Ft} ​​after being mapped by a learnable projection matrix. }, a pre-defined negative sample pool (Dneg), and real-time calculated signaling context confidence. With network preset hyperparameters (basic temperature coefficient) The system uses the following parameters during optimization: fault tolerance adjustment factor ρ, regularization balance coefficient λ, and weight adjustment coefficient η. As a positive sample benchmark, combined with dynamic modal weights With adaptive temperature coefficient Calculate the single-sample NCE contrastive loss term for each modality and then add the cross-modal L2 significance. Figure 1 A consistency regularization term is used to construct the total loss LSIP-CACL. The cross-modal projection matrix parameters are iteratively updated through backpropagation to minimize the loss value. Finally, after the model converges, the optimized multimodal features are aggregated, and a fixed-dimensional unified intent vector carrying the precise semantic alignment results of speech, video, text, and signaling is output. It serves directly as a standardized input for downstream hierarchical memory arbitration and dynamic routing distribution.

[0113] Signaling Context Adaptive Contrastive Loss Function LSIP-CACL with Telecom-Grade Signaling Intent Anchor Vector To stabilize the positive sample benchmark, projection features of each modality are calculated. and The cosine similarity is calculated and normalized using Softmax to dynamically generate modal weights. This enables adaptive focusing on modalities highly correlated with the signaling context; simultaneously, it introduces signaling context confidence. Constructing an adaptive temperature coefficient The system automatically smooths the loss optimization surface during roaming transitions or instantaneous signaling fluctuations, effectively suppressing model erroneous convergence. Building upon this, the InfoNCE comparison mechanism is used to narrow the distance between each modal feature and the signaling anchor point and push them away from the negative sample pool. A cross-modal L2 consistency regularization term is then applied to force heterogeneous features to cluster compactly within the semantic space. Finally, weighted summation and backpropagation are jointly used to optimize the cross-modal projection matrix, ensuring accurate convergence of heterogeneous features such as speech, video, and text within a unified semantic space, outputting a robust and highly consistent unified intent vector. .

[0114] The signaling intent processing module converges carrier-grade signaling context and heterogeneous multimodal media streams into a unified intent representation: This module first parses the SIP standard and 5G extended fields to generate a 256-dimensional signaling intent anchor vector. As a semantic benchmark, original features of speech / video / text are extracted synchronously and in parallel, and then... For positive samples, a Signaling Context Adaptive Contrastive Loss Function (SIP-CACL) is constructed. Through joint optimization of dynamic modality weight allocation, confidence adaptive temperature adjustment, and cross-modal consistency regularization, heterogeneous features are driven to converge accurately in a unified semantic space, ultimately outputting a unified intent vector of fixed dimensions. It provides a robust, interpretable, and carrier-grade reliable standardized intent input for downstream memory arbitration and dynamic routing.

[0115] S504, a layered memory and arbitration mechanism, intercepts write requests from each Agent and parses metadata to obtain SIP status, network QoS, and registration status. It calculates the comprehensive adjudication score of existing knowledge and newly submitted knowledge, and determines whether the difference in the comprehensive adjudication score is less than a preset conflict threshold. If the result is yes, the winning version is written to the shared space. If the result is no, the highest-scoring version is locked directly, and then the winning version is written to the shared space. The losing version is archived, and the memory synchronization time is broadcast to refresh the Agent cache.

[0116] The hierarchical memory and arbitration module comprises a dynamic arbitration engine and a hierarchical memory framework module. The dynamic arbitration engine is the core adjudication component for resolving cross-agent knowledge conflicts, responsible for intercepting write requests from the private library to the shared space and performing deterministic arbitration. The hierarchical memory framework module includes a read-only shared space and N private agent data spaces, ensuring data separation and privacy.

[0117] The dynamic arbitration algorithm uses signaling state modulation to jointly model the SIP session state machine, network slice QoS, alignment confidence and agent historical credibility, and construct a dynamic scoring function that adaptively evolves with the signaling stage.

[0118] Suppose that the i-th business agent (i∈{1,2,...,N}) submits a memory update request at time ti. The arbitration engine calculates its overall award score. as follows:

[0119] ;

[0120] in, ∈{0,1} is the user explicit declaration identifier, where 0 represents no declaration and 1 represents explicit declaration. Knowledge is the time difference between the knowledge generation and the present conversation (knowledge generation timestamp).

[0121] ∈[0,1] is the quality index of the current network slice QoS. It is the Sigmoid mapping function. It is the cosine similarity between the unified intent vector and the intent anchor vector of the contrastive learning alignment module. It is the historical credibility score of the intelligent agent. It is the priority of user registration status. The current SIP session state machine phase (current call session state) ∈{Early, Confirmed, Terminated} represents the period before the call, during the call, and after the call ends. Weighting coefficients. - It is modulated in real time by the SIP state machine, reflecting business context awareness. The dynamic weight update formula is as follows: ,in, These are the preset state modulation matrix parameters.

[0122] This arbitration algorithm differs fundamentally from existing technologies that rely on black-box scheduling logic that prioritizes last writes or hard-coded static priorities. Its core innovation lies in proposing a four-dimensional weighted joint scoring mechanism based on SIP session state machine modulation. This mechanism deeply couples user explicit declarations, session temporal decay, multimodal alignment confidence, and Agent historical credibility. Furthermore, the weight coefficients of each dimension are not fixed values ​​but evolve in real time through a pre-set modulation matrix as the SIP session lifecycle progresses. This ensures that the arbitration logic strictly follows the deterministic rule of user intent > session timeliness > semantic confidence > network state.

[0123] This algorithm is the first to directly map carrier-grade control plane parameters, including IMS registration priority and slice QoS, into arbitration weight factors. It quantifies network quality and timing differences through Sigmoid smoothing and exponential decay functions. Unlike the static rule paradigm in traditional architectures where signaling status and service routing are disconnected, this method gives conflict resolution strong interpretability. It combines dual-threshold hierarchical judgment and hash metadata to achieve complete archiving, enabling precise backtracking of decision paths based on single calls and agent credibility sliding self-evolution. It solves the systemic defects of existing technologies, such as random, unauditable, and unoptimizable resolution, and meets the carrier-grade high reliability and compliance audit requirements of communication services.

[0124] S505 features dynamic routing and service execution, combining network reachability, terminal capabilities, and intent confidence into a three-element framework. Routes are distributed to the corresponding Agents to execute tasks; execution success rates and user interaction feedback are collected, and the Agent's historical credibility score is updated via sliding.

[0125] S506 determines whether to capture SIP termination signaling. If the result is yes, it triggers full memory snapshot solidification and binds the call identifier and user identifier to complete the closed loop. If the result is no, it continues to listen and waits for the next signaling or write request.

[0126] Figure 6 This is a flowchart of dynamic arbitration according to an embodiment of this application, such as... Figure 6 As shown, the specific steps include:

[0127] Step S601: Intercept the write request and parse the metadata. The parsed content includes: source agent identifier, knowledge generation timestamp, intent alignment confidence, and data type label.

[0128] Step S602: Obtain signaling status and network context, including: current call session status, network slice quality indicators, and user registration status priority.

[0129] Step S603: Dynamic weight calculation. Obtain the weight modulation coefficients from the session query table and calculate the user claim weight, time-series decay weight, confidence weight, and registration status weight.

[0130] Step S604, Comprehensive scoring and conflict detection: Calculate the comprehensive adjudication score for both existing knowledge and newly submitted knowledge, and calculate the score difference between existing knowledge and newly submitted knowledge.

[0131] Step S605: Determine if the score difference is greater than a preset conflict threshold. If the result is no, directly lock the highest-scoring version and proceed to S606. If the result is yes, determine if the score difference is less than or equal to a soft arbitration trigger threshold. If the score difference is greater than the soft arbitration trigger threshold, directly lock the highest-scoring version. If the score difference is less than or equal to the soft arbitration trigger threshold, generate a conflict summary and send it to the terminal interface through the data channel, waiting for the user to manually select or confirm via voice. Determine if the user has completed confirmation. If the user has completed confirmation, set the user declaration identifier of the corresponding knowledge to the highest priority and return to step S604 to recalculate the comprehensive adjudication score. If the user has not completed confirmation, adjudicate directly according to the current score and proceed to S606.

[0132] Step S606: The arbitration is executed and the memory topology is synchronized. The winning version is written to the read-only shared space index, and the losing version is archived to the history chain along with the comprehensive arbitration score, metadata, and integrity hash. The memory synchronization event is broadcast to refresh the local cache of the relevant agents.

[0133] Step S607: Credibility feedback loop: Collect the agent's execution success rate and user interaction satisfaction, and update the agent's historical credibility score by sliding.

[0134] The arbitration engine intercepts requests from each agent to write session state to the private database, parses the source agent identifier, UTC timestamp, alignment confidence, and data type tag, and then queries the control plane for the current session's SIP state, network slice QoS metrics, and user IMS registration status to construct the real-time arbitration context. Next, it looks up the modulation coefficients based on the SIP session state and calculates dynamic weight parameters such as user claim weight, timing attenuation weight, confidence weight, and registration status weight.

[0135] Subsequently, the comprehensive adjudication score is calculated for the existing knowledge in the shared space and the newly submitted knowledge, and it is determined whether the difference between the two scores is less than the conflict threshold. If the score difference is greater than or equal to the conflict threshold, it is determined that there is no significant conflict and the highest-scoring version is directly locked to enter the adjudication execution stage. If the score difference is less than the conflict threshold, it is determined to be a knowledge conflict and it is further determined whether the difference is lower than the soft arbitration trigger threshold. If the difference is lower than the soft arbitration trigger threshold, a conflict summary is generated and sent to the terminal interface through the IMSData Channel for the user to manually select or confirm by voice.

[0136] If the user confirms, the user's claim identifier for the corresponding knowledge is set to the highest priority and the score is recalculated. If the user times out or does not confirm, the current score is used to make the decision directly. If the difference is higher than the soft arbitration trigger threshold, the highest-scoring version is locked directly. The winning version is then written to the read-only shared space index, and the losing version, along with its arbitration score, quadruple metadata, and hash root, is archived to the historical chain. At the same time, a memory synchronization event is broadcast to refresh the local cache of the relevant Agent. Then, the Agent's routing success rate and user interaction satisfaction are collected to slide and update the Agent's historical trustworthiness score to complete the trustworthiness self-evolution. Finally, if a SIP BYE signaling is captured, a full memory snapshot is triggered to solidify and bind the SIP Call-ID to complete the lifecycle loop. If no termination signaling is captured, the Agent remains in a listening state, waiting for the next write request.

[0137] This embodiment uses cross-domain preference conflict and millisecond-level synchronization of multiple service agents as an example to illustrate the implementation method of the hierarchical memory collaboration of the present invention. Figure 7 This is a flowchart of hierarchical memory synchronization and arbitration based on IMS signaling event-driven implementation according to an embodiment of this application, as follows: Figure 7 As shown, it includes:

[0138] S701, Establish Call. The user terminal initiates an INVITE signaling to the IMS core network to establish the call service;

[0139] S702, binding to the Confirmed state triggers a synchronization window. The SIP signaling parsing module captures the signaling and parses the standard header fields and memory synchronization extension fields, binds the session to the Confirmed state, and triggers the memory synchronization time window;

[0140] Step S703: Submit the "evening meeting" preference (including four-tuple metadata). The call agent identifies the user's "evening meeting" preference in a session isolation environment and generates a four-tuple write request carrying the source identifier, timestamp, alignment confidence, and data type label;

[0141] Step S704: A conflict is detected triggered by "Friday afternoon free". The dynamic arbitration engine intercepts the write request, detects that the shared space already has a "Friday afternoon free" preference recorded by the schedule agent, and determines that a cross-agent knowledge conflict has been triggered.

[0142] Step S705, Four-dimensional weighted scoring and adjudication. The arbitration engine retrieves the dynamic weights based on the current SIP status, performs a four-dimensional weighted scoring and adjudication (user explicit declaration > session sequence > alignment confidence > registration priority), and determines the winner based on call preference;

[0143] Step S706: The winning version is written to the index, and the losing version is archived. The winning version is written to the read-only shared space index, and the losing version, along with its arbitration score, quadruple metadata, and hash root, is asynchronously archived to the history chain.

[0144] Step S707: The arbitration engine generates a memory synchronization event and broadcasts it.

[0145] Step S708: After receiving the notification, the scheduling agent refreshes its local cache and recalculates the scheduling strategy.

[0146] In step S709, the terminal sends a BYE signaling to release the session, the state machine captures the Terminated state, triggers a full memory snapshot to solidify and bind the SIP Call-ID, and completes the lifecycle loop.

[0147] This embodiment uses cross-modal semantic convergence and unified intent output of compound instructions as an example to illustrate the intent alignment implementation method of the present invention. Figure 8 This is a flowchart of multimodal intent alignment based on a UMF media pipeline according to an embodiment of this application, as shown below. Figure 8 As shown, it includes:

[0148] Step S801: Terminal concurrently sends commands and media streams. The user terminal sends voice commands, touch gesture sequences, Data Channel text messages, and SIP signaling carrying extended fields in parallel.

[0149] Step S802: Generate intent anchor vector. The SIP signaling parsing module extracts the standard and extended header fields and generates a 256-dimensional signaling intent anchor vector through an MLP projection network. ;

[0150] Step S803: Inject RTP stream + Data channel, and perform zero-copy feature extraction at the transcoding stage. Audio, video, and text media streams are synchronously injected into the Unified Media Framework (UMF) pipeline along with the data channel, triggering zero-copy feature extraction at the transcoding / mixing stage.

[0151] Step S804: Inject the multimodal feature tensor into the intent anchor vector. The multimodal feature extraction module outputs acoustic sequence features, keyframe visual features, and explicit instruction semantic tensors in parallel, which are then concatenated and injected into the contrastive learning intent alignment module.

[0152] Step S805, SIP-CACL comparative learning, yields a unified intent vector. The alignment engine then... Using positive sample anchors, a signaling context adaptive contrastive loss function (SIP-CACL) is constructed, and the cross-modal projection matrix is ​​iteratively optimized to converge and generate a unified intent vector. ;

[0153] Step S806: The vector routing module outputs a unified intent (i.e., the dynamic route distribution module).

[0154] Step S807, Route Distribution. The dynamic route distribution module receives the unified intent vector. By combining network reachability and terminal capability matching, the task is accurately routed to the translation agent;

[0155] Step S808, Simultaneous Interpretation / Subtitle Response. The translation agent loads the intent context, performs real-time simultaneous interpretation / subtitle generation, and the service response stream is returned to the user terminal in real time.

[0156] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0157] This embodiment also provides a data processing apparatus for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated. The apparatus includes:

[0158] The extraction module is used to acquire multimodal data based on the session initiated by the terminal, and extract multimodal features based on the multimodal data;

[0159] The parsing and generation module is used to parse the Session Initiation Protocol (SIP) signaling to obtain the structured signaling context, and generate an intent anchor vector based on the structured signaling context.

[0160] The determination module is used to determine a unified intent vector based on multimodal features and intent anchor vectors;

[0161] The processing module is used to process multimodal data based on a unified intent vector.

[0162] In the embodiments of this application, the above-mentioned device may also include different modules, and the naming and functional division of the modules may be selected in different ways according to the actual situation, without specific restrictions.

[0163] It should be noted that the above modules can be implemented by software or hardware. For the latter, they can be implemented in the following ways, but are not limited to: all the above modules are located in the same processor; or, the above modules are located in different processors in any combination.

[0164] This application also provides a network device. Figure 9 This is a structural block diagram of a network device according to an embodiment of this application, such as... Figure 9 As shown, the network device 900 includes a receiver 901, a transmitter 902, and a processor 903. The network device 900 is used to perform the steps of the above method embodiments through at least one of the receiver 901, transmitter 902, and processor 903.

[0165] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when run.

[0166] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0167] Embodiments of this application also provide an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0168] In one exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0169] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0170] In one exemplary embodiment, the computer program product described above includes a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the methods described in various embodiments of this application.

[0171] Specific examples in this embodiment can be found in the examples described in the above embodiments and exemplary implementations, and will not be repeated here.

[0172] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0173] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.

Claims

1. A data processing method, executed by a network-side device, characterized in that, include: Multimodal data is acquired based on the session initiated by the terminal, and multimodal features are extracted based on the multimodal data; The SIP signaling of the session initiation protocol is parsed to obtain a structured signaling context, and an intent anchor vector is generated based on the structured signaling context. A unified intent vector is determined based on the multimodal features and the intent anchor vector; The multimodal data is processed based on the unified intent vector.

2. The method according to claim 1, characterized in that, Determining a unified intent vector based on the multimodal features and the intent anchor vector includes: Using the intent anchor vector as a positive sample benchmark, a structured signaling context adaptive contrastive loss function is constructed based on the intent anchor vector, the multimodal features, and the signaling context confidence. The total loss value of the signaling context adaptive contrastive loss function is minimized through backpropagation. After convergence, the unified intent vector is obtained by aggregation.

3. The method according to claim 2, characterized in that, The signaling context adaptive contrastive loss function constructed based on the intent anchor vector, the multimodal features, and the signaling context confidence includes one of the following: The signaling context adaptive contrastive loss function is constructed based on the intent anchor vector, the multimodal features, the modality adaptive weights, and the signaling context confidence; or An adaptive temperature coefficient is determined based on the signaling context confidence, and an adaptive contrastive loss function for the signaling context is constructed based on the intent anchor vector, the multimodal features, and the adaptive temperature coefficient.

4. The method according to claim 3, characterized in that, The signaling context adaptive contrastive loss function, constructed based on the intent anchor vector, the multimodal features, the modality adaptive weights, and the signaling context confidence, includes one of the following: Modality adaptive weights are calculated based on the multimodal features and the intent anchor vector, and the signaling context adaptive contrastive loss function is constructed based on the signaling context confidence, the modality adaptive weights, and the adaptive temperature coefficient. or Obtain the base temperature coefficient, and construct the signaling context adaptive contrastive loss function based on the intent anchor vector, the multimodal features, the modality adaptive weights, the base temperature coefficient, and the signaling context confidence.

5. The method according to claim 4, characterized in that, The signaling context adaptive contrastive loss function, constructed based on signaling context confidence, modality adaptive weights, and adaptive temperature coefficient, includes: The single-modal negative sample contrast error (NCE) loss term is calculated based on the signaling context confidence, the multimodal features, and the intent anchor vector. Determine the intent consistency regularization term; The signaling context adaptive contrastive loss function is constructed based on the single-modal NCE loss term, the modality adaptive weights, and the intent consistency regularization term.

6. The method according to claim 5, characterized in that, The single-modal negative sample contrast error (NCE) loss term, calculated based on signaling context confidence, the multimodal features, and the intent anchor vector, includes one of the following: An adaptive temperature coefficient is constructed based on the signaling context confidence. The single-modal NCE loss term is calculated based on the cosine similarity between the multimodal features and the intent anchor vector, and the adaptive temperature coefficient; or The dynamic margin of signaling confidence modulation is determined based on the signaling context confidence, and the single-mode NCE loss term is calculated based on the dynamic margin of signaling confidence modulation, the multimodal features, and the Euclidean distance of the intent anchor vector.

7. The method according to claim 5, characterized in that, The intent consistency regular expression includes one of the following: Calculate the sum of squared Euclidean distances between each modality feature in the multimodal features, and use it as the intent consistency regularization term; or Based on the intent anchor vector, the cosine similarity or KL divergence between each modality feature in the multimodal features is calculated and used as the intent consistency regularization term.

8. The method according to claim 4, characterized in that, The modality adaptive weights are calculated based on the multimodal features and the intent anchor vector, including one of the following: The cosine similarity between the multimodal features and the intention anchor vector is determined respectively, and the modality adaptive weights are calculated based on the cosine similarity. or The element-wise product of the multimodal features and the intention anchor vector is determined based on element-wise multiplication, and the modal adaptive weights are calculated based on the element-wise product.

9. The method according to claim 6, characterized in that, Constructing an adaptive temperature coefficient based on signaling context confidence includes one of the following: Obtain the base temperature coefficient, and construct the adaptive temperature coefficient based on the signaling context confidence and the base temperature coefficient; or Adaptive temperature modulation based on the feature variance and information entropy of the multimodal features, and the adaptive temperature coefficient determined according to the signaling context confidence.

10. The method according to claim 1, characterized in that, Parsing the SIP signaling yields the structured signaling context, which includes: Parse the standard header fields and 5G extension fields of the SIP signaling; By concatenating fields, validating and filling in missing values ​​according to a predefined schema, and hashing sensitive fields according to a privacy policy, the structured signaling context is output.

11. The method according to claim 10, characterized in that, The standard header fields include P-Asserted-Identity (user identity), P-Access-Network-Info (access location), and Privacy (privacy policy). The P-Asserted-Identity is used to form the static base dimension of the intent anchor vector, and the P-Access-Network-Info (access location) is used to form the spatial context dimension of the intent anchor vector. The 5G extended fields include P-5GNC-Mem-Sync-ID (memory synchronization identifier) ​​and P-Intent-Anchor-Profile (intent anchor metadata).

12. The method according to claim 11, characterized in that, Generating intent anchor vectors based on the structured signaling context includes: The structured signaling context is subjected to discrete field encoding and continuous field normalization to obtain the processed user identity and processed access location; The static base dimension of the intent anchor vector is generated based on the processed user identity, and the spatial context dimension of the intent anchor vector is generated based on the processed access location.

13. The method according to claim 1, characterized in that, Extracting multimodal features based on the multimodal data includes: The multimodal data is subjected to speech feature extraction, video feature extraction, and text feature extraction to obtain the multimodal features.

14. The method according to claim 1, characterized in that, Processing the multimodal data based on the unified intent vector includes: The target business agent and corresponding processing strategy are determined based on the unified intent vector. The multimodal data is processed using the target business intelligence agent and the corresponding processing strategy to obtain the processing result.

15. The method according to claim 14, characterized in that, The method further includes: Intercept the update request of the processing result and parse the metadata from the update request, wherein the metadata includes at least the target business intelligent agent identifier and the knowledge generation timestamp; Obtain the current call session status, network slice service quality (QoS) indicators, user registration status priority, and the historical credibility of the target business agent; Based on the target business agent identifier, the knowledge generation timestamp, the current call session status, the network slice QoS index, the user registration status priority, and the historical credibility of the target business agent, determine the arbitration comprehensive score corresponding to the update request submitted by the target business agent at the current moment; The update request is processed based on the comprehensive arbitration score.

16. The method according to claim 15, characterized in that, Processing the update request based on the arbitration comprehensive score includes: If the difference between the arbitration comprehensive score corresponding to the currently submitted update request and the arbitration comprehensive score corresponding to the previously submitted update request is less than a preset conflict threshold, the processing result is updated to the shared space. If the difference between the arbitration comprehensive score corresponding to the currently submitted update request and the arbitration comprehensive score corresponding to the previously submitted update request is greater than or equal to a preset conflict threshold, and less than or equal to a soft arbitration trigger threshold, a conflict summary request confirmation is sent to the terminal. If the terminal's confirmation information for the conflict summary is received, the processing result is updated based on the confirmation message, and the updated processing result is updated to the shared space.

17. A communication device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1 to 16.

18. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1 to 16.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program is configured to perform the steps of the method described in any one of claims 1 to 16 when executed.

Images