A software copyright protection detection method and system based on semantic analysis

Abstract

The present application provides a software copyright protection detection method and system based on semantic analysis. The method includes: in the interaction scenario between the user terminal and the software copyright protection detection system, collecting physical behavior signals and semantic feedback behavior data generated by the user operation trajectory; combining with a pre-trained semantic analysis model, performing coupled analysis on the physical behavior signals and semantic feedback behavior data to generate a user intention representation sequence; through a pre-trained reinforcement learning model, mapping the semantic feedback behavior and physical behavior in the user intention representation sequence into an association strength parameter; according to the association strength parameter, combining with a preset dynamic risk determination rule, generating a software copyright protection detection result including a copyright risk index. The present application improves the real-time detection accuracy of software copyright infringement behavior.

Description

Technical Field

[0001] This application relates to the technical field of semantic analysis, and particularly to a software copyright protection detection method and system based on semantic analysis. Background Art

[0002] With the diversification of software interaction methods, the scenarios where users interact with the system through multi-modal operations such as voice and gestures are becoming increasingly popular. Traditional copyright protection technologies are difficult to effectively identify infringement acts implemented through complex interaction behaviors. There is an urgent need for a detection method that can analyze the correlation between the semantics of user operations and physical behaviors in real time, and accurately identify abnormal operation patterns with infringement intentions while ensuring a normal user experience.

[0003] The current mainstream solution adopts a static feature analysis method based on operation logs. By extracting statistical features such as frequency and time sequence in software operation records, and combining with a predefined rule library for infringement risk matching. This method establishes a Markov model of the operation sequence, calculates the probability deviation between the user behavior path and the typical infringement pattern, and triggers an alarm when the deviation exceeds the threshold.

[0004] This solution only relies on the explicit features of operation logs and cannot capture the implicit correlation between semantic instructions and actual operations, resulting in a high missed detection rate for infringement acts disguised as normal operations. At the same time, the static rule library is difficult to adapt to the behavioral pattern variations generated by new interaction methods, and the false alarm rate increases in multi-modal scenarios such as voice-touch. Summary of the Invention

[0005] This application provides a software copyright protection detection method and system based on semantic analysis to solve the problem of low real-time detection accuracy of software copyright infringement acts in the prior art.

[0006] In a first aspect, this application provides a software copyright protection detection method based on semantic analysis, including:

[0007] In the interaction scenario between the user terminal and the software copyright protection detection system, collect the physical behavior signals and semantic feedback behavior data generated by the user operation trajectory;

[0008] Combined with a pre-trained semantic analysis model, perform coupled analysis on the physical behavior signals and the semantic feedback behavior data to generate a user intention representation sequence;

[0009] Through a pre-trained reinforcement learning model, map the semantic feedback behavior and physical behavior in the user intention representation sequence into association strength parameters;

[0010] According to the association strength parameters, combined with a preset dynamic risk determination rule, generate a software copyright protection detection result including a copyright risk index.

[0011] Optionally, mapping the semantic feedback behavior and the physical behavior in the user intention representation sequence into an association strength parameter by the pre-trained reinforcement learning model includes:

[0012] Segmentally encoding the time distribution feature of the semantic feedback behavior and the spatial distribution feature of the physical behavior respectively for each conversation turn to generate a joint feature representation;

[0013] Calculating a cooperation degree parameter between the semantic feedback behavior and the physical behavior within each conversation turn according to the joint feature representation;

[0014] Based on the cooperation degree parameter, iteratively optimizing the association relationship between the semantic feedback behavior and the physical behavior through the reinforcement learning model, and the iterative optimization process is a process of adjusting the matching weight between the semantic feature and the physical feature;

[0015] Performing a cross-turn fusion operation on the optimized matching weight and the cooperation degree parameter of the current conversation turn to generate an association strength parameter.

[0016] Optionally, the iteratively optimizing the association relationship between the semantic feedback behavior and the physical behavior through the reinforcement learning model based on the cooperation degree parameter includes:

[0017] Separating the time distribution feature corresponding to the semantic feedback behavior and the spatial distribution feature corresponding to the physical behavior from the joint feature representation, taking the time distribution feature as the semantic feature, and taking the spatial distribution feature as the physical feature;

[0018] Constructing a state representation vector reflecting the fluctuation of the cooperation degree parameter in the state space of the reinforcement learning model;

[0019] Processing the state representation vector through the policy network of the reinforcement learning model to generate a weight adjustment strategy;

[0020] Performing an iterative update operation on the matching weight between the semantic feature and the physical feature according to the weight adjustment strategy.

[0021] Optionally, the processing the state representation vector to generate a weight adjustment strategy includes:

[0022] Decomposing the state representation vector into the immediate feature of the current conversation turn and the memory features of a preset number of historical conversation turns to form a multi-level feature combination;

[0023] Performing a cross-turn correlation analysis on the multi-level feature combination, and determining the dynamic association pattern between the semantic feature and the physical feature according to the cross-turn correlation analysis result;

[0024] Generate dynamic adjustment parameters according to the dynamic association pattern;

[0025] Fuse the dynamic adjustment parameters with the coordination parameters of the current conversation turn to generate a weight adjustment strategy.

[0026] Optionally, the calculating the coordination parameter between the semantic feedback behavior and the physical behavior within each conversation turn according to the joint feature representation includes:

[0027] Extract the time interval sequence of the semantic feedback behavior and the spatial displacement sequence of the physical behavior from the joint feature representation;

[0028] Compare the change trends of the time interval sequence and the spatial displacement sequence within the same conversation turn for synchronization;

[0029] Generate a behavior coupling index according to the synchronization comparison result, and the behavior coupling index is obtained by quantifying the matching degree between the time interval change and the spatial displacement change;

[0030] Based on the behavior coupling index, combine the triggering density of the semantic feedback behavior and the trajectory coverage area of the physical behavior in the current conversation turn to calculate the coordination parameter.

[0031] Optionally, the coupling analysis of the physical behavior signal and the semantic feedback behavior data by combining a pre-trained semantic analysis model to generate a user intention representation sequence includes:

[0032] Input the physical behavior signal into the physical feature encoding layer of the semantic analysis model to output a structured physical feature sequence;

[0033] Input the semantic feedback behavior data into the semantic parsing layer of the semantic analysis model to output a standardized semantic feature sequence;

[0034] Establish a dynamic association mapping between the structured physical feature sequence and the standardized semantic feature sequence through the cross-attention coupling module of the semantic analysis model to generate behavior association features;

[0035] Use the verification module of the semantic analysis model to perform consistency verification on the behavior association features;

[0036] Combine the behavior association features that pass the verification in the order of conversation turn time to generate a user intention representation sequence.

[0037] Optionally, the generating a software copyright protection detection result including a copyright risk index according to the association strength parameter and combining a preset dynamic risk determination rule includes:

[0038] Arrange the association strength parameters in chronological order of the dialogue turns to form a parameter change curve;

[0039] Extract the key turning points in the parameter change curve as risk analysis nodes;

[0040] Set multiple risk determination intervals in the preset dynamic risk determination rule;

[0041] According to the distribution position of the risk analysis nodes in the parameter change curve, match the corresponding risk determination interval to determine the risk level;

[0042] Dynamically correct the risk level by combining the complexity of the semantic feedback behavior and the operation frequency of the physical behavior in the current dialogue turn;

[0043] Combine the mapping relationship between the preset risk level and the copyright risk index, and quantify the corrected risk level into the corresponding copyright risk index;

[0044] Combine the copyright risk index, the corrected risk level, and the software copyright protection strategy corresponding to the copyright risk index to generate a software copyright protection detection result.

[0045] In a second aspect, the present application provides a software copyright protection detection system based on semantic analysis, including:

[0046] An acquisition module for acquiring physical behavior signals and semantic feedback behavior data generated by a user operation trajectory in an interaction scenario between a user terminal and the software copyright protection detection system;

[0047] A first generation module for coupling and analyzing the physical behavior signals and the semantic feedback behavior data in combination with a pre-trained semantic analysis model to generate a user intention representation sequence;

[0048] A mapping module for mapping the semantic feedback behavior and the physical behavior in the user intention representation sequence into association strength parameters through a pre-trained reinforcement learning model;

[0049] A second generation module for generating a software copyright protection detection result including a copyright risk index according to the association strength parameters and in combination with a preset dynamic risk determination rule.

[0050] In a third aspect, the present application provides a computing device including a processor and a memory, where a computer program is stored in the memory, and the processor is configured to run the computer program to execute any one of the methods for software copyright protection detection based on semantic analysis in the first aspect.

[0051] Fourthly, the present application provides a computer storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, a software copyright protection detection method according to any one of the first aspect is implemented.

[0052] In the present application, a software copyright protection detection method based on semantic analysis is provided. The method includes: in an interaction scenario between a user terminal and a software copyright protection detection system, collecting physical behavior signals and semantic feedback behavior data generated by a user operation track; combining a pre-trained semantic analysis model, performing coupled analysis on the physical behavior signals and the semantic feedback behavior data to generate a user intention representation sequence; through a pre-trained reinforcement learning model, mapping the semantic feedback behavior and the physical behavior in the user intention representation sequence into an association strength parameter; according to the association strength parameter, combining a preset dynamic risk determination rule to generate a software copyright protection detection result including a copyright risk index.

[0053] The technical solution provided by the present application has the following beneficial effects:

[0054] The present application realizes synchronous capture of multi-modal user interaction data, providing a complete behavior observation basis for subsequent analysis. Establish a mapping relationship between semantic instructions and physical operations, eliminate the ambiguity of single-modal behavior analysis, and accurately restore the true intention of the user. Dynamically quantify the coordination degree between semantics and physical behaviors, and adapt to the interaction habit differences of different users through continuous learning and optimization. Output an interpretable infringement risk quantification index, support a hierarchical early warning strategy, and balance security protection and user experience.

[0055] Furthermore, the present application also fuses the time feature of the semantic behavior and the space feature of the physical behavior into a joint feature representation through segmented coding, and calculates a coordination degree parameter within a dialogue turn; based on the reinforcement learning model, iteratively optimize the semantic-physical feature matching weight, and finally generate a dynamic parameter reflecting the behavior association strength through cross-turn fusion operation. This process realizes deep coupling of semantic understanding and behavior analysis, enabling the machine to understand the consistency of "what to say" and "what to do" like a human.

[0056] Moreover, break through the limitations of traditional static rule detection, and establish a dynamic association evaluation model with time series perception ability. By quantifying the semantic-behavior deviation degree, effectively identify infringement behaviors disguised as normal operations (such as the voice command "save file" actually triggering data stealing operations).

[0057] These aspects or other aspects of the present application will be more clearly understood in the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0059] Figure 1 It is a flowchart of a software copyright protection detection method based on semantic analysis provided by an embodiment of the present application;

[0060] Figure 2 It is a schematic structural diagram of a software copyright protection detection system based on semantic analysis provided by an embodiment of the present application;

[0061] Figure 3 It is a schematic structural diagram of a computing device provided by an embodiment of the present application. Detailed implementation manners

[0062] To enable those skilled in the art of the present technology to better understand the solutions of the present application, the following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with the accompanying drawings in the embodiments of the present application.

[0063] In some processes described in the specification, claims and the above accompanying drawings of the present application, there are multiple operations that appear in a specific order. However, it should be clearly understood that these operations can be executed not in the order in which they appear in this article or in parallel. The operation numbers such as 101, 102, etc. are only used to distinguish different operations, and the numbers themselves do not represent any execution order. In addition, these processes can include more or fewer operations, and these operations can be executed in sequence or in parallel. It should be noted that the descriptions such as "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor do they limit that "first" and "second" are of different types.

[0064] Researchers have found that existing software copyright protection technologies are difficult to effectively identify infringement behaviors implemented through complex interaction behaviors, especially unable to accurately capture the implicit correlation between user semantic instructions and actual operations, resulting in a high missed detection rate for infringement behaviors disguised as normal operations. Based on this, a software copyright protection detection method based on semantic analysis is provided. This method can achieve end-to-end detection from user interaction intentions to infringement risk determination through multi-modal behavior data collection, semantic-physical behavior coupling analysis, and dynamic association strength evaluation. The technical solutions of the present application are applicable to intelligent software interaction scenarios that need to distinguish normal user operations from malicious infringement behaviors, such as professional application fields like design software and financial trading systems.

[0065] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0066] Figure 1 The flowchart of a software copyright protection detection method based on semantic analysis provided by an embodiment of the present application is as Figure 1 shown, and the method includes:

[0067] Step 101: In the interaction scenario between the user terminal and the software copyright protection detection system, collect the physical behavior signals and semantic feedback behavior data generated by the user operation track.

[0068] In this step, the user terminal refers to devices such as computers and mobile phones used by users. The physical behavior signals include mouse movement tracks, touch screen operation coordinates, key sequences, etc. The semantic feedback behavior data represents the instruction content input by the user through means such as voice and text.

[0069] In the embodiments of the present application, through the data acquisition module installed on the user terminal, the operation track of the user on the software interface is recorded in real time as the physical behavior signal. At the same time, the semantic instruction of the user is obtained through voice recognition or text input as the semantic feedback behavior data. Both types of data are provided with accurate timestamps to ensure synchronization. The physical behavior signal contains metadata such as coordinate positions and operation types, and the semantic feedback behavior data has been subjected to preliminary word segmentation and intention classification processing.

[0070] For example, in the design software usage scenario, the user says "Move the selected graphic 5 centimeters to the right" (semantic feedback) through voice, and at the same time drags the graphic on the touch screen with a finger (physical behavior). The system records the voice waveform and the text transcription result as semantic data, and records the touch coordinate sequence of the finger as physical behavior data, and both are marked with the same time interval.

[0071] Step 102: Combine the pre-trained semantic analysis model to perform coupled analysis on the physical behavior signal and the semantic feedback behavior data to generate a user intention representation sequence.

[0072] In this step, the semantic analysis model represents a neural network model that has been trained to understand natural language instructions. The user intention representation sequence represents a vector sequence that reflects the true intention of the user.

[0073] In the embodiments of the present application, the collected physical behavior signals are converted into spatial trajectory features, and the semantic feedback data is converted into text feature vectors. The attention mechanism of the semantic analysis model is used to establish the association between the two, analyze the matching degree in the time dimension, and output a representation sequence containing the intention confidence. The model will judge whether the physical behavior accurately executes the semantic instruction and generate an intention representation with time sequence marks.

[0074] For example, continuing with the previous example, the system converts the voice instruction of "move 5 centimeters to the right" into a movement intention vector, and at the same time converts the touch trajectory into displacement features. Through model calculation, the intention matching degree of this operation is 0.92 (full score 1.0), indicating that the user did execute the operation according to the instruction.

[0075] Step 103: Through a pre-trained reinforcement learning model, map the semantic feedback behavior and physical behavior in the user intention representation sequence into an association strength parameter.

[0076] In this step, the reinforcement learning model represents a machine learning model that optimizes decisions through a reward mechanism. The association strength parameter represents a numerical value that quantifies the degree of association between semantics and behavior.

[0077] In the embodiments of the present application, the reinforcement learning model receives the intention representation sequence as the state input, analyzes the historical interaction pattern through the policy network, dynamically adjusts the weight ratio of semantic and behavior features, and outputs an association strength parameter reflecting the credibility of the current operation. The model will automatically update the evaluation policy according to the interaction quality of multiple consecutive dialogue rounds.

[0078] For example, the system analyzes the recent 5 operation records and finds that the matching degree between the user's voice instruction and touch operation is stable above 0.9. Therefore, the association strength parameter of this operation is set to 0.93. If a low matching degree operation suddenly appears, this parameter will decrease.

[0079] Step 104: According to the association strength parameter, combined with a preset dynamic risk determination rule, generate a software copyright protection detection result containing a copyright risk index.

[0080] In this step, the dynamic risk determination rule represents the preset multi-level risk judgment conditions. The copyright risk index represents the infringement risk level comprehensively evaluated. The software copyright protection detection result is the comprehensive infringement risk assessment report finally output by the system, including three core elements: the copyright risk index, the risk level identifier, and the specific abnormal behavior type. Among them, the copyright risk index is a continuous value calculated by quantifying the abnormal fluctuation degree of the correlation strength parameter, reflecting the size of the infringement possibility of the current session; the risk level identifier maps the risk index to discrete levels such as "low risk", "medium risk", "high risk", etc., facilitating quick decision-making; the abnormal behavior type details the specific infringement feature patterns detected, such as "semantic-behavior mismatch", "abnormal operation frequency", etc., providing a basis for subsequent handling. This result contains both an intuitive risk rating and retains detailed detection process data, which can be used not only to block infringement operations in real time but also to provide evidence chain support for copyright traceability.

[0081] In the embodiment of the present application, the system maintains a risk determination rule library, which contains the risk levels corresponding to different strength parameter intervals. It monitors the change trend of the correlation strength parameter in real time. When it detects that the parameter continuously decreases or drops suddenly, it conducts a comprehensive evaluation in combination with features such as operation frequency and instruction complexity, and outputs a detection report with a risk level identifier.

[0082] For example, when the system detects that the correlation strength of a certain user's three consecutive operations suddenly drops from 0.9 to 0.4, and is accompanied by abnormal high-frequency operations, it is determined as high risk according to the rule, and a detection result containing a risk warning is generated, indicating that there may be automated infringement operations.

[0083] This method establishes a dynamic correlation evaluation system for user intentions and operation behaviors through multi-modal behavior collection and in-depth semantic analysis, and can accurately identify infringement behaviors disguised as normal operations. Compared with traditional solutions, it improves the detection accuracy in complex interaction scenarios and at the same time maintains a low false positive rate, providing an intelligent solution for software copyright protection. The system has the ability of continuous learning and can adapt to the evolution of new infringement means.

[0084] To solve the problem that it is difficult to accurately quantify the correlation degree between user semantics and operation behaviors in the prior art, in some embodiments, step 103: mapping the semantic feedback behavior and the physical behavior in the user intention representation sequence into a correlation strength parameter through a pre-trained reinforcement learning model, including:

[0085] Step 201: Segment and encode the time distribution feature of the semantic feedback behavior and the spatial distribution feature of the physical behavior respectively according to each dialogue turn to generate a joint feature representation.

[0086] In step 201, the time distribution feature refers to the triggering frequency and interval feature of semantic feedback behavior on the time axis. The spatial distribution feature refers to the coordinate change feature of physical behavior on the operation interface. The joint feature representation is a multi-dimensional vector formed by aligning the two types of features in time. Each dialogue turn is a technical unit naturally formed in the following way: Semantic feedback behavior data: Each complete semantic input of the user (such as a single voice command, text submission) is used as the starting point for turn division. Physical behavior signal: The user operation trajectory (such as a touch screen operation sequence) collected synchronously is aligned with the semantic input in time to form the physical behavior data of the same turn. Time window constraint: When the adjacent semantic feedback interval exceeds a preset duration (such as 10 seconds), it is automatically divided into a new dialogue turn.

[0087] In the embodiment of the present application, the system first marks the time stamps of the voice commands in each dialogue turn, extracts the command interval and duration as time features, and at the same time converts the touch trajectory into spatial features including position coordinates and moving speed. Finally, a joint feature representation containing spatio-temporal characteristics is generated through a feature splicing layer.

[0088] Step 202: Calculate the cooperation degree parameter between the semantic feedback behavior and the physical behavior in each dialogue turn according to the joint feature representation.

[0089] In step 202, the cooperation degree parameter is a scalar value that quantifies the matching degree between semantics and physical behavior within a single dialogue turn.

[0090] In the embodiment of the present application, the system calculates the cosine similarity between the time feature and the spatial feature in the joint feature representation as the basic cooperation degree, and then combines the operation completion degree of this turn for weighted correction to finally obtain a cooperation degree parameter within the range of 0-1.

[0091] Step 203: Based on the cooperation degree parameter, iteratively optimize the association relationship between the semantic feedback behavior and the physical behavior through a reinforcement learning model. The iterative optimization process is a process of adjusting the matching weights between semantic features and physical features.

[0092] In step 203, the matching weight is a dynamic parameter output by the reinforcement learning model and used to adjust the influence of different features on the final evaluation result.

[0093] In the embodiment of the present application, the reinforcement learning model uses the cooperation degree sequence within the historical window as the state input, generates a weight adjustment strategy for semantic and physical features through a policy network, and updates the network parameters using the policy gradient method based on the advantage function, enabling the model to dynamically adjust the feature weights according to the interaction scenario.

[0094] Step 204: Perform a cross-turn fusion operation on the optimized matching weight and the cooperation degree parameter of the current dialogue turn to generate an association strength parameter.

[0095] In step 204, the current conversation turn specifically refers to a single conversation turn being processed (i.e., the latest generated turn data); the current conversation turn refers to the latest independent conversation unit that the reinforcement learning model processes in real time. For example, it is the interaction turn of the latest voice command completed by the user and the corresponding mouse operation. Example: In the continuous conversation sequence [turn 1, turn 2, turn 3], if the system is processing turn 3, then the "current conversation turn" is turn 3. The cross-turn fusion operation is an algorithmic process that combines historical weights and current synergy degrees to calculate the final intensity parameter.

[0096] In the embodiment of the present application, the system adopts a weighted moving average algorithm to fuse and calculate the current synergy degree with the optimized weights of the past three turns. Among them, the weights of the recent turns are assigned higher influence, and finally an associated intensity parameter with temporal continuity is generated. Exemplarily, the associated intensity parameter Rt is obtained by weighted summation of the current synergy degree parameter Ct and the historical parameter Rt-1, where Rt = γCt+(1 - γ)Rt-1, and γ∈(0.7,0.9) is a dynamically adjusted coefficient; the conversion relationship between the synergy degree parameter and the associated intensity parameter is defined by the following features: The synergy degree parameter is an instant matching measure of semantic-physical behavior within a single conversation turn. The associated intensity parameter is a weighted cumulative quantity of cross-turn synergy degree parameters. The current turn synergy degree parameter accounts for the dominant weight (70%-90%), and the influence of the historical turn synergy degree parameter decreases exponentially through a decay factor. The synergy degree parameter is compressed to the [0,1] interval through an S-shaped function. The associated intensity parameter is mapped to the standard risk dimension through a linear transformation. The associated intensity parameter = α×current synergy + β×historical associated intensity, where α + β = 1, and α decreases dynamically as the conversation turn interval increases.

[0097] The following is a specific example:

[0098] The user completes operations in three consecutive dialogue turns in the design software: In the first turn, the user clearly issues the instruction "Enlarge the rectangle by two times" and accurately performs the two-finger enlargement operation. The system converts the voice instruction into a scaling intention vector, and at the same time analyzes the touch trajectory to obtain an actual magnification factor of 2.1 times. The cosine similarity between the calculated time feature (the delay from the instruction issuance to the operation completion is 1.2 seconds) and the spatial feature (the trajectory smoothness is 0.95) is 0.94. After weighting with the operation accuracy, the cooperation degree parameter is obtained as 0.93; in the second turn, the user says "Rotate by 45 degrees" but the actual rotation angle is 30 degrees. After the system detects the angle deviation, the cooperation degree parameter drops to 0.65. The reinforcement learning model adjusts the semantic feature weight from 0.9 to 0.7 according to the historical data (the average cooperation degree of the previous 5 turns is 0.9); in the third turn, the user corrects the operation to a 40-degree rotation. The system calculates the association strength parameter 0.82 using the weighting formula (the current cooperation degree × 0.6 + the historical average weight × 0.4), where the numerical calculations are all based on the preset normalization processing rules and the sliding window statistical algorithm.

[0099] In the embodiment of the present application, this solution establishes a fine-grained behavior feature association evaluation mechanism, enabling the system to accurately perceive the consistency change between the user's intention and the execution behavior, timely discover abnormal operation patterns, and provide accurate quantitative basis for copyright protection. Compared with the static rule method, it improves the recognition ability of hidden infringement behaviors.

[0100] To further improve the accuracy of semantic and behavior association optimization, in some embodiments, step 203: Based on the cooperation degree parameter, the reinforcement learning model iteratively optimizes the association relationship between the semantic feedback behavior and the physical behavior, including:

[0101] Step 301: Separate the time distribution feature corresponding to the semantic feedback behavior and the spatial distribution feature corresponding to the physical behavior from the joint feature representation, and use the time distribution feature as the semantic feature and the spatial distribution feature as the physical feature.

[0102] In step 301, the semantic feature is the time series feature extracted from the voice or text instruction, including the instruction interval and the response delay. The physical feature is the spatial motion feature extracted from the operation trajectory, including the coordinate change rate and the trajectory smoothness.

[0103] In the embodiment of the present application, the system splits the joint feature representation into independent time features and spatial features through the feature decoupling module. The time feature retains the instruction trigger moment and the duration, and the spatial feature retains the key control point coordinates and the moving direction of the operation trajectory.

[0104] Step 302: In the state space of the reinforcement learning model, construct a state representation vector reflecting the fluctuation of the cooperation degree parameter.

[0105] In step 302, the state representation vector is a feature combination that reflects the change trend of the cooperation degree, including the current value, the historical average value, and the fluctuation amplitude.

[0106] In the embodiment of the present application, the system constructs a three-dimensional state vector including the cooperation degree of the current round, the average value of the cooperation degrees of the past three rounds, and the standard deviation of the cooperation degrees of the last five rounds, so as to characterize the current interaction quality state of the system.

[0107] Step 303: Process the state representation vector through the policy network of the reinforcement learning model to generate a weight adjustment policy.

[0108] In step 303, the weight adjustment policy is an adjustment scheme output by the policy network, including the weight change direction and amplitude of semantic features and physical features, and is used to adjust the association relationship between the semantic features and the physical features.

[0109] In the embodiment of the present application, the policy network uses a fully connected layer structure to process the state vector and outputs an adjustment policy including two components: the semantic feature weight adjustment amount (in the range of -0.1 to +0.1) and the physical feature weight adjustment amount (in the range of -0.1 to +0.1), and the sum of the two remains constant.

[0110] Step 304: Perform an iterative update operation on the matching weights between the semantic features and the physical features according to the weight adjustment policy.

[0111] In step 304, the iterative update operation is a process of dynamically adjusting the feature weights according to the policy, and the new weight = the original weight + the adjustment amount.

[0112] In the embodiment of the present application, the system maintains the current weight values of semantic and physical features (both are initially 0.5), and fine-tunes them according to the output of the policy network in each round, but limits the maximum adjustment amplitude in a single round not to exceed 0.15 to ensure the smooth change of the weights.

[0113] The following is a specific example:

[0114] When the user performs graphic editing in the design software, the system first converts the speech waveform of the "double the magnification of the rectangle" instruction in the first round into time distribution features (including an instruction duration of 0.8 seconds and a response delay of 0.4 seconds), and at the same time converts the two-finger operation trajectory on the touch screen into spatial distribution features (including a starting coordinate difference of 3.2 cm and a trajectory curvature of 0.02). At this time, the weights of both the initial semantic features and physical features are 0.5; when there is a deviation between the "rotate 45 degrees" instruction in the second round and the actual 30-degree rotation, the system constructs a state representation vector including the current cooperation degree of 0.65, the average cooperation degree of the previous three rounds of 0.85, and a fluctuation amplitude of 0.2. The policy network calculates based on this vector that the weight of the semantic features should be reduced by 0.15 (the new weight 0.35 = the original weight 0.5 - 0.15), and the weight of the physical features increases by 0.15 accordingly; after the user corrects the rotation angle to 40 degrees in the third round, the system recalculates to obtain a cooperation degree of 0.8, and the state vector is updated to [0.8, 0.8, 0.15]. The policy network outputs an adjustment strategy with the semantic weight increased by 0.1 (the final weight 0.45 = 0.35 + 0.1).

[0115] In the embodiments of the present application, this solution enables the system to automatically balance the contribution degrees of semantic and behavioral features according to the user's real-time interaction quality by establishing a dynamic weight adjustment mechanism, which not only ensures a high matching degree for normal operations but also can quickly respond to abnormal behavior patterns, effectively improving the adaptability and robustness of infringement detection.

[0116] To further improve the accuracy of the weight adjustment strategy, in some embodiments, step 303: the processing of the state representation vector to generate a weight adjustment strategy includes:

[0117] Step 401: Decompose the state representation vector into the immediate features of the current conversation turn and the memory features of a preset number of historical conversation turns to form a multi-level feature combination.

[0118] In step 401, the immediate features refer to the cooperation degree parameters and their derivative features newly generated in the current turn. The memory features refer to the feature patterns with continuous influence extracted from historical turns. Relationship with historical conversation turns: Historical turn data (such as turn 1, turn 2) needs to be combined, but "the current conversation turn" only refers to the independent data of the latest turn (turn 3). The multi-level feature combination specifically refers to the combination composed of the following two fixed levels: The first level: the immediate features of the current conversation turn (the latest single-turn data), the second level: the memory features of a preset number of historical conversation turns (historical multi-turn data). The multi-level feature combination is a feature representation system that structurally integrates the user interaction behavior features according to different time dimensions.

[0119] In the embodiments of the present application, the system adopts a sliding window mechanism, splitting the state representation vector into immediate features such as the real-time collaboration degree and response speed in the current round, as well as memory features such as the average collaboration degree and trend slope in the past five rounds, and forming a feature combination including near, medium, and far three-layer time scales through feature splicing.

[0120] Step 402: Perform cross-round correlation analysis on the multi-level feature combination, and determine the dynamic association pattern between the semantic feature and the physical feature according to the cross-round correlation analysis result.

[0121] In step 402, the dynamic association pattern is a feature law reflecting the interaction relationship between semantics and physical behavior in the time dimension.

[0122] In the embodiments of the present application, the system calculates the mutual information between the immediate feature and each layer of memory feature, identifies the correlation pattern between the semantic feature weight and the physical feature weight on short-term and long-term time scales, and when it detects that the importance of the recent semantic feature increases while the long-term trend decreases, it is marked as the "weakening of semantic dominance" association pattern.

[0123] Step 403: Generate dynamic adjustment parameters according to the dynamic association pattern.

[0124] In step 403, the dynamic adjustment parameter is a quantization index guiding the direction and amplitude of weight change. It includes the increase amplitude and decrease direction of the matching weight, and sets the adjustment amplitude threshold by analyzing the fluctuation range of the collaboration degree parameter in historical dialogue rounds.

[0125] In the embodiments of the present application, the system selects the corresponding basic adjustment amount from the preset parameter library according to the identified association pattern type, and then performs dynamic scaling in combination with the degree of deviation of the current collaboration degree from the historical average value, generating an adjustment parameter including a symbol (+ / −) and an amplitude value (0 - 0.2).

[0126] Step 404: Perform policy fusion on the dynamic adjustment parameter and the collaboration degree parameter of the current dialogue round to generate a weight adjustment policy.

[0127] In step 404, policy fusion is a decision-making process that combines dynamic adjustment with the current state.

[0128] In the embodiments of the present application, the system fuses the dynamic adjustment parameter and the current collaboration degree parameter through weighted summation, where the collaboration degree parameter modulates the adjustment amplitude as a credibility factor, and the finally generated policy includes a semantic weight change amount and a physical weight reverse change amount, and the absolute values of the two are equal.

[0129] The following is a specific example:

[0130] When a user edits an engineering drawing using software, the system first processes the instruction "draw a straight line with a length of 10 cm" in the first round. It converts the voice instruction into time features (including an instruction clarity score of 0.9 and a response time of 1.5 seconds), and at the same time converts the mouse trajectory into spatial features (including an actual drawing length of 9.8 cm and a straight line deviation of 0.03). The initial semantic and physical feature weights are each set to 0.5. When in the second round the user issues the instruction "thicken the line to 3 mm" but the actual operation is set to 1 mm, the system decomposes the state representation vector into: immediate layer features [current cooperation degree 0.6, operation accuracy 0.65], short-term memory layer features [average cooperation degree in the past 3 rounds 0.88, trend change rate -0.12], and long-term memory layer features [user historical benchmark cooperation degree 0.9, operation stability 0.85]. By calculating the weighted correlation of each layer of features (immediate layer weight 0.6, short-term layer 0.3, long-term layer 0.1), it identifies the dynamic association pattern of "sudden drop in operation accuracy". According to the preset pattern-parameter mapping table, the basic adjustment amount corresponding to this pattern is a semantic weight of -0.12. Then, combined with the degree to which the current cooperation degree deviates from the historical average by 0.3, the final dynamic adjustment parameter is calculated as -0.144 according to the formula (basic adjustment amount × deviation coefficient 1.2). Finally, this parameter and the current cooperation degree of 0.6 are weighted and fused (adjustment parameter × 0.7 + cooperation degree × 0.3) to generate an actual adjustment strategy with a semantic weight of -0.1, reducing the semantic weight from 0.5 to 0.4 and correspondingly increasing the physical weight to 0.6. All numerical calculations are based on the preset feature weighting formula and dynamic adjustment rule library.

[0131] In the embodiment of the present application, this solution enables the weight adjustment strategy to take into account both the immediate interaction state and the long-term behavior pattern by establishing a multi-time-scale feature analysis mechanism, promptly responding to abnormal operations and avoiding over-adjustment, ensuring the system maintains a stable infringement recognition ability in complex interaction scenarios, and at the same time reducing interference with normal user operations.

[0132] To further improve the accuracy of calculating the cooperation degree parameter, in some embodiments, step 202: calculating the cooperation degree parameter between the semantic feedback behavior and the physical behavior within each dialogue turn according to the joint feature representation includes:

[0133] Step 501: Extract the time interval sequence of the semantic feedback behavior and the spatial displacement sequence of the physical behavior from the joint feature representation.

[0134] In step 501, the time interval sequence refers to the sequence composed of the time differences between semantic instructions. The spatial displacement sequence refers to the sequence of the moving distances of consecutive coordinate points in the operation trajectory.

[0135] In the embodiment of the present application, the system separates the instruction trigger timestamp sequence and the operation coordinate sequence from the joint feature representation, calculates the time difference between adjacent instructions as the time interval sequence, and calculates the Euclidean distance between consecutive coordinate points as the spatial displacement sequence.

[0136] Step 502: Compare the synchronization of the change trends of the time interval sequence and the spatial displacement sequence within the same dialogue turn.

[0137] In step 502, the same dialogue turn refers to the time range within the same independent dialogue interaction unit, which is used to define the data alignment boundary between semantic feedback behavior and physical behavior. Example: When the user says "zoom in on the picture" (semantic) and performs a pinch gesture (physical) in the Nth dialogue turn, both data come from the time period of turn N. The current dialogue turn refers to the latest dialogue turn being processed by the system in real time, which is used to identify the time reference when calculating the cooperation degree parameter. The same dialogue turn is a subset of the current dialogue turn: when the processing system is in turn N, the "same dialogue turn" specifically refers to the data alignment within turn N, and the current dialogue turn refers to the whole of turn N. The synchronization comparison is a process of aligning and analyzing the change trends in the time dimension and the spatial dimension.

[0138] In the embodiment of the present application, the system uses the dynamic time warping algorithm to non-linearly align the time interval sequence and the spatial displacement sequence, and calculates the matching degree of the two sequences in terms of shape similarity and change phase.

[0139] Step 503: Generate a behavior coupling index according to the synchronization comparison result, and the behavior coupling index is obtained by quantifying the matching degree between the time interval change and the spatial displacement change.

[0140] In step 503, the behavior coupling index is a numerical value that quantifies the temporal correlation strength between semantic and physical behaviors. The time interval sequence refers to the sequence of semantic feedback behavior interval data extracted from the joint feature representation and arranged in chronological order, for example: [the interval Δt1 between instruction A and instruction B, the interval Δt2 between instruction B and instruction C,...] (raw data layer). The time interval change refers to the analysis of the dynamic change characteristics of the sequence, including: the difference between adjacent intervals (such as Δt2 - Δt1), the change direction (increase / decrease), and the change rate (difference / time unit) (feature analysis layer). The spatial displacement sequence refers to the sequence of spatial coordinate changes of physical behavior at consecutive time points, for example: [the displacement d1 from position coordinates (x1, y1) to (x2, y2), the displacement d2 from (x2, y2) to (x3, y3),...] (raw data layer). The spatial displacement change refers to the quantification of the dynamic characteristics of the displacement sequence, including: the increase or decrease of the displacement amount (such as d2 - d1), the consistency of the movement direction (such as the trajectory turning angle), and the displacement acceleration (the change in displacement amount per unit time) (feature analysis layer).

[0141] In the embodiment of the present application, the system calculates the double-sequence correlation coefficient after synchronous comparison as the basic coupling index, and then combines the number of matching extreme points of the sequences for weighted correction to obtain the behavior coupling index within the range of 0-1.

[0142] Step 504: Based on the behavior coupling index, combine the trigger density of the semantic feedback behavior and the trajectory coverage area of the physical behavior in the current conversation turn to calculate the synergy parameter.

[0143] In step 504, the trigger density is the number of semantic instructions per unit time. The trajectory coverage area is the influence range of the physical behavior on the operation interface.

[0144] In the embodiment of the present application, the system performs weighted fusion of the behavior coupling index with the trigger density (number of instructions / round duration) and the trajectory coverage area (area of the circumscribed rectangle of the operation trajectory), and finally generates a synergy parameter. Exemplarily, the scenario is set as follows: The user interacts through voice commands (semantic feedback behavior) and stylus operations (physical behavior) in a graphics editing software. The system continuously records data for 5 conversation rounds: Input data preparation, Round 1: Semantic time interval sequence: [Interval of 1.5 seconds between command "A" and command "B"], Physical space displacement sequence: [Total length of the stylus movement trajectory is 4 cm], Round 2: Semantic time interval sequence: [Interval of 1.2 seconds between command "B" and command "C"], Physical space displacement sequence: [Total length of the trajectory is 5 cm], Round 3: Semantic time interval sequence: [Interval of 0.8 seconds between command "C" and command "D"], Physical space displacement sequence: [Total length of the trajectory is 7 cm]. Current round (Round 4): Semantic trigger density: 3 instructions per second, Physical trajectory coverage area: 8 cm². Synchronization comparison result: In Rounds 1 - 3, the shortening of the time interval (1.5 s → 1.2 s → 0.8 s) is positively correlated with the increase in displacement (4 cm → 5 cm → 7 cm). Calculated values of the behavior coupling index: [Round 1: 0.6, Round 2: 0.7, Round 3: 0.9] (range 0 - 1, the larger the value, the higher the synergy). Historical round trend: The behavior coupling index continuously rises (0.6 → 0.7 → 0.9), indicating an improvement in the user's operation proficiency. Average change slope: +0.15 / round. Combining the characteristics of the current round, trigger density correction: The current semantic instructions are dense (3 times per second), exceeding the historical average (1.5 times per second) → Trigger correction factor α = 1.2, Trajectory area correction: The current physical operation area (8 cm²) is larger than the historical maximum trajectory (7 cm²) → Area correction factor β = 0.9, Synergy parameter calculation, Basic value: Latest value of the historical behavior coupling index (0.9) × Trend coefficient (1 + 0.15) = 1.035. Final value: Synergy parameter = Basic value × α × β = 1.035 × 1.2 × 0.9 ≈ 1.12 (after normalization, the value is 0.92, range 0 - 1).

[0145] The following is a specific example:

[0146] When the user performs the "adjust graphic transparency" operation in the design software, the system first extracts the time interval sequence of this round (including the 1.5-second interval between the main instruction "set transparency to 50%" and the supplementary instruction "reduce by 10%"), and at the same time extracts the spatial displacement sequence of the touch screen sliding operation (sliding distance of 4.2 cm, including 3 key coordinate points); by comparing the two sequences through the dynamic time warping algorithm, the calculated trend similarity is 0.88 (full score 1.0), and 2 characteristic inflection points in the sequence are detected to match exactly, and the behavior coupling index is corrected to 0.91; the trigger density of the current round is 2 instructions / 3-second operation duration = 0.67, and the trajectory coverage area is 8.4 square centimeters (sliding trajectory width 2 cm × length 4.2 cm). According to the preset weight distribution formula (coupling index × 0.6 + trigger density × 0.2 + trajectory area coefficient × 0.2), where the area coefficient = 8.4 / 10 (maximum reference area 10 square centimeters), the final calculated synergy parameter = 0.91 × 0.6 + 0.67 × 0.2 + 0.84 × 0.2 = 0.85. This value and the synergy calculated in subsequent rounds are jointly input into the reinforcement learning model for generating the association strength parameter.

[0147] In the embodiment of the present application, through multi-dimensional feature fusion calculation, the synergy parameter can comprehensively reflect the association characteristics of semantics and physical behavior in terms of time sequence, frequency and space, capturing both macroscopic behavior patterns and microscopic operation details, providing a more reliable feature basis for subsequent infringement determination.

[0148] To further improve the accuracy of user intention recognition, in some embodiments, step 102: combining the pre-trained semantic analysis model to perform coupling analysis on the physical behavior signal and the semantic feedback behavior data to generate a user intention representation sequence, including:

[0149] Step 601: Input the physical behavior signal into the physical feature encoding layer of the semantic analysis model to output a structured physical feature sequence.

[0150] In step 601, the structured physical feature sequence is to convert the original operation trajectory into a feature vector sequence with spatio-temporal structure.

[0151] In the embodiment of the present application, the physical feature encoding layer uses a temporal convolutional network to process the operation trajectory, extracts a three-dimensional feature vector including position coordinates, moving speed and acceleration, divides it into equal-length segments according to a 100-millisecond time window, and outputs a time-tagged feature sequence.

[0152] Step 602: Input the semantic feedback behavior data into the semantic parsing layer of the semantic analysis model to output a standardized semantic feature sequence.

[0153] In step 602, the standardized semantic feature sequence converts speech or text instructions into a semantic vector sequence with a unified structure.

[0154] In the embodiment of the present application, the semantic parsing layer tokenizes and encodes the instructions into 768-dimensional vectors through a pre-trained language model, and aligns them according to the instruction timestamps into a sequence with the same time resolution as the physical features.

[0155] Step 603: Through the cross-attention coupling module of the semantic analysis model, establish a dynamic association mapping between the structured physical feature sequence and the standardized semantic feature sequence to generate behavior association features.

[0156] In step 603, the cross-attention coupling module is the core component for realizing the dynamic interaction between semantics and physical features. The dynamic association mapping means that the semantic analysis model establishes an adaptive association relationship between semantic instructions and physical operations in real time through the cross-attention mechanism. The behavior association features are the output results of the dynamic association mapping and are high-order representations that fuse semantic intentions and physical behaviors.

[0157] In the embodiment of the present application, this module calculates the attention weights of semantic features on physical features, generates physical behavior correction features guided by semantics at each time point, and at the same time calculates the reverse attention of physical features on semantic features to output a two-way feature representation of behavior association.

[0158] Step 604: Use the verification module of the semantic analysis model to perform consistency verification on the behavior association features.

[0159] In step 604, the verification module is used to ensure the logical consistency between semantics and physical features.

[0160] In the embodiment of the present application, this module includes two sub-modules: a temporal logic check and a spatial constraint check. The former detects whether the instruction execution order is reasonable, and the latter verifies whether the operation trajectory conforms to the spatial relationship described in the instruction. Specifically, at the instruction execution layer, verify whether the temporal deviation between the trajectory mutation points in the physical features and the instruction issuance moments in the semantic features is within a preset tolerance interval; at the behavior logic layer, verify whether the correlation between consecutive semantic instructions and the coherence of the corresponding physical behavior trajectories meet the preset matching rules.

[0161] Step 605: Combine the behavior association features that pass the verification in the chronological order of the dialogue turns to generate a user intention representation sequence.

[0162] In step 605, the dialogue turn time sequence refers to the sequence of dialogue units naturally formed during the interaction between the user and the system. Each turn corresponds to a complete semantic-physical behavior interaction cycle, which is strictly sorted according to the chronological order of occurrence. Example: Turn 1 (t1 - t2): User voice command "Open file" + mouse click on the file icon; Turn 2 (t2 - t3): User voice command "Zoom in on the picture" + two-finger zoom on the touch screen; Turn 3 (t3 - t4): User voice command "Save" + keyboard shortcut operation, where t1 < t2 < t3 < t4 strictly advances in time.

[0163] In the embodiment of the present application, the system concatenates the verified features in the order of the dialogue turn time sequence. Each turn feature includes three dimensions: semantic intention label, operation quality score, and execution time efficiency index.

[0164] The following is a specific example:

[0165] The user issues a voice command "Increase the picture brightness by 20%" (semantic feedback) in the design software, and at the same time slides the brightness adjustment slider on the touch screen with a finger (physical behavior). The system first inputs the touch trajectory into the physical feature encoding layer to extract a structured physical feature sequence including the starting coordinates (120, 300), ending coordinates (220, 300), and moving speed of 2.5 cm / s; at the same time, it inputs the voice command into the semantic parsing layer to output a standardized semantic feature sequence including the operation type (brightness adjustment), amplitude parameter (+20%), and time efficiency requirement (execute immediately). The cross-attention coupling module calculates and finds that the brightness adjustment amount corresponding to the actual slider movement distance is 18%, which has a deviation from the 20% of the command, and generates a behavior association feature [operation type matching degree 1.0, amplitude matching degree 0.9 (calculation formula: 18 / 20), time efficiency matching degree 0.95] through dynamic association mapping. The verification module detects that the time taken from the end of the command to the completion of the operation in this turn is 1.2 seconds (within the preset reasonable range of 0 - 2 seconds). After the consistency check passes, a user intention representation sequence is finally generated: [brightness adjustment command, comprehensive matching degree 0.92 (weighted calculation: type weight 0.4×1.0 + amplitude weight 0.4×0.9 + time efficiency weight 0.2×0.95), execution status mark "compliant"], where all values are generated based on the preset feature weighting formula and verification rules, and are used together with the data of subsequent turns for the calculation of the association strength parameter.

[0166] In the embodiment of the present application, this solution enables the system to accurately restore the user's true operation intention by establishing a deep coupling analysis mechanism for multi-modal features, understanding both the surface instruction semantics and the actual operation connotation, providing a high-precision intention understanding basis for copyright infringement determination.

[0167] To further improve the accuracy of copyright risk determination, in some embodiments, step 104: generating a software copyright protection detection result including a copyright risk index according to the association strength parameter in combination with a preset dynamic risk determination rule, includes:

[0168] Step 701: Arrange the association strength parameters in chronological order of conversation turns to form a parameter change curve.

[0169] In step 701, the parameter change curve is a visual representation of the association strength parameter in the time dimension, reflecting the evolution trend of the user behavior pattern.

[0170] In the embodiments of the present application, the system constructs a line chart with the conversation turn as the abscissa and the association strength parameter as the ordinate, and eliminates the instantaneous fluctuations through sliding window smoothing processing, retaining the trend change characteristics. Exemplarily, during the interaction process where the user continuously completes 5 conversation turns, the association strength parameters recorded by the system are in sequence: the 1st turn 0.85 (the user clearly issues the instruction "open file A" and accurately clicks on the corresponding file), the 2nd turn 0.72 (the user's ambiguous instruction "adjust that" combined with a slow slider drag), the 3rd turn 0.65 (the instruction "save" conflicts with the actually triggered deletion operation), the 4th turn 0.41 (high-frequency meaningless voice noise accompanied by random touch screen), the 5th turn 0.38 (the operation timing of detecting the characteristics of an automated script). After connecting these parameters in turn order, the formed parameter change curve shows a monotonically decreasing trend from 0.85 to 0.72 to 0.65 to 0.41 to 0.38.

[0171] Step 702: Extract the key turning points in the parameter change curve as risk analysis nodes.

[0172] In step 702, the key turning point is an extreme point where the change rate in the parameter curve exceeds the threshold, identifying the moment of mutation of the behavior pattern. Specifically, the parameter points in the association strength parameter change curve that meet the following conditions: extreme point: the point where the parameter value reverses directionally relative to the previous N conversation turns, including: the local highest point where the continuous rise turns into a fall, the local lowest point where the continuous fall turns into a rise, mutation point: the point where the change amplitude of the parameter value within a single conversation turn exceeds the third preset threshold, platform breakthrough point: the point where the parameter value first breaks through the stable interval after remaining stable for continuous K turns.

[0173] In the embodiments of the present application, the system calculates the first derivative of the curve, and selects the points where the absolute value of the derivative is greater than 0.15 and lasts for more than two turns, and at the same time satisfies that the difference between the parameter value of this point and the mean values before and after exceeds 0.1.

[0174] Step 703: Set multiple risk determination intervals in the preset dynamic risk determination rule.

[0175] In step 703, the risk determination interval is divided according to the parameter range preset based on the characteristics of the infringement behavior. Each interval corresponds to different abnormal types of the matching of semantic behavior and physical behavior, including: short-term sudden drop interval: identifying the abnormal situation where the decline amplitude of the association strength parameter exceeds the first preset threshold within 3 consecutive dialogue turns; continuous deviation interval: identifying the abnormal situation where the association strength parameter is lower than the second preset threshold for more than 5 consecutive dialogue turns;

[0176] In the embodiment of the present application, the system sets three intervals: safe area (0.8 - 1.0), observation area (0.5 - 0.8), and risk area (0 - 0.5), and each interval corresponds to a different abnormal pattern library and disposal strategy.

[0177] Step 704: According to the distribution position of the risk analysis node in the parameter change curve, match the corresponding risk determination interval to determine the risk level.

[0178] In step 704, the risk level is the initially determined infringement threat level.

[0179] In the embodiment of the present application, the system counts the frequency of the turning points falling into each interval. When three consecutive turning points appear in the risk area, it is determined as a high risk, and a single turning point combined with the frequency in the observation area determines a medium or low risk.

[0180] Step 705: Dynamically correct the risk level by combining the complexity of the semantic feedback behavior and the operation frequency of the physical behavior in the current dialogue turn.

[0181] In step 705, the complexity of the semantic feedback behavior refers to the complexity of the user instruction in terms of logical structure and execution requirements. The operation frequency of the physical behavior refers to the number of physical operations triggered by the user per unit time, which is used to detect automated scripts or abnormal operation patterns. Dynamic correction is a process of adjusting the risk level based on real-time interaction characteristics.

[0182] In the embodiment of the present application, the semantic complexity is calculated by the number of nested instruction layers (weight +0.1 for each layer), the operation frequency is calculated by the number of actions per unit time (weight -0.15 if it exceeds the benchmark value by 50%), and the final correction amount = semantic complexity weight + operation frequency weight.

[0183] Step 706: Combine the mapping relationship between the preset risk level and the copyright risk index to quantify the corrected risk level into the corresponding copyright risk index.

[0184] In step 706, the copyright risk index is a quantitative representation of the risk level.

[0185] In the embodiments of the present application, the system linearly maps the corrected risk level into an exponent, and the specific value is determined by interpolation within the interval according to the correction amount. Exemplarily, in the interaction scenario where the user operates the design software through the touch screen and voice commands: Initial risk level determination: It is detected that the correlation strength parameter drops by 35% in 4 consecutive conversation turns (exceeding 30% of the first preset threshold), and it matches the short-term sudden drop interval, and the initial risk level is determined to be a medium risk. Dynamic correction process: Semantic complexity: The current voice command is "Move the selected layer 5 pixels to the right and then add a blur effect", which contains two nested operation commands, and the correction amplitude is increased by 1 level according to the rules. Operation frequency: It is detected that the finger sliding operation is triggered 8 times, while there is only 1 semantic command (exceeding twice the ratio), and the correction amplitude is reduced by 0.5 level according to the rules. The finally obtained corrected risk level is: Medium (Level 2) +1 - 0.5 = Level 2.5. Quantification conversion example: The corresponding rule between the risk level and the exponent: Level 1 = risk exponent 3, Level 2 = risk exponent 60, Level 3 = risk exponent 90. Perform a linear quantization calculation on Level 2.5: (60 + 90) / 2 = risk exponent 75.

[0186] Step 707: Combine the copyright risk exponent, the corrected risk level, and the software copyright protection strategy corresponding to the copyright risk exponent to generate a software copyright protection detection result.

[0187] In step 707, the software copyright protection strategy is a disposal plan for the risk exponent.

[0188] In the embodiments of the present application, the policy library includes three-level response measures: enhanced log record (exponent > 50), secondary operation verification (exponent > 70), and forced session termination (exponent > 90).

[0189] The following is a specific example:

[0190] The user completes multiple operation rounds continuously in the design software: First, execute the instruction of "move the graphic 5 cm to the right" (associated strength parameter 0.93), then perform the operation of "rotate the graphic 30 degrees" (parameter 0.89), and then three abnormal operations suddenly occur, and the parameters drop to 0.45, 0.38, and 0.33 in sequence. After the system connects the parameters into a change curve, it detects a key turning point in the third round (the parameter drop amplitude is 0.48, exceeding the preset threshold of 0.3), and this point and the subsequent two points all fall into the preset high-risk interval (0 - 0.5), and it is initially determined to be a high-risk level of 85 points; the semantic instruction of the current round is the complex "combine graphics and add texture" (including 3 layers of sub-operations, adjusted by +10 points for each layer, and the complexity correction is +30 points), but it is detected that the operation frequency reaches 5 times per second (the benchmark value is 2 times, and if it exceeds 150%, it is calculated by -15 points for every 50%, and the frequency correction is -45 points), and the final risk level is adjusted to 70 points (85 + 30 - 45); according to the preset linear mapping rule (the high-risk interval of 60 - 100 points corresponds to the risk index of 70 - 100), 70 points are converted into a risk index of 85, triggering the protection strategy of "operation verification + log record", where all values are calculated through the preset weighted formula and interval mapping table.

[0191] In the embodiment of the present application, this solution establishes a multi-dimensional risk assessment system, enabling the copyright protection system to not only capture the abnormal trend of macroscopic behaviors but also respond to the changes in microscopic interaction characteristics, realizing the closed-loop management from risk perception to disposal strategies, and effectively balancing infringement prevention and normal user experience.

[0192] Figure 2 The structural schematic diagram of a software copyright protection detection system based on semantic analysis provided by the embodiment of the present application is as Figure 2 shown, and this system includes:

[0193] The acquisition module 21 is used to acquire the physical behavior signals and semantic feedback behavior data generated by the user operation trajectory in the interaction scenario between the user terminal and the software copyright protection detection system.

[0194] The first generation module 22 is used to perform coupled analysis on the physical behavior signals and the semantic feedback behavior data in combination with a pre-trained semantic analysis model, and generate a user intention representation sequence.

[0195] The mapping module 23 is used to map the semantic feedback behavior and physical behavior in the user intention representation sequence into associated strength parameters through a pre-trained reinforcement learning model.

[0196] The second generation module 24 is used to generate a software copyright protection detection result including a copyright risk index according to the associated strength parameter and in combination with a preset dynamic risk determination rule.

[0197] Figure 2 The described software copyright protection detection system based on semantic analysis can execute Figure 1 The software copyright protection detection method based on semantic analysis described in the illustrated embodiment, and its implementation principle and technical effects will not be elaborated. For the software copyright protection detection system based on semantic analysis in the above embodiment, the specific ways for each module and unit to execute operations have been described in detail in the embodiment related to this method, and will not be elaborated here.

[0198] In a possible design, Figure 2 The software copyright protection detection system based on semantic analysis in the illustrated embodiment can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0199] The storage component 31 stores one or more computer instructions, where the one or more computer instructions are called and executed by the processing component 32.

[0200] The processing component 32 above Figure 1 The software copyright protection detection method based on semantic analysis in the illustrated embodiment.

[0201] Among them, the processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components for executing the above method.

[0202] The storage component 31 is configured to store various types of data to support the operations of the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic memory, flash memory, magnetic disks, or optical disks.

[0203] Of course, the computing device may necessarily further include other components, such as an input / output interface, a display component, a communication component, etc.

[0204] The input / output interface provides an interface between the processing component and the peripheral interface module, and the peripheral interface module may be an output device, an input device, etc.

[0205] The communication component is configured to facilitate communication between the computing device and other devices in a wired or wireless manner, etc.

[0206] Among them, the computing device may be a physical device or an elastic computing host provided by a cloud computing platform, etc. At this time, the computing device may refer to a cloud server, and the above-mentioned processing component, storage component, etc. may be basic server resources leased or purchased from a cloud computing platform.

[0207] The embodiment of the present application further provides a computer storage medium storing a computer program, and when the computer program is executed by the computer, it can implement the above Figure 1 A software copyright protection detection method based on semantic analysis shown in the embodiment.

[0208] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0209] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

[0210] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0211] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of each embodiment of the present application.

Claims

1. A software copyright protection detection method based on semantic analysis, characterized in that, Including: In the interaction scenario between the user terminal and the software copyright protection detection system, collect the physical behavior signals and semantic feedback behavior data generated by the user's operation trajectory; Combined with a pre-trained semantic analysis model, perform coupled analysis on the physical behavior signals and the semantic feedback behavior data to generate a user intention representation sequence; Through a pre-trained reinforcement learning model, map the semantic feedback behavior and physical behavior in the user intention representation sequence into association strength parameters; According to the association strength parameters, combined with a preset dynamic risk determination rule, generate a software copyright protection detection result including a copyright risk index; The step of mapping the semantic feedback behavior and physical behavior in the user intention representation sequence into association strength parameters through a pre-trained reinforcement learning model includes: Segment and encode the time distribution feature of the semantic feedback behavior and the spatial distribution feature of the physical behavior separately according to each dialogue turn to generate a joint feature representation; According to the joint feature representation, calculate the cooperation degree parameter between the semantic feedback behavior and the physical behavior within each dialogue turn; Based on the cooperation degree parameter, use the reinforcement learning model to iteratively optimize the association relationship between the semantic feedback behavior and the physical behavior. The iterative optimization process is a process of adjusting the matching weight between semantic features and physical features; Perform cross-turn fusion operation on the optimized matching weight and the cooperation degree parameter of the current dialogue turn to generate an association strength parameter.

2. The method according to claim 1, wherein The step of iteratively optimizing the association relationship between the semantic feedback behavior and the physical behavior based on the cooperation degree parameter through the reinforcement learning model includes: Separate the time distribution feature corresponding to the semantic feedback behavior and the spatial distribution feature corresponding to the physical behavior from the joint feature representation, and use the time distribution feature as semantic features and the spatial distribution feature as physical features; In the state space of the reinforcement learning model, construct a state representation vector reflecting the fluctuation of the cooperation degree parameter; Process the state representation vector through the policy network of the reinforcement learning model to generate a weight adjustment strategy; According to the weight adjustment strategy, perform iterative update operations on the matching weight between the semantic features and the physical features.

3. The method according to claim 2, wherein The step of processing the state representation vector to generate a weight adjustment strategy includes: Decompose the state representation vector into the immediate features of the current dialogue turn and the memory features of a preset number of historical dialogue turns to form a multi-level feature combination; Perform cross-turn correlation analysis on the multi-level feature combination, and determine the dynamic association pattern between the semantic features and the physical features according to the cross-turn correlation analysis result; Generate a dynamic adjustment parameter according to the dynamic association pattern; Perform policy fusion on the dynamic adjustment parameter and the cooperation degree parameter of the current dialogue turn to generate a weight adjustment strategy.

4. The method according to claim 1, wherein The step of calculating the cooperation degree parameter between the semantic feedback behavior and the physical behavior within each dialogue turn according to the joint feature representation includes: Extract the time interval sequence of the semantic feedback behavior and the spatial displacement sequence of the physical behavior from the joint feature representation; Synchronously compare the change trends of the time interval sequence and the spatial displacement sequence within the same conversation turn; Generate a behavior coupling index according to the synchronous comparison result, and the behavior coupling index is obtained by quantifying the matching degree between the time interval change and the spatial displacement change; Based on the behavior coupling index, combine the triggering density of the semantic feedback behavior and the trajectory coverage area of the physical behavior in the current conversation turn to calculate a synergy parameter.

5. The method according to claim 1, characterized in that Combined with a pre-trained semantic analysis model, perform coupling analysis on the physical behavior signal and the semantic feedback behavior data to generate a user intention representation sequence, including: Input the physical behavior signal into the physical feature encoding layer of the semantic analysis model to output a structured physical feature sequence; Input the semantic feedback behavior data into the semantic parsing layer of the semantic analysis model to output a standardized semantic feature sequence; Through the cross-attention coupling module of the semantic analysis model, establish a dynamic association mapping between the structured physical feature sequence and the standardized semantic feature sequence to generate behavior association features; Use the verification module of the semantic analysis model to perform consistency verification on the behavior association features; Combine the behavior association features that pass the verification in the order of conversation turn time to generate a user intention representation sequence.

6. The method according to claim 1, characterized in that According to the association strength parameter, combine a preset dynamic risk determination rule to generate a software copyright protection detection result including a copyright risk index, including: Arrange the association strength parameters in the order of conversation turn time to form a parameter change curve; Extract the key turning points in the parameter change curve as risk analysis nodes; Set multiple risk determination intervals in the preset dynamic risk determination rule; According to the distribution position of the risk analysis node in the parameter change curve, match the corresponding risk determination interval to determine the risk level; Dynamically correct the risk level by combining the complexity of the semantic feedback behavior and the operation frequency of the physical behavior in the current conversation turn; Combine the mapping relationship between the preset risk level and the copyright risk index to quantify the corrected risk level into the corresponding copyright risk index; Combine the copyright risk index, the corrected risk level, and the software copyright protection strategy corresponding to the copyright risk index to generate a software copyright protection detection result.

7. A software copyright protection detection system based on semantic analysis, characterized in that, Including: An acquisition module for acquiring physical behavior signals and semantic feedback behavior data generated by a user operation trajectory in an interaction scenario between a user terminal and a software copyright protection detection system; A first generation module for performing coupling analysis on the physical behavior signal and the semantic feedback behavior data by combining a pre-trained semantic analysis model to generate a user intention representation sequence; A mapping module for mapping the semantic feedback behavior and the physical behavior in the user intention representation sequence into an association strength parameter through a pre-trained reinforcement learning model; A second generation module for generating a software copyright protection detection result including a copyright risk index according to the association strength parameter and combining a preset dynamic risk determination rule; Mapping the semantic feedback behavior and the physical behavior in the user intention representation sequence into an association strength parameter through a pre-trained reinforcement learning model, including: Segmenting and encoding the time distribution feature of the semantic feedback behavior and the spatial distribution feature of the physical behavior separately according to each dialogue turn to generate a joint feature representation; Calculating a cooperation degree parameter between the semantic feedback behavior and the physical behavior within each dialogue turn according to the joint feature representation; Based on the cooperation degree parameter, iteratively optimizing the association relationship between the semantic feedback behavior and the physical behavior through a reinforcement learning model, and the iterative optimization process is a process of adjusting the matching weight between the semantic feature and the physical feature; Performing a cross-turn fusion operation on the optimized matching weight and the cooperation degree parameter of the current dialogue turn to generate an association strength parameter.

8. A computing device, characterized in that, Including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a software copyright protection detection method based on semantic analysis according to any one of claims 1 to 6.

9. A computer storage medium, characterized in that, Stored with a computer program, when the computer program is executed by a computer, it implements a software copyright protection detection method based on semantic analysis according to any one of claims 1 to 6.

Images