Interactive service channel optimization method based on artificial intelligence

Abstract

The invention relates to the field of network interactive services, and discloses an artificial intelligence-based interactive service channel optimization method, which comprises the following steps of: collecting input contents of a user and carrying out semantic coding to generate an intention vector of a current round; acquiring a historical intention path of a user, and generating a path dependence feature vector based on the historical path; judging the difference between the current intention vector and a historical path predicted value, and when the difference exceeds a threshold value, performing disturbance correction on the current intention vector to obtain a corrected intention vector; fusing the corrected intention vector and the path dependence feature vector to obtain a fused feature vector; and mapping the fusion feature vector to a preset stable subspace to obtain a subspace projection feature vector. According to the method, the user input is converted into the semantic vector, and the weighted calculation is performed in combination with the historical intention path, so that the accurate modeling of the user intention is realized, and the effect of efficiently identifying the complex user demand is obtained.

Description

Technical Field

[0001] This invention relates to the field of online interactive services, and more particularly to a method for optimizing interactive service channels based on artificial intelligence. Background Technology

[0002] With the rapid development of information technology and artificial intelligence, intelligent service systems have been widely applied across various industries. Especially in customer service, intelligent recommendations, and intelligent customer service, the introduction of AI technology has significantly improved work efficiency and service quality. However, while these systems can provide a convenient user experience to some extent, they still face many challenges, particularly limitations in dynamically adapting to user behavior and accurately predicting user needs. Existing technologies often rely on fixed rules and models to identify user needs, but these methods do not truly adjust service responses according to changes in user behavior, lacking flexibility and personalization.

[0003] In traditional intelligent service systems, user input is typically directly translated into corresponding responses. This approach largely relies on rule-driven or template-based models, where the system determines user needs and provides appropriate answers by matching specific rules. While this method has certain advantages in fixed scenarios, it often fails to handle complex and dynamically changing user behaviors. User needs are frequently variable, and traditional methods struggle to cope with these changes, resulting in delayed and inaccurate responses.

[0004] Furthermore, many intelligent service systems focus only on immediate responses to current user input, neglecting the long-term accumulation and influence of users' historical behavior and preferences. In fact, user behavior is time-series and historically dependent; current needs are often influenced by past behaviors. If a system ignores these factors, it often cannot perform effective reasoning and response. Summary of the Invention

[0005] To address the above shortcomings, this invention provides an AI-based method for optimizing interactive service channels, aiming to improve the inaccurate response and poor adaptability of existing intelligent service systems in terms of dynamic recognition of user intent, updating of behavioral preferences, and generation of service strategies.

[0006] In a first aspect, the present invention provides the following technical solution: an artificial intelligence-based method for optimizing interactive service channels, comprising:

[0007] Collect user input and perform semantic encoding to generate the intent vector for the current round;

[0008] Obtain the user's historical intent path and generate a path dependency feature vector based on the historical path;

[0009] Determine the difference between the current intent vector and the historical path prediction value. When the difference exceeds a threshold, perform perturbation correction on the current intent vector to obtain the corrected intent vector.

[0010] The modified intent vector is fused with the path-dependent feature vector to obtain a fused feature vector;

[0011] The fused feature vectors are mapped to a preset stable subspace to obtain the subspace projection feature vectors;

[0012] The user behavior preference profile is updated based on the subspace projection feature vector to generate an updated behavior preference vector;

[0013] When the updated behavior preference vector meets the convergence condition, a service strategy is generated based on the behavior preference vector to optimize the user's service channel response.

[0014] Preferably, the historical intent path consists of multiple intent vectors prior to the current round, and weighted calculations are performed based on weight values ​​set according to time order to generate the path-dependent feature vector, wherein the intent weight value of more recent time points is higher than the weight value of more distant time points.

[0015] Preferably, the disturbance correction includes:

[0016] Predict the expected vector of the current semantic input based on the historical intent path;

[0017] Calculate the difference vector between the current intent vector and the expected vector;

[0018] The difference vector is multiplied by the perturbation adjustment matrix and then added to the current intention vector to obtain the corrected intention vector.

[0019] Preferably, the stable subspace is constructed by dimensionality reduction or clustering of large-scale historical user behavior data, and the mapping process includes projecting the fused feature vector onto the nearest point in the stable subspace in the sense of Euclidean distance.

[0020] Preferably, the behavior preference vector is updated using a linear fusion method, which combines the behavior preference vector of the previous round with the current round subspace projection feature vector according to a fixed update coefficient to generate the updated behavior preference vector.

[0021] Preferably, the convergence condition includes: the difference between the current behavior preference vector and the previous behavior preference vector is within a preset threshold, and the current behavior preference vector is located within the stable subspace.

[0022] Preferably, the service strategy is based on the decision result generated by the current behavior preference vector input to the strategy selection model. The decision result is used to determine whether to perform manual transfer, recommend preset content, change service priority, or switch service channels.

[0023] Secondly, the present invention provides the following technical solution: an interactive service channel optimization system based on artificial intelligence, comprising:

[0024] The input acquisition module is used to receive user input information;

[0025] The intent encoding module is used to convert user input into the current intent vector;

[0026] The path modeling module is used to construct historical intent paths and generate path-dependent feature vectors;

[0027] The perturbation correction module is used to determine semantic deviations based on historical path prediction results and to perturb and correct the current intent vector.

[0028] The subspace mapping module is used to map the fused feature vectors to a stable subspace;

[0029] The profile update module is used to update the behavior preference vector based on the projection results;

[0030] The strategy generation module is used to generate service response strategies based on the converged behavior preference vector.

[0031] Thirdly, the invention provides the following technical solution: a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned artificial intelligence-based interactive service channel optimization method.

[0032] Fourthly, the present invention provides the following technical solution: a readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned artificial intelligence-based interactive service channel optimization method.

[0033] The present invention has the following beneficial effects:

[0034] 1. In this invention, by converting user input into semantic vectors and combining them with historical intent paths for weighted calculation, accurate modeling of user intent is achieved, resulting in the ability to efficiently identify complex user needs. By optimizing the semantic understanding process, errors in intent recognition are reduced, and the system's ability to process diverse user inputs is improved.

[0035] 2. In this invention, by introducing a disturbance correction module, automatic correction can be performed when a significant difference is detected between the input intent and the predicted intent. This achieves the ability to provide accurate service even in interference environments, effectively eliminating input noise and errors. This innovative improvement enhances the system's robustness, enabling it to respond effectively to uncertain inputs.

[0036] 3. In this invention, a stable subspace mapping and dynamic profile update mechanism are used to achieve deep learning and adaptive adjustment of user behavior and preferences, resulting in the ability to accurately identify and predict changes in user needs. This technological improvement enables the system to better adapt to long-term changes in user needs and optimizes personalized recommendation and service strategies.

[0037] 4. In this invention, by generating the optimal service strategy based on user profile updates and convergence judgments, the service channel response strategy is automatically and dynamically adjusted according to user behavior, thereby improving the accuracy and efficiency of service response. This technological improvement solves the problems of slow or inaccurate response in traditional service strategies, enhancing the system's intelligence and automation level. Attached Figure Description

[0038] Figure 1 This is a flowchart of the interactive service channel optimization method based on artificial intelligence proposed in this invention;

[0039] Figure 2 This is a system architecture diagram of the AI-based interactive service channel optimization system proposed in this invention. Detailed Implementation

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Example 1

[0042] Reference Figure 1 In the first embodiment of the present invention, the present invention provides an artificial intelligence-based interactive service channel optimization method, comprising:

[0043] Collect user input and perform semantic encoding to generate the intent vector for the current round;

[0044] Specifically, the system receives natural language information input by users through interactive channels (such as mobile apps, web pages, and voice recognition interfaces). User input can be a simple sentence or text from a multi-turn dialogue. When the user input is brief, the system directly performs semantic analysis through the text processing module. If the input text is long, it will be segmented into sentences and words to ensure that each part of the information is effectively extracted.

[0045] After processing user input, the system needs to semantically encode the input. This step is crucial because natural language has rich grammatical structures and semantic polysemy. To accurately understand user intent, the system employs deep learning-based semantic encoding techniques, such as BERT (Bidirectional Encoder Representations from Transformers) or other similar pre-trained models. These models learn contextual information from large corpora, enabling them to not only understand vocabulary but also capture complex relationships between words.

[0046] Specifically, the user's input text is converted into fixed-dimensional word vectors through a word embedding layer, and then processed by a semantic encoding model. The result of semantic encoding is a vector containing user intent information, called an intent vector. These vectors represent the user's true intent in the current dialogue turn, including information such as user emotions, needs, and instructions. The system uses this vector to further analyze and identify the user's intent.

[0047] In the formula, let's assume the user input text is... ,in Indicates the first The intent vector generated after semantic encoding of a word or character can be represented as:

[0048] ;

[0049] in, It is the final intent vector, representing the semantic information of the current round, and Encoder is the deep learning model used.

[0050] The core of this process lies in handling the diversity and complexity of user input. Deep learning models, especially those based on the Transformer architecture, utilize self-attention to capture both local and global information in the text simultaneously. For example, input text may sometimes contain ambiguous words or phrases that traditional methods struggle to handle accurately. Deep learning models, however, can mitigate the impact of such ambiguity through contextual understanding.

[0051] Each input word is mapped to a fixed-size vector space. These word vectors undergo multiple processing steps, ultimately resulting in a fixed-dimensional vector. This vector not only represents the meaning of a single word but also comprehensively represents the word's context within a sentence and its relationships with other words.

[0052] This process ensures that the system can understand complex language structures, reducing its strict reliance on grammar and sentence patterns and focusing more on semantic understanding. In this way, the system improves its ability to recognize user intent, obtaining accurate intent vector representations for both short commands and complex queries.

[0053] Obtain the user's historical intent path and generate a path-dependent feature vector based on the historical path. The historical intent path consists of multiple intent vectors from previous rounds, and weights are set according to time order to generate a path-dependent feature vector. The weight of intents at more recent time points is higher than that at more distant time points.

[0054] Specifically, a user's historical intent path consists of multiple intent vectors from previous rounds. Each time a user interacts, the system generates an intent vector for the current round and stores it in the historical intent path. This historical path includes not only the user's intent in a single conversation but also the intents from previous rounds of conversation. The historical intent path provides a complete user behavior trajectory, helping the system understand changes in the user's needs at different points in time.

[0055] In practice, the construction of historical intent paths is a dynamically updated process. As users continue to interact with the system, new intent vectors are added to the historical path, forming a sequence containing multiple intent vectors. Suppose a user... Intent vectors were generated at each time step. So, the historical path It includes points in time. up to the current moment The set of all intent vectors:

[0056] ;

[0057] The key value of historical paths lies in their ability to reflect long-term trends in user behavior and needs. To effectively utilize the information in historical intent paths, the system weights each intent vector in the path according to its chronological order, generating a path-dependent feature vector. The weighting follows a simple principle: intents from more recent times have higher weights, while intents from more distant times have lower weights. The significance of this weighting strategy is that the current user's intent is often more influenced by recent interactions, while the influence of earlier intents gradually weakens.

[0058] Therefore, assigning higher weight to more recent intents can better capture real-time changes in user needs. Assuming each historical intent vector... The weight in the historical path is ,and It is a decreasing function related to time intervals, and the feature vector of the weighted historical path can usually be calculated in the following way. :

[0059] ;

[0060] in, , It is a hyperparameter that controls the decay rate of the time weight. Thus, at the current time point... Corresponding path-dependent feature vector It is the result of a weighted average of historical intent vectors.

[0061] The generation of path-dependent feature vectors relies on dynamically weighted analysis of users' historical behavior. In this way, the system can deeply learn the evolution patterns of users' needs over multiple rounds of interaction. For example, if a user has consistently expressed a need for a product feature or function in previous rounds, the system can identify the strength and trend of this need from the historical path and adjust the response strategy for the current conversation accordingly.

[0062] This weighted calculation method is based on the assumption that "recent impact is greater," which helps the system better understand the underlying intent behind current user needs. For example, when a user repeatedly asks about the details of a certain function in multiple rounds of dialogue, the system can perceive this need through the feature vectors of historical paths and provide more accurate service responses for subsequent dialogues accordingly.

[0063] The difference between the current intent vector and the historical path prediction value is determined. When the difference exceeds a threshold, the current intent vector is perturbed and corrected to obtain a corrected intent vector. The perturbation correction includes:

[0064] Predict the expected vector of the current semantic input based on the historical intent path;

[0065] Calculate the difference vector between the current intent vector and the expected vector;

[0066] The modified intention vector is obtained by multiplying the difference vector by the perturbation adjustment matrix and then adding it to the current intention vector.

[0067] Specifically, when the system obtains the intent vector for the current round... In this case, the first step is to predict the "expected vector" for that round based on the user's historical intent path. This expected vector is inferred based on the user's previous multi-round intent data; that is, the system infers the user's most likely needs and intentions at this moment based on historical interaction records.

[0068] The prediction process relies on long-term patterns and trends in user behavior. Specifically, the system uses historical path-dependent feature vectors. By combining machine learning models (such as recurrent neural networks, Transformers, or other time-series models), the expected intent vector for the current round is generated. This vector represents the predicted user intent based on historical paths, and can be expressed as:

[0069] ;

[0070] Among them, the Predictor is a prediction model learned from historical paths, used to infer the possible intentions of the current user.

[0071] Once the current intent vector is obtained, the difference vector between the current intent vector and the expected vector is calculated. and the expected vector of historical path prediction The next step is to calculate the differences between them. (Difference vector) This represents the deviation between the current user input and the ideal input predicted by the system. The purpose of calculating the difference is to identify any perturbations; a large difference indicates a significant deviation between the current input and the expected input. The difference vector can be calculated as follows:

[0072] ;

[0073] Here, It is the intent vector of the user's current input. It is a predicted intent vector based on historical path prediction. Using this difference vector, the system can quantify the degree of anomaly in the input.

[0074] When the difference vector When the input exceeds a preset threshold, the system considers the current user's input to have a certain degree of abnormality or perturbation. At this point, the system needs to perform perturbation correction on the current intent vector. The core idea of ​​perturbation correction is to adjust the current intent vector through an adjustment matrix to make it more consistent with the system's prediction of user needs.

[0075] Specifically, disturbance correction includes the following two steps:

[0076] 1. Calculate the adjustment matrix: The system needs to dynamically adjust the adjustment matrix based on the differences between historical paths and current intentions. This adjustment matrix is ​​typically obtained through training and represents how the current intent vector is adjusted based on the difference vector. The adjustment matrix is ​​usually a learnable weight matrix that can adjust each dimension of the vector.

[0077] The calculation method is as follows:

[0078] ;

[0079] Among them, the Adjuster is a regulation function trained on historical data, used to calculate how to correct the current intent based on the difference vector.

[0080] 2. Perform correction operations: adjust the difference vector. With adjustment matrix After multiplication, add it back to the current intent vector. The corrected intent vector is obtained. .

[0081] The formula for the correction process is:

[0082] ;

[0083] Thus, the corrected intent vector It is the user intent after perturbation correction, which has eliminated the deviation from the expectation.

[0084] The modified intent vector and the path-dependent feature vector are fused to obtain the fused feature vector;

[0085] Specifically, in the preceding steps, the system has already generated a revised intent vector. This vector reflects the true needs of the current user intent, while eliminating the influence of abnormal input and disturbances. Meanwhile, path-dependent feature vectors... This feature vector is obtained through a weighted calculation of historical intent paths, encompassing the user's behavior and demand patterns across multiple past interactions. This feature vector not only reflects the user's long-term behavioral trends but also provides contextual information relevant to the current input. The modified intent vector and path-dependent feature vector model the user's intent and behavior from different perspectives, representing the user's "immediate intent" and "historical demand" in the current interaction. Fusing these two vectors helps improve the system's comprehensive understanding of user needs.

[0086] The process of fusing the revised intent vector with the path-dependent feature vector aims to create a more accurate description of user needs by integrating current immediate demands with historical behavioral patterns. Essentially, this process combines the two feature vectors through weighting, concatenation, or other appropriate methods to obtain a unified, higher-quality feature vector. The basic idea behind this fusion process is that current user needs (i.e., the revised intent vector) and historical user behavioral patterns (i.e., the path-dependent feature vector) are not independent; they are complementary and interactive. Therefore, fusing these two features can improve the system's accuracy in predicting user needs, ensuring that the system can comprehensively consider both immediate user feedback and long-term needs.

[0087] The fusion method used in this implementation is weighted summation, which is calculated directly on the modified intent vector. and path-dependent feature vectors The weighted sums are then performed to form a new fused feature vector. The weighting coefficients can be set according to the actual situation; a common practice is to give higher weight to the current intent (corrected intent).

[0088] ;

[0089] in, and These are the weight coefficients for correcting the intent vector and the path-dependent feature vector, typically... This is because the influence of current intentions is usually stronger than the influence of historical paths.

[0090] The fused feature vector is mapped to a preset stable subspace to obtain the subspace projection feature vector. The stable subspace is constructed by dimensionality reduction or clustering of large-scale historical user behavior data. The mapping process includes projecting the fused feature vector to the nearest point in the stable subspace in the sense of Euclidean distance.

[0091] Specifically, in the fifth step of this invention, the system maps the fused feature vector to a preset stable subspace, thereby obtaining a subspace projection feature vector within that subspace. This process aims to further improve the stability and discriminativeness of the vector representation, ensuring that user features are not affected by noise or abnormal input in high-dimensional space.

[0092] The so-called "stable subspace" refers to a low-dimensional representation space constructed through the analysis of large-scale historical user behavior data. This space preserves the core distribution characteristics and typical intent patterns in user behavior and has the following two key attributes:

[0093] 1. Low-dimensional representation: Through dimensionality reduction techniques such as Principal Component Analysis (PCA), Autoencoder, t-SNE, or UMAP, the system can extract the main representation directions from high-dimensional user behavior data, thereby constructing a space with lower dimensionality but higher information density. 2. Behavioral stability: In large amounts of historical interaction data, user behaviors and intentions often cluster in specific regions or patterns, exhibiting relatively stable patterns. Clustering methods (such as K-Means, GMM, DBSCAN, etc.) can identify these stable cluster centers, and the space they form can be used as the system's stable intention representation subspace.

[0094] Therefore, a stable subspace can be formalized as a subspace with dimension . ( Linear or nonlinear subspaces:

[0095] ;

[0096] in, It is a set of transformation matrices or basis vectors learned through dimensionality reduction or clustering, representing the directional basis of the stable subspace. Mapping process: The nearest point projection in the Euclidean distance sense fuses the feature vectors. The process of mapping to a stable subspace is essentially about finding elements in the stable subspace that are related to... The point with the closest Euclidean distance is the orthogonal projection point of the vector into the stable subspace. .

[0097] This projection exhibits minimal distortion, meaning it improves the stability and generalization ability of the representation while altering the original features to a minimum.

[0098] The mapping method is as follows: If the stable subspace is a linear space, it is determined by the orthogonal basis matrix. If we represent the projection point, then the projection point is:

[0099] ;

[0100] here, It is the projection operator of the fused vector in the stable subspace.

[0101] If the stable subspace is a nonlinear space composed of multiple cluster centers (e.g., the centers of a Gaussian mixture distribution), then the system will fuse vectors. Project to the nearest cluster center or construct a nonlinear projection function using kernel methods. Thus, we obtain:

[0102] ;

[0103] The user behavior preference profile is updated based on the subspace projection feature vector to generate an updated behavior preference vector. The update of the behavior preference vector adopts a linear fusion method, which combines the previous round behavior preference vector with the current round subspace projection feature vector according to a fixed update coefficient to generate the updated behavior preference vector.

[0104] Specifically, a user behavior preference profile is typically represented by a vector, denoted as . A user profile is an abstract representation of a user's behavioral characteristics, preferences, and intent types over a specific time period. Its core function is to serve as a crucial basis for subsequent recommendation, ranking, retrieval, and interaction strategy selection. A user profile should progressively absorb explicit or implicit feedback from each interaction, reflecting the dynamic evolution of their behavior. It should integrate feature information obtained from real-time interactions with long-term user behavior data to maintain semantic consistency and discriminative power.

[0105] To ensure that the behavior preference vector retains historical information while also reflecting the features generated by the new round of interactions, this invention employs a linear fusion method for profile updating. Specifically, the system updates the profile using the behavior preference vector from the previous round. The subspace projection feature vector of the current round By performing weighted combination, a new behavioral preference vector is obtained. :

[0106] ;

[0107] in: This is the update coefficient, used to control the fusion ratio between historical profiles and current features; a larger coefficient indicates a higher fusion ratio. A smaller value indicates that the system places more emphasis on historical behavior (conservative updates); The value indicates that the system is more sensitive to the current interaction (aggressive update); : The behavioral preference vector saved after the last interaction; : Subspace projection feature vector generated by the current interaction.

[0108] This linear fusion method is simple and efficient. It can adaptively update user profiles without introducing additional complex models, and can be continuously iterated and optimized in each round of interaction.

[0109] When the updated behavior preference vector meets the convergence criteria, a service strategy is generated based on this vector to optimize the user's service channel response. The convergence criteria include: the difference between the current behavior preference vector and the previous behavior preference vector is within a preset threshold, and the current behavior preference vector lies within a stable subspace. The service strategy is based on the decision result generated by the strategy selection model, which is input into the current behavior preference vector. This decision result is used to determine whether to perform manual transfer, recommend preset content, change service priority, or switch service channels.

[0110] Specifically, to ensure sufficient stability and reliability of the behavioral preference vector when generating service policies, the system introduces a dual convergence judgment mechanism:

[0111] Current behavioral preference vector Compared with the previous round vector The Euclidean distance between them is less than a preset threshold.

[0112] ;

[0113] This condition ensures that user preferences do not fluctuate significantly across consecutive rounds, exhibiting behavioral consistency. Furthermore, the current behavioral preference vector should reside in the aforementioned stable subspace. Inside, that is, satisfying:

[0114] ;

[0115] This condition ensures that the semantic location of the current user profile is within a high-density, reliable historical behavior pattern region, thereby enhancing the accuracy of service strategy determination. The system can assess the degree of fit of the preference vector to the stable subspace through projection error, further determining its attribution.

[0116] Once the above convergence conditions are met, the system inputs the current behavior preference vector into the pre-trained policy selection model. The model outputs a set of corresponding service response policies based on the pattern matching relationship between the vector and historical data.

[0117] Strategy selection models are typically built using neural network structures, decision tree models, or reinforcement learning frameworks. They output the optimal combination of response actions based on the semantic location and dimensionality of the action vectors. Their core objective is to make service responses more closely aligned with users' actual needs, thereby improving service efficiency and satisfaction.

[0118] The generated strategies include, but are not limited to, the following categories:

[0119] If the system determines whether a user has entered a complex or high-intent expression stage based on their current behavioral characteristics, it can trigger a human service transfer to improve the service resolution rate.

[0120] If the behavioral preference vector highly matches a certain type of preset content or FAQ knowledge point, the system can directly push content selected from the recommendation module to avoid repeated questions and waste of resources.

[0121] For highly active or high-value users, behavioral preferences can trigger service priority adjustments, such as increasing their processing priority in the service queue or allocating more computing resources in a multi-tasking system.

[0122] When the preference vector indicates a change in the user's preference channel (such as a shift from text to voice), the system can switch service channels based on policy judgments to achieve multimodal conversions such as voice interaction, video guidance, or mixed text and image responses.

[0123] Once a service strategy is generated and executed, the system will monitor user feedback during that round of interaction to evaluate whether the strategy response has achieved the expected goals. If the system detects that the user has re-entered a state of preference fluctuation, or that the strategy execution has failed to effectively address the user's needs, the system will terminate the current strategy path and re-enter the behavioral preference iteration and stabilization subspace projection process to complete closed-loop adaptive optimization.

[0124] This strategy-driven feedback loop mechanism ensures the system's self-regulation capability in dynamic environments, enabling it to cope with the variability, complexity, and personalized needs of user behavior.

[0125] Example 2:

[0126] Reference Figure 2 In a second embodiment of the present invention, the present invention provides an artificial intelligence-based interactive service channel optimization system, comprising:

[0127] The input acquisition module is used to receive user input information;

[0128] The intent encoding module is used to convert user input into the current intent vector;

[0129] The path modeling module is used to construct historical intent paths and generate path-dependent feature vectors;

[0130] The perturbation correction module is used to determine semantic deviations based on historical path prediction results and to perturb and correct the current intent vector.

[0131] The subspace mapping module is used to map the fused feature vectors to a stable subspace;

[0132] The profile update module is used to update the behavior preference vector based on the projection results;

[0133] The strategy generation module is used to generate service response strategies based on the converged behavior preference vector.

[0134] Specifically, the system's operation begins with the input acquisition module, whose core function is to receive and process user input, including various forms such as text and voice. The user's input is converted into a standardized data format, serving as the basis for subsequent processing.

[0135] Next, the intent encoding module performs semantic analysis on the user input, transforming it into a current intent vector. This process typically relies on pre-trained deep learning models, such as BERT or other natural language processing (NLP) techniques, to capture the deep semantics of the user input, thereby providing accurate input for subsequent understanding and decision-making.

[0136] Based on the semantic patterns of historical interactions, the path modeling module constructs historical paths of user intent and generates path-dependent feature vectors related to historical interactions. These feature vectors reflect the user's behavioral trends and preference patterns in past interactions, providing valuable background information for the system's decision-making.

[0137] When there is a discrepancy between the user's current input and historical path features, the perturbation correction module will correct the intent vector. The role of this module is to ensure that the system can adjust the vector representation of the current intent in a timely manner based on the semantic discrepancy between the intent judgment predicted by the historical path and the current input, so as to avoid misunderstanding or inappropriate response.

[0138] The corrected intent vector, along with the path feature vector, is then fed into the subspace mapping module. This module projects the fused feature vector of intent and path information into a stable subspace, ensuring that behavioral analysis and decision-making results are unaffected by noise or bias. In this way, the system can efficiently and stably handle users' changing needs, avoiding unnecessary fluctuations.

[0139] As user interaction continues, the system constantly updates user behavioral preferences. During this process, the profile update module, based on the projection results from the subspace mapping module, updates the user's behavioral preference vector in a timely manner, making it more accurately reflect the user's latest needs and changing interests. This module ensures the continuity and accuracy of the user profile by weightedly fusing historical preferences and current features.

[0140] Once the behavioral preference vector stabilizes and meets the convergence condition, the system uses the strategy generation module to generate corresponding service response strategies. These strategies are generated based on the analysis results of the current behavioral preference vector, and they determine how the system will interact with the user in the next step, such as whether manual transfer is needed, what content to recommend, adjusting service priorities, or switching to other service channels.

[0141] Throughout operation, the system continuously provides dynamic feedback on user needs, ensuring personalized and timely service. With each user interaction, the system is continuously adjusted and optimized to ensure that the final service strategy perfectly matches the user's behavioral patterns and changing needs.

[0142] Example 3

[0143] In a third embodiment of the present invention, based on the same inventive concept, the present invention proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the artificial intelligence-based interactive service channel optimization method of the above embodiments.

[0144] Example 4

[0145] In the fourth embodiment of the present invention, based on the same inventive concept, the present invention proposes a computer, which includes: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory to execute the artificial intelligence-based interactive service channel optimization method of the above embodiment.

[0146] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0147] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing an interactive service channel based on artificial intelligence, characterized in that, The method comprises the following steps: Collecting the input content of the user and performing semantic encoding to generate an intent vector of the current round; Obtaining the historical intent path of the user, and generating a path-dependent feature vector based on the historical path; Judging the difference between the current intent vector and the historical path prediction value, and when the difference exceeds a threshold, performing perturbation correction on the current intent vector to obtain a corrected intent vector; Fusing the corrected intent vector with the path-dependent feature vector to obtain a fused feature vector; Mapping the fused feature vector to a preset stable subspace to obtain a subspace projection feature vector; Updating the user behavior preference portrait based on the subspace projection feature vector to generate an updated behavior preference vector; When the updated behavior preference vector meets the convergence condition, generating a service strategy according to the behavior preference vector for optimizing the user's service channel response mode. 2.The artificial intelligence-based interactive service channel optimization method of claim 1, wherein, The historical intent path is composed of multiple intent vectors before the current round, and a weight value is set according to the time sequence for weighted calculation to generate the path-dependent feature vector, wherein the intent weight value of the recent time point is higher than that of the remote time point. 3.The artificial intelligence-based interactive service channel optimization method of claim 1, wherein, The perturbation correction includes: Predicting the expected vector of the current semantic input based on the historical intent path; Calculating the difference vector between the current intent vector and the expected vector; Multiplying the difference vector by a perturbation adjustment matrix and adding it to the current intent vector to obtain the corrected intent vector. 4.The artificial intelligence-based interactive service channel optimization method of claim 1, wherein, The stable subspace is constructed by dimension reduction or clustering of large-scale historical user behavior data, and the mapping process includes projecting the fused feature vector to the nearest point in the stable subspace in the sense of Euclidean distance. 5.The artificial intelligence-based interactive service channel optimization method of claim 1, wherein, The update of the behavior preference vector adopts a linear fusion method, which combines the last round behavior preference vector and the current round subspace projection feature vector according to a fixed update coefficient to generate an updated behavior preference vector. 6.The artificial intelligence-based interactive service channel optimization method of claim 1, wherein, The convergence condition includes that the difference between the current behavior preference vector and the last round behavior preference vector is within a preset threshold, and the current behavior preference vector is located in the stable subspace. 7.The artificial intelligence-based interactive service channel optimization method of claim 1, wherein, The service strategy is based on the decision result of the strategy selection model input by the current behavior preference vector, which is used to determine whether to perform artificial transfer, recommend preset content, change service priority or switch service channel.

8. An artificial intelligence based interactive service channel optimization system characterized in that, The method for optimizing the interactive service channel based on artificial intelligence according to any one of claims 1-7 comprises: An input collection module for receiving user input information; An intent encoding module for converting user input into a current intent vector; A path modeling module for constructing a historical intent path and generating a path-dependent feature vector; A perturbation correction module for judging semantic deviation based on historical path prediction results and performing perturbation correction on the current intent vector; A subspace mapping module for mapping the fused feature vector to a stable subspace; A portrait updating module for updating the behavior preference vector based on the projection result; A strategy generation module for generating a service response strategy according to the converged behavior preference vector.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the method for optimizing the interactive service channel based on artificial intelligence according to any one of claims 1-7.

10. A readable storage medium, characterized by, The readable storage medium has a computer program stored thereon, and the computer program is executed by the processor to implement the method for optimizing an interactive service channel based on artificial intelligence according to any one of claims 1 to 7.

Images