Knowledge graph-based user intent reasoning system

Abstract

The application relates to the technical field of knowledge graph reasoning, in particular to a user intention reasoning system based on a knowledge graph. The system obtains user business state data and generates a state vector and an intention validity vector, constructs a heterogeneous graph model containing state nodes and intention nodes, analyzes an excitation weight based on a fault level and a historical business handling time length, and obtains an intention activation probability vector based on a business background through a random walk algorithm; a confidence factor is dynamically adjusted based on the amount of valid information in the text, the text semantics and the business background reasoning result are fused, and a hierarchical strategy is executed according to the intention comprehensive evaluation index after fusion. The application converts business background common sense into a calculable probability feature by constructing a state-intention heterogeneous graph and introducing a dynamic fusion mechanism, effectively reduces the failure problem of NLP technology in the semantic fuzzy and multi-state coupling scene, and improves the intention recognition accuracy and service efficiency in the complex business scene.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph reasoning technology, and more specifically to a user intent reasoning system based on knowledge graphs. Background Technology

[0002] With the rapid development of mobile internet services, telecom operators and internet service providers face a massive influx of user inquiries. Current intelligent customer service systems primarily rely on natural language processing (NLP) technology to semantically classify user-input text and thus identify user intent. This single-modal approach is highly effective when handling clearly stated, standard questions, but it has limitations when facing complex business scenarios.

[0003] When users encounter network failures or emergencies, they often only provide vague words or emojis, resulting in a lack of semantic meaning in the text. Furthermore, user intent is often strongly correlated with their current objective business status, such as account arrears or regional network failures, as well as business execution costs. However, existing methods often separate intent recognition from business status queries, lacking a mechanism to map heterogeneous business data into probabilistic features for intent recognition. In addition, intent recommendations are typically based solely on text matching, ignoring the impact of business processing time on user experience. During the cold start phase when user intent is unclear, they fail to prioritize easy-to-use automated services, reducing the initial response rate and system service throughput. Summary of the Invention

[0004] To address the technical problems in existing technologies where relying solely on text semantic analysis fails to accurately identify user intent in scenarios involving ambiguous user expressions, missing textual information, and coupled multi-service states, resulting in low intent recognition accuracy, numerous invalid interactions, and low service efficiency, this invention aims to provide a user intent reasoning system based on a knowledge graph. The specific technical solution adopted is as follows: This invention provides a user intent reasoning system based on knowledge graphs, the system comprising: The state awareness module is used to acquire the user's current business state data and network environment data, and obtain the user's business state state vector through binary mapping; based on the preset intent pre-rule library, it performs validity verification on the candidate intent set and generates an intent validity vector. The heterogeneous graph reasoning module is used to construct a heterogeneous graph model containing state nodes and intent nodes. In the heterogeneous graph model, based on the fault level and historical business processing time in the business state data, the activation weight of state nodes on intent nodes is analyzed in combination with the intent validity vector. Using the state vector as the initial probability distribution, the random walk algorithm is used to perform probability propagation and iterative convergence in the heterogeneous graph model to obtain the intent activation probability vector based only on the business background. The dual-stream fusion decision module is used to perform semantic recognition on the text data input by the user to obtain a semantic classification probability vector; determine the text confidence factor based on the effective information content in the text data; fuse the semantic classification probability vector and the intent activation probability vector based on the text confidence factor to obtain an intent comprehensive evaluation index; and respond to the graded strategy according to the intent comprehensive evaluation index.

[0005] Furthermore, the method for obtaining the state vector includes: A set containing business status elements with two or more dimensions is pre-constructed; user business operation data and network log data are collected and matched with each dimension of business status in the state set. For a successfully matched business status dimension, the corresponding position in the state vector is assigned a value of 1; otherwise, the corresponding position in the state vector is assigned a value of zero, thus generating a state vector.

[0006] Furthermore, the method for obtaining the intent validity vector includes: The pre-defined intent pre-rule library stores the business execution pre-rules corresponding to each candidate intent. The state vector is substituted into the pre-rules of each candidate intent and Boolean logic operation is performed. If the operation result is true, it is marked as a valid identifier in the intent validity vector; if the operation result is false, it is marked as an invalid identifier in the intent validity vector, thus obtaining the intent validity vector.

[0007] Furthermore, the method for obtaining the activation weights includes: For each intent node currently marked as a valid identifier, for any directed edge from the state node to the valid identifier intent node, determine the fault gain factor based on the fault level corresponding to the state node of the directed edge; determine the duration damping factor based on the historical business processing duration corresponding to the intent node of the directed edge. The excitation weight of the directed edge is determined by combining the fault gain factor and the duration damping factor. The excitation weight is positively correlated with the fault gain factor and negatively correlated with the duration damping factor. For each intent node currently marked as invalid, set the activation weight of the directed edge pointing to the invalid intent node to the blocking value.

[0008] Furthermore, the method for obtaining the intent activation probability vector includes: The initial probability distribution of the random walk is obtained by normalizing the state vector; the initial probability distribution is then input into the heterogeneous graph model for probability propagation, and the change in the probability vector is calculated after each round of propagation. When the change is less than the preset convergence threshold, the iteration stops. The probability values ​​corresponding to the intention nodes in the heterogeneous graph model after the iteration stops convergence are extracted and integrated to obtain the intention activation probability vector.

[0009] Furthermore, the method for obtaining the semantic classification probability vector includes: The text data input by the user is segmented and features are extracted to obtain a text feature vector; the text feature vector is input into a pre-trained natural language processing model, and the matching degree between the text feature vector and each standard intent in the candidate intent set is output. The matching degrees are integrated to obtain a semantic classification probability vector.

[0010] Furthermore, the method for obtaining the text confidence factor includes: The system filters out invalid information from the user-input text data, extracts valid content from the text, and counts the amount of valid information. Based on a preset mapping rule between information amount and confidence level, the system substitutes the counted amount of valid information into the mapping rule to match and obtain the corresponding text confidence level factor.

[0011] Furthermore, the method for obtaining the comprehensive evaluation index of intent includes: The semantic classification probability vector is weighted using the text confidence factor; the intent activation probability vector is weighted using the difference between the constant 1 and the text confidence factor; and the weighted sum is used as the comprehensive intent evaluation index.

[0012] Furthermore, the response grading strategy based on the comprehensive evaluation index of intent includes: Sort the comprehensive evaluation indicators of intent in descending order; calculate the difference between the highest and second highest comprehensive evaluation indicators of intent. If the highest intent comprehensive evaluation index exceeds the execution threshold and the difference exceeds the discrimination threshold, the intent business corresponding to the highest intent comprehensive evaluation index will be executed directly. If the highest intent comprehensive evaluation index exceeds the clarification threshold but does not meet the direct execution conditions, then a clarification option containing a preset number of intents corresponding to the highest intent comprehensive evaluation index will be generated. Otherwise, the strategy of transferring to human assistance will be implemented.

[0013] Furthermore, it also includes a parameter update module, which is used to: monitor and record the actual time spent by the user on the recommended intent after the business is completed; and update the historical business processing time parameter of the corresponding intent in the heterogeneous graph model by using an exponentially weighted moving average algorithm, combining the historical business processing time with the actual processing time of this time.

[0014] The present invention has the following beneficial effects: This invention improves intent recognition capabilities under fuzzy semantics by acquiring user business status data and generating state vectors and intent validity vectors. It facilitates the analysis and utilization of implicit correlations between business states (such as faults or overdue payments) and intents, constructing a heterogeneous graph model containing state nodes and intent nodes. It then analyzes activation weights based on fault levels and historical business processing times, and obtains intent activation probability vectors based on business context through a random walk algorithm. By incorporating historical processing times and fault level analysis, it prioritizes short-duration, highly automated services during the cold start phase, reducing user operation barriers and the reliance on human customer service. Finally, it dynamically adjusts confidence factors based on the effective information content of the text, achieving intelligent fusion of text semantics and business context reasoning results, adapting to different user input scenarios. A tiered strategy is implemented based on the comprehensive intent evaluation index after fusion, making intent responses more aligned with actual needs, further improving the system's service adaptability and user experience. This invention, by constructing a state-intent heterogeneous graph and introducing a dynamic fusion mechanism, transforms business background common sense into computable probabilistic features, effectively reducing the failure problem of NLP technology in semantically fuzzy and multi-state coupled scenarios, and improving the accuracy and efficiency of intent recognition in complex business scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a structural diagram of a knowledge graph-based user intent reasoning system provided in one embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a user intent reasoning system based on a knowledge graph proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details a specific solution for a knowledge graph-based user intent reasoning system provided by this invention.

[0020] Please see Figure 1 The diagram illustrates a structure of a knowledge graph-based user intent reasoning system according to an embodiment of the present invention. The system includes: a state awareness module 101, a heterogeneous graph reasoning module 102, a dual-stream fusion decision module 103, and a parameter update module 104.

[0021] The state awareness module 101 is used to acquire the user's current business state data and network environment data, and obtain the state vector of the user's business state through binarization mapping; based on the preset intent pre-rule library, it performs validity verification on the candidate intent set and generates an intent validity vector.

[0022] To transform the dispersed and heterogeneous service status information of users into a form that can be computed by subsequent graph models, it is necessary to construct a unified state representation space and acquire real-time user status data. First, relevant data reflecting the overall picture of user services is acquired. In this embodiment of the invention, the system reads user CRM (Customer Relationship Management) system data and local network operation and maintenance logs in parallel through standardized API interfaces to obtain the user's current service status data and network environment data, such as account balance, remaining data allowance, base station alarm information, and optical attenuation values.

[0023] Since fragmented business and network data cannot be directly processed by subsequent algorithms, they need to be transformed into a standardized vector form. Therefore, a state vector of the user's business status is obtained through binarization mapping. In this embodiment of the invention, a set containing business status elements with two or more dimensions is pre-constructed. This set covers all key dimensions affecting business processing and can be expanded or adjusted according to actual business needs. For example, the set of business status elements may include multiple dimensions such as overdue payment status, excessive traffic status, broadband optical attenuation alarm status, and 5G on / off status.

[0024] The system collects user business operation data and network log data, and matches them with various business states in a state set. The matching operation is performed according to preset feature identification rules, specifically checking whether each data field returned by the interface contains the feature identifier of each state in the state set. For a successfully matched business state dimension, the corresponding position in the state vector is assigned a value of 1; otherwise, the corresponding position is assigned a value of 0, generating a state vector. The state vector is a sparse binary vector, which can objectively record the user's real-time business situation. For example, if the system collects data showing a user account in arrears or a broadband optical attenuation alarm, and all other preset states fail to match, then the positions corresponding to arrears and optical attenuation alarms in the state vector are assigned a value of 1, and the other positions are assigned a value of 0.

[0025] To ensure the system doesn't recommend services that are currently impossible to process, such as recommending data packages to users whose accounts have been suspended, the system uses pre-defined hard logic rules in the knowledge graph for verification. Typically, the execution of each service intent requires certain preconditions to be met; intent recommendations detached from the service status are meaningless. Therefore, a pre-defined intent precondition rule library is established. This rule library is a pre-configured data resource library stored in the system, specifically storing the service execution precondition rules corresponding to each candidate intent. For example, the precondition rule for the intent to reactivate a mobile phone is that the user is in arrears and has not cancelled the account; the precondition rule for the intent to report broadband repair is that the user has broadband service and there is a network fault. In this embodiment of the invention, the state vector is substituted into the precondition rules of each candidate intent for Boolean logic operations. If the operation result is true, it is marked as valid in the intent validity vector; if the operation result is false, it is marked as invalid in the intent validity vector, thus obtaining the intent validity vector.

[0026] Boolean logic operations, based on AND, OR, and NOT logical relationships, are used to determine whether a user's current service status meets the preconditions for intent execution. Specifically, the component values ​​related to each intent's precondition rules are extracted from the status vector and substituted into the corresponding logical expressions to complete the calculation. For example, if the user's status vector shows outstanding fees and no account cancellation, and the result of the calculation using the precondition rule for mobile phone reactivation is true, then the mobile phone reactivation intent is marked as valid in the intent validity vector. If the user does not have broadband service, and the result of the calculation using the precondition rule for broadband repair is false, then the broadband repair intent is marked as invalid.

[0027] This completes the quantitative representation of user business status and the initial screening of candidate intents, providing more reasonable and accurate basic data for subsequent heterogeneous graph inference.

[0028] The heterogeneous graph reasoning module 102 is used to construct a heterogeneous graph model containing state nodes and intent nodes. In the heterogeneous graph model, based on the fault level and historical business processing time in the business state data, the activation weight of the state node on the intent node is analyzed in combination with the intent validity vector. Using the state vector as the initial probability distribution, the random walk algorithm is used to perform probability propagation and iterative convergence in the heterogeneous graph model to obtain the intent activation probability vector based only on the business background.

[0029] In real-world business scenarios, a user's intent is often implicitly related to their objective environment (such as network failures or account status) and is also affected by the cost of business processing (such as time consumption). Therefore, simple rule matching cannot quantify this complex push-pull effect.

[0030] In this embodiment of the invention, a graph model capable of representing the relationship between state and intent can be constructed to achieve probabilistic reasoning about potential user intents. A heterogeneous graph model is constructed, comprising two core types of nodes: state nodes and intent nodes. Nodes are connected by directed edges, with edge weights reflecting the strength of the connection. This serves as a vehicle to reflect the stimulating effect of business states on user intents and the potential relationships between intents. While constructing edges pointing from state nodes to intent nodes, the co-occurrence frequency of consecutive intents performed by the same user in a single session is statistically analyzed based on historical business processing logs. Directed edges pointing from intent nodes to intent nodes are then constructed in the heterogeneous graph model, with the co-occurrence frequency used as the edge weight.

[0031] Furthermore, by combining the objective characteristics of the business status, historical data of business processing, and the validity status of the intent, the correlation strength between the status node and the intent node is quantitatively analyzed, that is, the activation weight of the status node on the intent node is analyzed. In this embodiment of the invention, for each intent node currently marked as valid, for any directed edge from the status node to the valid intent node, that is, the directed edge is the correlation edge representing the activation effect of a certain business status on a certain valid intent, the weight of the edge directly reflects the probability of the business status triggering the user to generate the intent.

[0032] First, a fault gain factor is determined based on the fault level in the business state data corresponding to the directed edge state node. In this embodiment of the invention, according to a preset mapping relationship between fault level and gain factor, different fault severity levels are converted into corresponding numerical gain factors. The higher the fault level, the larger the corresponding gain factor value, reflecting the positive reinforcement effect of the fault urgency of the business state on the intent stimulation effect. The more severe the fault, the easier it is to trigger the relevant user intent. Specifically, the fault level is preset to a discrete value from 0 to 5, where 0 represents no fault and 5 represents a severe fault. The higher the fault level value, the higher the fault urgency. The product of the fault level and the fault weight adjustment coefficient is used as the fault gain factor. The fault weight adjustment coefficient is preset to 0.5. The higher the fault level, the larger the calculated fault gain factor value, thereby realizing the positive reinforcement of the fault level on the intent stimulation effect. That is, the more severe the fault, the stronger the stimulation effect of the corresponding business state on the relevant intent.

[0033] Secondly, the duration damping factor is determined based on the historical processing time corresponding to the directed edge intent node. In this embodiment of the invention, according to the preset mapping relationship between historical processing time and damping factor, different processing times are converted into corresponding numerical damping factors. The longer the processing time, the larger the corresponding damping factor value, reflecting the reverse inhibition effect of the execution cost of processing on the intent stimulation effect. The more cumbersome the processing, the lower the likelihood of the user choosing the intent. Specifically, after dimensionless processing of the historical processing time of the intent, the processed historical processing time is added to a constant 1 and logarithmic operation with the natural constant as the base is used to obtain the duration damping factor. The longer the historical average processing time of a certain intent, the larger the calculated duration damping factor value, thereby realizing the reverse inhibition of the intent stimulation effect of processing time. That is, in the cold start phase, in order to optimize the system service throughput and the first-question resolution rate, the duration damping factor is used to reduce the weight of high-time-consuming services and prioritize recommending highly automated and short-time-consuming service intents to users. It should be noted that the dimensionless processing eliminates the influence of dimensions. The specific means of eliminating the influence of dimensions are techniques well known to those skilled in the art, such as Z-score standardization, and are not limited here.

[0034] Finally, the activation weight of the directed edge is determined by combining the fault gain factor and the duration damping factor. The activation weight is positively correlated with the fault gain factor and negatively correlated with the duration damping factor, representing the activation intensity of the state on the intent after comprehensively considering the urgency of the business and the execution cost. In a specific embodiment of the present invention, the sum of constant 1 and the fault gain factor is used as the numerator, and the sum of constant 1 and the duration damping factor is used as the denominator to obtain the activation weight. The more severe the fault and the shorter the processing time, the higher the activation weight of the business state on the intent, and the greater the probability of triggering the user to generate the intent.

[0035] It should be noted that for each intent node currently marked as invalid, that is, the intent does not meet the execution preconditions in the user's current business state and has no actual recommendation meaning, the activation weight of the directed edge pointing to the invalid intent node is set to the blocking value. In this way, the invalid intent is logically shielded in the graph model to prevent the invalid intent from participating in subsequent probabilistic reasoning. The blocking value can be set to 0 to ensure that this type of edge has no probability propagation capability.

[0036] By activating weights to assign weights to edges in the heterogeneous graph model, the graph model can accurately represent the correlation strength between business states and valid intentions. Furthermore, starting from the user's current business state, a random walk algorithm is used to propagate probability within the heterogeneous graph model, allowing the probability flow to spread from state nodes to intention nodes, thus achieving probabilistic inference of intentions based on business context. The state vector needs to be normalized first to meet the mathematical requirements of the random walk algorithm for the initial probability distribution.

[0037] In this embodiment of the invention, the state vector is normalized to obtain the initial probability distribution of the random walk. Normalization involves scaling each component value in the state vector proportionally so that the sum of all component values ​​is 1, ensuring that it conforms to the basic characteristics of a probability distribution. In an extreme case, if all component elements are 0 during state vector normalization, a preset minimum positive number, such as e, is added to each dimension. -6 Perform smoothing.

[0038] In one specific embodiment of the present invention, the initial probability distribution is input into the heterogeneous graph model for probability propagation. Multiple rounds of iterative calculation are performed using a random walk algorithm with restart. During the iteration process, the restart probability is set to 0.15 to simulate the user's behavior of jumping back to the initial business state and reselecting, preventing excessive dissipation of the probability flow in the graph. After each iteration, a new full-graph node probability vector is obtained. The change in the L2 norm of the probability vectors between adjacent iterations is calculated to determine whether the probability distribution tends to stabilize.

[0039] The iteration stops when the change is less than a preset convergence threshold. The probability values ​​corresponding to the intent nodes in the heterogeneous graph model after the iteration converges are extracted and integrated to obtain the intent activation probability vector. This vector reflects the probability of each valid intent being triggered based solely on the user's current business context. In this embodiment, the convergence threshold can be preset according to the algorithm's accuracy requirements and computational efficiency; it is a very small value, such as 1×10⁻⁶. -6 Extract the probability value corresponding to each intent node to form a vector with the same dimension as the number of valid intents, which can represent the potential priority of each valid intent in the current business context.

[0040] At this point, the intent reasoning based on the business context has been completed, and background probability features with noise resistance have been obtained.

[0041] The dual-stream fusion decision module 103 is used to perform semantic recognition on the text data input by the user to obtain a semantic classification probability vector; determine the text confidence factor based on the effective information content in the text data; fuse the semantic classification probability vector and the intent activation probability vector based on the text confidence factor to obtain an intent comprehensive evaluation index; and respond to the grading strategy according to the intent comprehensive evaluation index.

[0042] Furthermore, after obtaining the intent activation probability vector based on the business context, it is also necessary to combine the user's explicit text input to achieve a fusion analysis of subjective input and objective background. This is because relying solely on business background reasoning cannot reflect the user's explicit demands, and relying solely on text semantic analysis cannot cope with the situation of ambiguous user input. Therefore, a fusion decision is made from two dimensions: text semantic recognition and background probability reasoning, to achieve accurate intent recognition in all scenarios.

[0043] In this embodiment of the invention, the user-input text data is segmented and feature extracted to obtain a text feature vector. Segmentation involves dividing continuous text content into independent lexical units, and feature extraction involves converting these lexical units into numerical vectors that represent the semantics of the text. The text feature vector is input into a pre-trained natural language processing model. Through the model's feature matching and probability calculation functions, the matching degree between the text feature vector and each standard intent in the candidate intent set is output. The matching degrees are then integrated to obtain a semantic classification probability vector. It should be noted that the segmentation can use the Jieba segmentation algorithm to divide continuous text content into independent lexical units, and the feature extraction can use the Word2Vec model to convert the lexical units into fixed-dimensional numerical text feature vectors. The BERT natural language processing model is used for matching analysis. These processing methods are well-known techniques to those skilled in the art and will not be elaborated upon here.

[0044] Considering that user-input text data may contain semantic ambiguity and missing information, and that text with different amounts of information has different reference value for intent decision-making, in this embodiment of the invention, invalid information filtering is performed on the user-input text data to extract the effective content in the text and count the amount of effective information. Invalid information filtering involves removing stop words, punctuation marks, emoticons, and other content in the text that has no actual semantic meaning. In one specific embodiment of the invention, the amount of effective information can be quantified by counting the length of effective characters after removing invalid information. The longer the length of effective characters, the higher the amount of effective information in the text.

[0045] To dynamically adjust the weight of text semantics and business context in intent decision-making, allowing the system to adaptively switch decision criteria based on the validity of text input, a mapping rule between information content and confidence level is pre-defined. A one-to-one correspondence is established between different effective information content ranges and corresponding confidence factor values; the higher the information content, the larger the corresponding confidence factor value. For example, the pre-defined mapping rule between information content and confidence level is as follows: if the effective character length is less than 2, it is determined to be low-information-content text, and the corresponding text confidence factor is 0.2; if the effective character length is greater than or equal to 2, it is determined to be high-information-content text, and the corresponding text confidence factor is 0.7.

[0046] Furthermore, based on the preset mapping rules between information content and confidence level, the statistical effective information content can be substituted into the mapping rules to match the corresponding text confidence level factor, which is the coefficient value used to adjust the weight ratio of text semantics and business background. It reflects the reference weight of user text input in intent decision-making. The higher the effective information content, the larger the text confidence level factor and the higher the reference weight of text semantics.

[0047] After obtaining the semantic classification probability vector, intent activation probability vector, and text confidence factor, the inference results of text semantics and business context can be weighted and fused to obtain an evaluation index that comprehensively reflects the user's intent. In this embodiment of the invention, the semantic classification probability vector is weighted using the text confidence factor, and the intent activation probability vector is weighted using the difference between a constant 1 and the text confidence factor. The weighted sum is used as the comprehensive intent evaluation index, which is to sum the corresponding components of the two probability vectors one by one after weighting them respectively. The resulting vector, composed of the comprehensive intent scores in order, reflects the comprehensive matching degree of each candidate intent combined with text semantics and business context, and characterizes the probability that each candidate intent is the user's true intent.

[0048] By obtaining the comprehensive intent evaluation index, the system can comprehensively consider the user's explicit textual requests and objective business context, enabling the system to execute differentiated intent response strategies based on the degree of comprehensive matching, thereby improving the accuracy and efficiency of intent services. In this embodiment of the invention, the comprehensive intent evaluation index is sorted in descending order, and the difference between the highest and second-highest comprehensive intent evaluation index is calculated. That is, the comprehensive intent evaluation indices of each intent are sorted from high to low, the first and second ranked comprehensive intent evaluation indices are extracted, and the numerical difference between the two is calculated. This difference reflects the degree of distinction between the highest-rated intent and other intents.

[0049] If the highest intent comprehensive evaluation index exceeds the execution threshold and the difference exceeds the discrimination threshold, it indicates that the user's true intent is clear and unique, without ambiguity. The system can directly execute the corresponding business operation, thus directly executing the intent business corresponding to the highest intent comprehensive evaluation index without needing to clarify with the user, thereby improving service efficiency. In this embodiment of the invention, the execution threshold can be set to 0.85, and the discrimination threshold can be set to 0.15. These are preset values ​​based on business needs, and implementers can adjust them according to the actual intent recognition accuracy, without any restrictions.

[0050] If the highest comprehensive intent evaluation index exceeds the clarification threshold but does not meet the direct execution conditions, it indicates that the user's intent has a certain tendency but is not unique and is ambiguous. It is necessary to confirm the user's true intent. Therefore, clarification options containing a preset number of intents corresponding to the highest comprehensive intent evaluation indices are generated, displayed to the user, and the user is guided to choose, improving the accuracy of intent recognition. In this embodiment of the invention, the clarification threshold can be set to 0.4. If the threshold is lower than 0.4, it indicates that the user's intent cannot be recognized. The preset number can be set to 3 according to business needs. That is, after the comprehensive intent evaluation indices are sorted in descending order, the intents corresponding to the top three comprehensive intent evaluation indices are displayed as cards for the user to choose from.

[0051] Otherwise, if the highest intent comprehensive evaluation index is lower than the clarification threshold, it means that the system cannot effectively identify the user's true intent. In this case, the system will execute the strategy of transferring the user's conversation to a human customer service representative for further service processing to ensure that the user's needs are effectively responded to.

[0052] The parameter update module 104 is used to monitor and record the actual time spent on the user's recommended intent after the business is completed; and to update the historical business processing time parameter of the corresponding intent in the heterogeneous graph model by using the exponential weighted moving average algorithm, combining the historical business processing time with the actual processing time of this time.

[0053] Ultimately, considering that the processing time of business transactions can change with factors such as business process optimization, system upgrades, and network status, in order to ensure that the duration damping factor in the heterogeneous graph model can reflect the current business execution cost in real time and ensure the accuracy and timeliness of intent reasoning, in this embodiment of the invention, a parameter update module is used to dynamically update the historical business processing time parameters. By collecting the time data of successfully processed transactions in real time, the resistance parameters in the graph model are smoothly updated using the Exponential Weighted Moving Average (EWMA) algorithm. This enables the system to learn in a closed loop and dynamically evolve parameters in response to changes in the business environment, achieving adaptive perception of changes in the business environment and ensuring that the weight calculation of the heterogeneous graph model always fits the actual business situation.

[0054] In summary, this invention improves intent recognition capabilities under fuzzy semantics by acquiring user business status data and generating state vectors and intent validity vectors. This facilitates the analysis and utilization of implicit correlations between business status (such as faults or arrears) and intents, constructing a heterogeneous graph model containing state nodes and intent nodes. It then analyzes activation weights based on fault levels and historical business processing times, and obtains intent activation probability vectors based on business context through a random walk algorithm. By incorporating historical processing times and fault level analysis, it prioritizes short-duration, highly automated services during the cold start phase, reducing user operation barriers and the reliance on human customer service. Finally, it dynamically adjusts the confidence factor based on the effective information content of the text, achieving intelligent fusion of text semantics and business context reasoning results to adapt to different user input scenarios. A tiered strategy is implemented based on the fused intent comprehensive evaluation index, making intent responses more aligned with actual needs and further improving the system's service adaptability and user experience. This invention constructs a state-intent heterogeneous graph and introduces a dynamic fusion mechanism to transform common business background knowledge into computable probabilistic features, effectively reducing the failure problem of NLP technology in semantically ambiguous and multi-state coupled scenarios, and improving the accuracy of intent recognition and service efficiency in complex business scenarios.

[0055] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0056] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A user intent reasoning system based on knowledge graphs, characterized in that, The system includes: The state awareness module is used to acquire the user's current business state data and network environment data, and obtain the user's business state state vector through binary mapping; based on the preset intent pre-rule library, it performs validity verification on the candidate intent set and generates an intent validity vector. The heterogeneous graph reasoning module is used to construct a heterogeneous graph model containing state nodes and intent nodes. In the heterogeneous graph model, based on the fault level and historical business processing time in the business state data, the activation weight of state nodes on intent nodes is analyzed in combination with the intent validity vector. Using the state vector as the initial probability distribution, the random walk algorithm is used to perform probability propagation and iterative convergence in the heterogeneous graph model to obtain the intent activation probability vector based only on the business background. The dual-stream fusion decision module is used to perform semantic recognition on the text data input by the user to obtain a semantic classification probability vector; determine the text confidence factor based on the effective information content in the text data; fuse the semantic classification probability vector and the intent activation probability vector based on the text confidence factor to obtain an intent comprehensive evaluation index; and respond to the graded strategy according to the intent comprehensive evaluation index.

2. The user intent reasoning system based on knowledge graphs according to claim 1, characterized in that, The method for obtaining the state vector includes: A set containing business status elements with two or more dimensions is pre-constructed; user business operation data and network log data are collected and matched with each dimension of business status in the state set. For a successfully matched business status dimension, the corresponding position in the state vector is assigned a value of 1; otherwise, the corresponding position in the state vector is assigned a value of zero, thus generating a state vector.

3. The user intent reasoning system based on knowledge graphs according to claim 1, characterized in that, The method for obtaining the intent validity vector includes: The pre-defined intent pre-rule library stores the business execution pre-rules corresponding to each candidate intent. The state vector is substituted into the pre-rules of each candidate intent and Boolean logic operation is performed. If the operation result is true, it is marked as a valid identifier in the intent validity vector; if the operation result is false, it is marked as an invalid identifier in the intent validity vector, thus obtaining the intent validity vector.

4. The user intent reasoning system based on knowledge graphs according to claim 3, characterized in that, The method for obtaining the activation weight includes: For each intent node currently marked as a valid identifier, for any directed edge from the state node to the valid identifier intent node, determine the fault gain factor based on the fault level corresponding to the state node of the directed edge; determine the duration damping factor based on the historical business processing duration corresponding to the intent node of the directed edge. The excitation weight of the directed edge is determined by combining the fault gain factor and the duration damping factor. The excitation weight is positively correlated with the fault gain factor and negatively correlated with the duration damping factor. For each intent node currently marked as invalid, set the activation weight of the directed edge pointing to the invalid intent node to the blocking value.

5. The user intent reasoning system based on knowledge graphs according to claim 1, characterized in that, The method for obtaining the intent activation probability vector includes: The initial probability distribution of the random walk is obtained by normalizing the state vector; the initial probability distribution is then input into the heterogeneous graph model for probability propagation, and the change in the probability vector is calculated after each round of propagation. When the change is less than the preset convergence threshold, the iteration stops. The probability values ​​corresponding to the intention nodes in the heterogeneous graph model after the iteration stops convergence are extracted and integrated to obtain the intention activation probability vector.

6. The user intent reasoning system based on knowledge graphs according to claim 1, characterized in that, The method for obtaining the semantic classification probability vector includes: The text data input by the user is segmented and features are extracted to obtain a text feature vector; the text feature vector is input into a pre-trained natural language processing model, and the matching degree between the text feature vector and each standard intent in the candidate intent set is output. The matching degrees are integrated to obtain a semantic classification probability vector.

7. The user intent reasoning system based on knowledge graphs according to claim 1, characterized in that, The method for obtaining the text confidence factor includes: The system filters out invalid information from the user-input text data, extracts valid content from the text, and counts the amount of valid information. Based on a preset mapping rule between information amount and confidence level, the system substitutes the counted amount of valid information into the mapping rule to match and obtain the corresponding text confidence level factor.

8. The user intent reasoning system based on knowledge graphs according to claim 1, characterized in that, The methods for obtaining the comprehensive intent evaluation indicators include: The semantic classification probability vector is weighted using the text confidence factor; the intent activation probability vector is weighted using the difference between the constant 1 and the text confidence factor; and the weighted sum is used as the comprehensive intent evaluation index.

9. The user intent reasoning system based on knowledge graph according to claim 1, characterized in that, The response grading strategy based on the comprehensive evaluation indicators of intent includes: Sort the comprehensive evaluation indicators of intent in descending order; calculate the difference between the highest and second highest comprehensive evaluation indicators of intent. If the highest intent comprehensive evaluation index exceeds the execution threshold and the difference exceeds the discrimination threshold, the intent business corresponding to the highest intent comprehensive evaluation index will be executed directly. If the highest intent comprehensive evaluation index exceeds the clarification threshold but does not meet the direct execution conditions, then a clarification option containing a preset number of intents corresponding to the highest intent comprehensive evaluation index will be generated. Otherwise, the strategy of transferring to human assistance will be implemented.

10. The user intent reasoning system based on knowledge graphs according to claim 1, characterized in that, It also includes a parameter update module, which is used to: monitor and record the actual time spent on the user's recommended intent after the business is completed; and update the historical business processing time parameter of the corresponding intent in the heterogeneous graph model by using an exponentially weighted moving average algorithm, combining the historical business processing time with the actual processing time of this time.

Images