Intelligent communication data retrieval method and platform based on semantic analysis

Abstract

The invention discloses a communication data intelligent retrieval method and platform based on semantic analysis, and relates to the technical field of natural language processing, and the method comprises the following steps: after receiving an original query text of a user, binding an association retrieval permission strategy based on a user ID, loading a user portrait to complement semantics to obtain a plurality of semantic equivalence complement statements, and storing the semantic equivalence complement statements in a database; the method comprises the following steps: extracting feature words of communication data, constructing a plurality of hierarchical retrieval feature trees in combination with a domain knowledge graph, then matching a distributed retrieval unit, carrying out dynamic superposition retrieval by taking a retrieval permission strategy as a constraint to obtain a plurality of refined result sets, carrying out descending order arrangement and integration on the refined result sets according to metadata timestamps, and outputting a communication data time series retrieval sequence. According to the method and the device, the technical problems of low retrieval efficiency and difficulty in accurately positioning the communication data caused by single semantic understanding in traditional communication data retrieval are solved, and the technical effects of accurate semantic understanding of the communication data, orderly and efficient retrieval and accurate positioning of the communication data meeting the intention of the user are achieved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and platform for intelligent retrieval of communication data based on semantic parsing. Background Technology

[0002] In real-world communication scenarios, efficient retrieval of massive amounts of communication data is crucial for business analysis, troubleshooting, and responding to user needs. Accurately acquiring target data is a core prerequisite for achieving these applications. Current technologies for communication data retrieval largely rely on traditional methods such as keyword matching, which play a role in structured data retrieval scenarios. However, with the increasing scale and semantic complexity of communication data, traditional retrieval technologies reveal limitations. They are prone to retrieval bias due to limited semantic understanding, making it difficult to accurately locate target data. The resulting results are disordered and redundant, failing to meet the demands for accurate retrieval and efficient application of communication data. Summary of the Invention

[0003] This application provides a semantic parsing-based intelligent retrieval method and platform for communication data, which solves the technical problems of low retrieval efficiency and difficulty in accurately locating communication data due to the single semantic understanding in traditional communication data retrieval.

[0004] The first aspect of this application provides a semantic parsing-based intelligent retrieval method for communication data. The method includes: upon receiving an original query text submitted by a user, triggering a user ID-based association retrieval permission strategy binding; loading a user profile based on the user ID to perform semantic completion of the original query text, obtaining multiple semantically equivalent completion statements; extracting feature words from the multiple semantically equivalent completion statements to obtain multiple sets of retrieval feature words; performing semantic association divergence of the multiple sets of retrieval feature words based on a domain knowledge graph to construct multiple hierarchical retrieval feature trees; performing hierarchical narrowing retrieval task matching based on the node weight features of the multiple hierarchical retrieval feature trees to locate multiple distributed retrieval units; using the retrieval permission strategy as a constraint, driving the multiple distributed retrieval units to perform top-down conditional dynamic superposition retrieval of the multiple hierarchical retrieval feature trees, outputting multiple refined result sets; retrieving the metadata timestamps of the multiple refined result sets, arranging and integrating them in descending order, and outputting a time-series retrieval sequence of communication data.

[0005] A second aspect of this application provides a semantic parsing-based intelligent retrieval platform for communication data. The system includes: an associated retrieval permission policy execution module, used to trigger an associated retrieval permission policy binding based on the user ID after receiving the original query text submitted by the user; a semantic equivalence completion statement acquisition module, used to load a user profile based on the user ID to perform semantic completion of the original query text, obtaining multiple semantic equivalence completion statements; a retrieval feature word acquisition module, used to extract feature words from the multiple semantic equivalence completion statements to obtain multiple sets of retrieval feature words; and a hierarchical retrieval feature tree construction module, based on a domain knowledge graph... The semantic associations of multiple sets of search feature words are diverged to construct multiple hierarchical search feature trees; a distributed search unit positioning module is used to perform hierarchical narrowing search task matching based on the node weight features of the multiple hierarchical search feature trees to locate multiple distributed search units; a refined result set acquisition module is used to drive the multiple distributed search units to perform top-down conditional dynamic superposition search of the multiple hierarchical search feature trees under the constraint of the search permission policy, and output multiple refined result sets; a time-series search sequence acquisition module is used to retrieve the metadata timestamps of the multiple refined result sets, sort them in descending order and integrate them, and output a time-series search sequence of communication data.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application, upon receiving the original query text submitted by the user, triggers a user ID-based associated retrieval permission strategy binding. Based on the user ID, it loads a user profile to complete the semantics of the original query text, obtaining multiple equivalent statements. After extracting retrieval feature words, it constructs a hierarchical retrieval feature tree using a domain knowledge graph. It then matches distributed retrieval units and executes a top-down, dynamically superimposed retrieval based on permission policies. Finally, it retrieves and integrates the result metadata (timestamps) in descending order, thereby accurately obtaining communication data that matches the user's intent. This makes the communication data retrieval results more accurate, secure, and user-friendly, achieving the technical effects of precise semantic understanding of communication data, orderly and efficient retrieval, and accurate location of communication data that matches the user's intent. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0008] Figure 1 This is a flowchart illustrating the intelligent retrieval method for communication data based on semantic parsing provided in an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of the structure of the intelligent communication data retrieval platform based on semantic parsing provided in the embodiments of this application.

[0010] Figure labeling: 1. Module for executing the associated retrieval permission strategy; 2. Module for obtaining semantic equivalence completion statements; 3. Module for obtaining retrieval feature words; 4. Module for constructing hierarchical retrieval feature trees; 5. Module for locating distributed retrieval units; 6. Module for obtaining refined result sets; 7. Module for obtaining time-series retrieval sequences. Detailed Implementation

[0011] This application provides a semantic parsing-based intelligent retrieval method and platform for communication data, which solves the technical problems of low retrieval efficiency and difficulty in accurately locating communication data due to the single semantic understanding in traditional communication data retrieval.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, a semantic parsing-based intelligent retrieval method for communication data includes: Step A100: After receiving the original query text submitted by the user, trigger the binding of the associated retrieval permission policy based on the user ID.

[0015] In this embodiment of the application, the original query text is the search-related text submitted by the user, which serves as the basic input for subsequent operations such as semantic completion and feature word extraction.

[0016] Specifically, after receiving the user's original query text, real-time request attributes and the user ID are extracted from the request context, operation risk tags are constructed by associating with the security audit library, the initial retrieval permissions are matched according to the user's identity role, and the risk tags are degraded and corrected. Then, real-time risk control adaptation is performed according to the operation environment status, and finally, an associated retrieval permission policy is output to implement the binding of the associated retrieval permission policy based on the user ID. The specific steps are described in detail in A110 - A140.

[0017] Step A200: Load the user portrait according to the user ID to perform semantic completion of the original query text, and obtain multiple semantically equivalent completion statements.

[0018] In the embodiment of the present application, the user portrait is the real-time user portrait features retrieved from the user portrait database through the user ID, specifically including historical retrieval features, business preference features, and semantic habit features.

[0019] Optionally, the real-time user portrait features are obtained from the portrait database with the user ID as the query key, and are spliced with the original query text into a structured input. After multi-branch semantic completion by the semantic completion model, multiple semantically equivalent completion statements corresponding to multiple personalized retrieval intention vectors are generated, thereby completing the semantic completion of the original query text based on the user portrait. The specific steps are described in detail in A210 - A220.

[0020] Step A300: Extract feature words of the multiple semantically equivalent completion statements to obtain multiple groups of retrieval feature words.

[0021] In an embodiment of the present application, first, text preprocessing is performed on the obtained multiple semantically equivalent completion statements to filter out stop words and punctuation: first, load the preset stop word list, which contains words without actual retrieval meaning such as "the", "and", "already", etc., traverse each word in the statement sentence by sentence, delete the matched stop words, and at the same time remove punctuation marks such as commas and periods to obtain a clean statement text without redundant interference.

[0022] Then, feature word extraction is performed. The TF-IDF word frequency weighting method is used to extract keywords, calculate the word frequency (TF) of each word in each clean statement and calculate the inverse document frequency (IDF) of the word. Multiply TF and IDF to obtain the word weight. This process is the same as that in step A224 - 4 and will not be elaborated here. Then, sort in descending order according to the weight, and take the top 3 - 5 words as the retrieval feature words of the statement. Among them, the retrieval feature words refer to the words that can accurately represent the core retrieval intention of the statement. For example, "last 3 days" extracted from retrieving the base station communication data in the last 3 days is a time - type retrieval feature word that limits the retrieval time range, "base station" is an object - type retrieval feature word that limits the object to which the retrieval data belongs, and "communication data" is a type - type retrieval feature word that limits the category of the retrieval data. They are all keyword vocabularies that reflect the user's retrieval requirements.

[0023] Finally, the search feature words are grouped, and the search feature words extracted for each semantic equivalence completion statement are grouped together. For example, the first semantic equivalence completion statement's "last 3 days", "base station", and "communication data" are grouped into the first group, and the second semantic equivalence completion statement's "last 3 days", "core base station", and "communication records" are grouped into the second group, and so on, to obtain multiple groups of search feature words.

[0024] By preprocessing text to remove interference, extracting core feature words using TF-IDF, and grouping and organizing by sentence, the system achieves the effect of accurately extracting multiple sets of search feature words that reflect the core search needs of users from multiple semantically equivalent completion sentences.

[0025] Step A400: Based on the domain knowledge graph, perform semantic association divergence of the multiple sets of search feature words to construct multiple hierarchical search feature trees.

[0026] Specifically, the first set of search feature words is linked with the nodes of the domain knowledge graph to construct the original semantic layer search feature. Using this feature as the root node, the semantic association divergence of the preset extension level is performed to obtain the first level tree node set. Based on this node set, the first initial level tree is constructed. The first historical recurrence frequency set of the node set is called locally to calculate the initial weight set of the first feature word. After the weight set is mapped and loaded into the initial tree, the weight threshold pruning is performed according to the preset level decay rule. Finally, the first level search feature tree is constructed. This achieves the purpose of constructing multiple level search feature trees by semantic association divergence of multiple sets of search feature words based on the domain knowledge graph. The specific steps are explained in detail in A410-A450.

[0027] Step A500: Perform hierarchical narrowing retrieval task matching based on the node weight features of the multiple hierarchical retrieval feature trees to locate multiple distributed retrieval units.

[0028] Specifically, firstly, the weight set of the first node of the first-level retrieval feature tree is extracted. Based on this weight set, the depth of the first tree and the entropy of the first weight distribution are calculated. After weighted normalization, the first retrieval difficulty coefficient is output and the task is classified accordingly to obtain the first retrieval task level. Then, the first distributed retrieval unit is matched and located according to the retrieval task level. This realizes the hierarchical narrowing retrieval task matching based on the node weight features of multiple levels of retrieval feature trees to locate multiple distributed retrieval units. The specific steps are explained in detail in A510-A540.

[0029] Step A600: Using the retrieval permission policy as a constraint, drive the multiple distributed retrieval units to perform top-down conditional dynamic superposition retrieval of the multiple hierarchical retrieval feature trees, and output multiple refined result sets.

[0030] Specifically, constrained by the retrieval permission policy, the first distributed retrieval unit initiates an initial query based on the top-level node feature words of the first-level retrieval feature tree, outputting the first-level retrieval results. Then, the second and third-level node feature words are dynamically added sequentially to perform secondary and tertiary queries. This process is repeated traversing the retrieval feature tree to iterate and superimpose feature word conditions until the first global retrieval result set is output. Finally, the retrieval permission policy is used as a result desensitization constraint to remove unauthorized fields and output the first refined result set. This completes the retrieval of multiple levels of retrieval feature trees by multiple distributed retrieval units. The specific steps are explained in detail in A610-A650.

[0031] Step A700: Retrieve the metadata timestamps of the multiple refined result sets, sort and integrate them in descending order, and output the time-series retrieval sequence of communication data.

[0032] In this embodiment, the metadata timestamp corresponds to the generation or storage time of the communication data and is the core basis for realizing the chronological arrangement.

[0033] Specifically, after multiple distributed retrieval units perform top-down dynamic condition overlay retrieval of multiple hierarchical retrieval feature trees and output multiple refined result sets that conform to the retrieval permission policy, the metadata timestamp associated with each refined result set is retrieved. Then, according to the descending order of time from newest to oldest, all refined result sets with timestamps are sorted sequentially. Finally, the sorted result sets are integrated as a whole to ensure that the communication data retrieval results in different refined result sets are connected in a unified time logic. Finally, a well-structured and clearly ordered time sequence of communication data retrieval is output, making it convenient for users to intuitively obtain the latest communication data retrieval results.

[0034] Furthermore, step A100 in the method provided in this application embodiment includes: A110: Extract the real-time request attributes and the user ID from the request context of the original query text, wherein the real-time request attributes include user identity role and operating environment status.

[0035] A120: Retrieve historical unauthorized operation records from the security audit database associated with the user ID, and construct an operation risk label.

[0036] A130: After matching the initial search permissions according to the user's identity and role, dynamically downgrade the initial search permissions based on the operation risk label, and output the corrected search permissions.

[0037] A140: Perform real-time risk control to adapt the modified search permissions according to the operating environment status, and output the associated search permission strategy.

[0038] In this embodiment of the application, the security audit database is an encrypted database that stores historical retrieval operation records.

[0039] Specifically, after receiving the original query text submitted by the user, the system first parses the request context of the original query text. The request context contains various accompanying information when the user initiates the search operation. For example, the data packet of the search request's transmission protocol carries a user ID used to identify the user's identity. It also contains identity-related information and environmental information when the user initiates the operation. Based on this, the user's identity role can be identified, such as communication network maintenance personnel or ordinary individual users, as well as the operating environment status. The operating environment status specifically includes the login IP geofence and device security authentication level. For example, the system will determine whether the user's login IP is within the enterprise's preset secure geographical area and whether the search device used by the user has passed security authentication processes such as two-factor authentication. By extracting this information, basic data support is provided for the construction of subsequent permission policies.

[0040] Next, after extracting the user ID, it is associated with the security audit database. The database is then used to retrieve the user's past search operation records, with a focus on filtering for historical unauthorized access records. These records represent instances where the user exceeded their valid search permissions during past communication data retrieval operations. The security audit database stores and updates all users' search behavior data in real time, including the scope of each user's access permissions, the actual content retrieved, and whether there were any attempts to access or obtain communication data beyond their authorized scope. For example, if a user has attempted to retrieve communication logs from other departments outside their responsibilities three times in the past six months, a corresponding operational risk label is created for that user based on these historical unauthorized access records. Common label types include high risk, medium risk, and low risk.

[0041] Furthermore, the classification rules for the above three risk labels are mainly based on a comprehensive judgment of the frequency of unauthorized operations, the severity of unauthorized behavior, and potential impact within a preset time range. The low-risk label applies to users who have performed fewer than two unauthorized operations in the past six months, and whose unauthorized attempts did not involve highly sensitive communication data, such as core network topology data or user privacy data, and did not pose a substantial threat to data security. The medium-risk label applies to users who have performed two to three unauthorized operations in the past six months, or whose unauthorized behavior involves moderately sensitive data, such as departmental communication log summaries. While these operations did not result in data leakage, they demonstrate a clear intent to breach authorization boundaries. The high-risk label applies to users who have performed more than three unauthorized operations in the past six months, or whose unauthorized behavior involves highly sensitive data, such as user call content or encrypted communication protocol data, or whose unauthorized attempts have partially succeeded in obtaining data, such as more than three attempts to retrieve core communication failure records from other departments, or one successful unauthorized acquisition of sensitive data. If a user meets the criteria for different risk levels simultaneously, the higher risk level will be prioritized.

[0042] After completing the construction of operational risk tags, corresponding initial search permissions are matched based on the extracted user roles. Specifically, based on the actual business needs of communication data management and the preset security specifications, different user roles correspond to different initial search permissions. For example, the initial search permission for communication network maintenance personnel is to view the real-time communication data and historical logs of all base stations in their area of ​​responsibility, while the initial search permission for ordinary individual users is only to view their own call records, SMS records, and other personal communication data.

[0043] Subsequently, based on the established low, medium, and high operational risk labels, the initial search permissions are dynamically downgraded in a differentiated manner: for users with low operational risk labels, only slight restrictions are imposed on the initial search permissions, usually by slightly narrowing the search scope of non-core communication data, or making limited adjustments to marginal permissions such as the duration of historical data backtracking, in order to minimize the impact on the user's normal search needs. For example, for ordinary users whose initial permissions are to view their own call records, SMS records, and other personal communication data for the past 3 months, the system will slightly downgrade their permissions, adjusting them to only allow them to view their own call records and SMS records for the past 1 month, reducing the scope of access to unnecessary data.

[0044] Then, for users with a medium-risk operation risk label, their initial search permissions will be moderately reduced. Specifically, this means narrowing the scope of core communication data searches or restricting key search functions such as data export and detailed viewing. While ensuring basic usage, risk control is strengthened. For example, if an operations and maintenance personnel initially have the permission to view the basic communication logs of 5 base stations in their area of ​​responsibility, the system will moderately downgrade their permissions, allowing them to view only the basic communication logs of 3 of the main base stations they are responsible for, thus reducing the scale of data access. For users with a high-risk operation risk label, their initial search permissions will be strictly restricted, significantly reducing the scope of core communication data searches or disabling important search permissions such as real-time data viewing and access to sensitive fields. By significantly reducing the permission level, the system aims to minimize the risk of unauthorized operations. For example, if an operations and maintenance personnel initially have the permission to view the data of all base stations in their area of ​​responsibility, the system will significantly downgrade their permissions, allowing them to view only the communication data of the 3 specific base stations they are directly responsible for, strictly limiting data access boundaries. Finally, the system will output corrected search permissions corresponding to the risk level.

[0045] Finally, based on the operational environment status extracted in the above steps, real-time risk control adaptation is performed on the modified search permissions. For example, if a user's modified search permission is to view communication data from three specific base stations, the system will check their operational environment status. If it finds that the user's login IP is not within the preset security geofence (i.e., the user usually logs into the system from the company's intranet but is logging in from a public network elsewhere), or that the user's device has a low security authentication level (e.g., the device does not have two-factor authentication enabled and unpatched security vulnerabilities are detected), the modified search permissions will be further adjusted. For example, the user may be restricted to viewing only the communication data from these three base stations within the last 24 hours, rather than all historical data. After this series of adaptations and adjustments, the final associated search permission policy is output, clearly defining the scope of communication data, data types, and operational restrictions that the user can access in this search operation.

[0046] By extracting multi-dimensional information from the request context, constructing risk tags by associating with the security audit library, dynamically adjusting permissions based on identity and risk, and performing real-time risk control adaptation according to the operating environment, a coherent process is achieved to accurately bind associated search permission strategies for users. This effectively prevents unauthorized operations and ensures the security of communication data while guaranteeing the normal search needs of legitimate users.

[0047] Furthermore, step A200 in the method provided in this application embodiment includes: A210: Using the user ID as the query key, query the user profile database to retrieve real-time user profile features.

[0048] A220: The original query text and real-time user profile features are concatenated into a structured input, and multi-branch semantic completion is performed through a semantic completion model to generate multiple semantically equivalent completion statements corresponding to multiple personalized search intent vectors.

[0049] Optionally, during the process of loading a user profile based on the user ID, after receiving the original query text submitted by the user, the user ID is used as the core query key to access the pre-built user profile database. This database is a structured data set used to store the user's personalized characteristics. Internally, a unique index is established through the user ID, and dynamic data related to the user's search behavior, business attributes, and semantic habits are stored in association.

[0050] Next, based on the uniqueness of the user ID, a precise match is performed in the index system of the database to quickly locate the profile data entry corresponding to the user, and then the real-time user profile features are retrieved and extracted from it. The real-time user profile features are not static data, but a multi-dimensional feature set that is dynamically updated with user behavior, specifically including historical retrieval features, business preference features and semantic habit features, which are explained in detail in step A211.

[0051] Next, the real-time user profile features are converted into multi-dimensional dynamic feature vectors with dynamic detached attributes. Word segments are extracted from the original query text and intent anchors are inserted to generate a query intent sequence. Then, based on this sequence, the multi-dimensional dynamic feature vectors are updated by random perturbation of vector weights based on probability distribution and concatenated to construct multiple personalized search intent vectors. Finally, the original query text and these personalized search intent vectors are used as structured inputs and multi-branch semantic completion is performed by the semantic completion model to generate multiple semantically equivalent completed sentences. The specific steps are explained in detail in A221-A224.

[0052] Furthermore, step A210 in the method provided in this application embodiment includes: A211: The real-time user profile features include historical retrieval features, business preference features, and semantic habit features.

[0053] Specifically, historical search features are user search behavior data recorded in the user profile database. These features include high-frequency search keywords, which are core words that users repeatedly use in multiple searches; commonly used filter condition combinations, which are the conditions that users usually use when performing searches, such as setting the time range to the last 7 days and the data type to real-time communication traffic; and repeated query patterns, which are the periodic or repetitive search behavior patterns of users, such as regularly searching for the department's voice call statistics from the previous week every Monday.

[0054] Business preference characteristics reflect users' search tendencies based on their own business needs or job responsibilities, including frequently accessed communication data types, data categories that users prioritize when searching, such as voice call records, SMS content, or corporate email communication data; and key contacts / departments that users focus on, i.e. specific objects that users frequently associate with when searching, such as a customer service representative who often focuses on the communication records of the complaint handling department or key customers.

[0055] Semantic habit features reflect users' personalized language habits in their search expressions. The core is a personalized terminology list, which consists of the unique expressions users habitually use when searching. For example, most users would use "call" to refer to VoIP voice communication at the technical level, and "message" to refer to SMS service data. Through this search and extraction process, real-time user profile features that comprehensively and accurately depict users' search needs and preferences are obtained. This provides a specific and personalized data foundation for subsequent steps, ensuring that subsequent semantic completion operations accurately match users' potential search intentions.

[0056] Furthermore, step A220 in the method provided in this application embodiment includes: A221: The real-time user profile features are converted into multi-dimensional dynamic feature vectors, wherein the multi-dimensional dynamic feature vectors have dynamic detached properties.

[0057] A222: After extracting words from the original query text, intent anchor characters are inserted to generate a query intent sequence.

[0058] A223: Based on the query intent sequence, after updating the multi-dimensional dynamic feature vector by random perturbation of the vector weights based on probability distribution, the multiple personalized search intent vectors are constructed by concatenating the vectors.

[0059] A224: The original query text and multiple personalized search intent vectors are used as structured inputs, and multi-branch semantic completion is performed through the semantic completion model to generate the multiple semantically equivalent completed statements.

[0060] Specifically, the real-time user profile features are first transformed into vectors using the features acquired in the previous steps. These features include three core dimensions: historical search features, business preference features, and semantic habit features. These feature data are then standardized and converted into multi-dimensional dynamic feature vectors using word embedding and feature embedding techniques. Each dimension of the vector corresponds one-to-one with the category of the real-time user profile feature, and the value of each dimension represents the importance of the corresponding feature. The dynamic detached attribute in this vector means that the weights of each dimension are not fixed but dynamically adjusted according to changes in the user's recent search behavior and the current search scenario, ensuring that the vector reflects the user's search needs and preferences in real time. The specific process is as follows: Step a: When standardizing real-time user profile features, adopt corresponding standardization methods for different types of feature data to ensure that the feature data format is uniform and the numerical range is suitable for subsequent embedding processes. For numerical features in historical retrieval features, such as retrieval frequency and usage count, the Min-Max normalization method is adopted: First, the maximum and minimum values ​​of the numerical feature in the user profile database are obtained. After normalization, the value is normalized by formula = (original value - minimum value) / (maximum value - minimum value), which maps the original value to the [0,1] interval, eliminating the influence of different numerical magnitudes on subsequent vector weights. For categorical features in business preference features, such as frequently accessed communication data types including voice, SMS, and email, and key departments of interest including operations and maintenance departments and customer service departments, as well as categorical data in semantic habit features, such as the technical categories corresponding to personalized terms, one-hot encoding is used to generate an independent binary feature bit for each category. When a user feature belongs to a certain category, the corresponding feature bit is assigned a value of 1, and the rest are assigned a value of 0. For example, if a user frequently accesses voice communication data, the feature bit corresponding to voice is 1, and the feature bits corresponding to SMS and email are 0, avoiding interference from the disorder of categorical data on feature embedding.

[0061] Step b: When converting the standardized feature data into a multi-dimensional dynamic feature vector, the method of assigning weights to fixed-dimensional corresponding feature categories can be followed: First, set the total dimensions of the vector based on the three core dimensions of the real-time user profile features, including historical search features, business preference features, and semantic habit features. For example, set 3 dimensions for historical search features, corresponding to high-frequency search keywords, common filter condition combinations, and repeated query patterns; set 2 dimensions for business preference features, corresponding to frequently accessed communication data types and key objects of interest; and set 2 dimensions for semantic habit features, corresponding to personalized terms and common expression styles, forming a total of 7-dimensional vectors. Next, assign values ​​to each dimension: standardized numerical features can be directly used as the corresponding dimension weights, while standardized categorical features are assigned a value of 0.9 for the matching category dimension and 0.1 for the non-matching category dimension. By using the method of dimension-corresponding features + direct value filling, a multi-dimensional dynamic feature vector can be quickly formed. Subsequently, the weights of each dimension can be adjusted as user characteristics change to reflect dynamism.

[0062] Step c: When implementing the dynamic detached attribute of multi-dimensional dynamic feature vectors, a method of recent behavior statistics + scenario factor adaptation is adopted based on the acquired real-time user profile features: First, the user's search behavior data for the past 7 days is statistically analyzed periodically. For example, the number of times high-frequency keywords are used in historical search features and the access frequency of various communication data in business preference features are statistically analyzed. For feature dimensions with high usage frequency, the weight of the corresponding dimension is increased by 0.1 for each additional high-frequency use according to a preset ratio, with an upper limit of 0.9. For feature dimensions with low usage frequency, the weight is decreased by 0.05 for each less frequent use according to a preset ratio, with a lower limit of 0.1. Then, key information of the current search scenario is extracted, such as whether the login device is an office computer or a personal mobile phone, and whether the search time is working hours or non-working hours. The preset scenario factor is 1.0 for office computer / working hours and 0.7 for personal mobile phone / non-working hours. The adjusted weights are multiplied by the corresponding scenario factors to obtain the final dynamic weights, so that the weights of each dimension of the vector change with the user's recent search behavior and the current scenario, thereby realizing the dynamic detached attribute.

[0063] Next, after constructing the multi-dimensional dynamic feature vector, the original query text undergoes structured processing. Specifically: First, the original query text submitted by the user is segmented into words, breaking down continuous text into independent semantic units, i.e., word segmentation results, ensuring accurate extraction of core information from the text. Then, intent anchors are inserted into the word segmentation results. Intent anchors are symbols used to mark the user's core search intent within the word segmentation results. They can be categorized into time anchors, object anchors, and action anchors, corresponding to time information, search object information, and action information in the text, respectively. Their function is to provide clear intent guidance for subsequent vector weight adjustments. For example, if the original query text is "search for voice communication records of core departments in the last 3 days," and the word segmentation results are "search," "last 3 days," "core departments," and "voice communication records," action anchors can be inserted at the "search" position, time anchors at "last 3 days," object anchors at "core departments," and object anchors at "voice communication records," ultimately generating a query intent sequence containing anchor markers.

[0064] Next, based on the query intent sequence, weight updates and vector concatenation operations are performed on the multi-dimensional dynamic feature vectors. First, the generated query intent sequence is loaded, and the core information marked by the intent anchors is parsed, such as the time anchor corresponding to the last 3 days and the object anchor corresponding to voice communication records in the example above. The dimensions of the multi-dimensional dynamic feature vectors associated with this core information are then determined. Next, a probability distribution-based random perturbation update method for vector weights is adopted. Based on the core information of the query intent sequence, within a preset probability distribution range, such as a normal distribution, the vector weights of the associated dimensions are randomly adjusted. This adjustment is guided by the information marked by the intent anchors, ensuring that the adjusted weights better match the current search intent. For example, if a query intent sequence contains an object anchor for voice communication records, the system will randomly perturb and update the weights of the communication data type dimension corresponding to the business preference features in the multi-dimensional dynamic feature vectors within a normal distribution range. Finally, the vectors with different weight combinations obtained after multiple perturbation updates are concatenated to generate multiple differentiated personalized search intent vectors, each corresponding to a potential user search intent direction.

[0065] Furthermore, when determining the range of the normal distribution, the initial weights of the associated dimensions in the multi-dimensional dynamic feature vector (such as the initial weights of 0.9 for matching category and 0.1 for non-matching category in step b above) can be set in conjunction with the degree of correlation between the weights and the search intent. If the core information marked by the intent anchor is strongly correlated with a certain dimension, the mean of the normal distribution can be set to 0 and the standard deviation to 0.1 to ensure that the adjustment range is moderate and does not deviate from the core range of the initial weights. If the correlation is weak, the standard deviation can be reduced to 0.05 to avoid excessive disturbance. In specific operation, the associated dimensions are first locked according to the intent anchor. For example, the time anchor corresponds to the repeated query pattern dimension of historical search features. A random number is generated within the set normal distribution. This random number is added to the current weight of the dimension. Finally, the result is truncated, such as being limited to between 0 and 1, to prevent the weight from exceeding the effective range. This completes the random disturbance update guided by intent, ensuring that the updated weights are more in line with the current search intent.

[0066] Finally, the method constructs a first intent identifier based on the first personalized retrieval intent vector and concatenates it with the original query text to generate a first structured input matrix. It drives the semantic completion model to perform bundle-width adjustable search decoding on the matrix to output a first semantically equivalent candidate set. Then, it extracts time entities and core actions from the original query text to construct core entity conservation rules. Finally, it uses the rules as constraints to calculate the semantic similarity between the candidate set and the original query text, and selects the first semantically equivalent completed statement. The steps A224-1-A224-4 are described in detail.

[0067] Through the above steps, the user's personalized features are deeply integrated with the current search intent, providing multi-directional and accurate input for the subsequent semantic completion model, thereby covering a variety of potential search intents of the user.

[0068] Furthermore, step A224 in the method provided in this application embodiment includes: A224-1: Construct a first intent identifier based on the first personalized search intent vector, and generate a first structured input matrix by concatenating the first intent identifier with the original query text.

[0069] A224-2: Drive the semantic completion model to perform bundle-width adjustable search decoding on the first structured input matrix and output the first semantic equivalence candidate set.

[0070] A224-3: Extract time entities and core actions from the original query text, and construct core entity conservation rules.

[0071] A224-4: Using the core entity conservation rule as a constraint, perform semantic similarity calculation between the first semantic equivalence candidate set and the original query text to filter and obtain the first semantic equivalence complete statement.

[0072] Specifically, firstly, a first intent identifier is constructed based on the first personalized search intent vector. Any personalized search intent vector reflecting a user's specific search tendency is transformed into a fixed-dimensional dense vector using Word2Vec technology as the first intent identifier. This identifier clearly defines the structured direction of semantic completion, preventing the model from deviating from the user's personalized needs. Next, it is concatenated with the word segmentation vector of the original query text, either row-wise or column-wise. The word segmentation vector is processed using the jieba word segmentation tool to obtain a numerical vector format with the same characteristics as the first intent identifier. After concatenation, a first structured input matrix is ​​generated, providing structured input data for the semantic completion model.

[0073] Next, the semantic completion model is driven to perform bundle-width adjustable search decoding on the matrix. The semantic completion model is constructed using a unidirectional LSTM neural network: the input layer receives the first structured input matrix and converts the matrix data into a sequence format that the model can process; the hidden layer captures the semantic associations of the text in the input sequence layer by layer through the unidirectional LSTM network, such as the association between the original query text and the user's personalized intent; the output layer outputs the probability distribution of words that may be generated at each position through a fully connected layer and a softmax activation function. During the training process, a matching dataset of query-semantic completion statements in the communication field is used, such as retrieving base station fault data → retrieving 5G base station fault communication data in the past 7 days, etc. The model parameters are iteratively optimized using the cross-entropy loss function, allowing the model to gradually acquire the ability to generate semantically equivalent statements. The search decoding stage adopts the bundle search method, setting the bundle width to 3, that is, each decoding retains the 3 candidate segments with the highest semantic scores, dynamically adjusting the generation process, and finally outputting the first semantically equivalent candidate set containing multiple candidate statements.

[0074] Subsequently, time entities and core actions are extracted from the original query text to construct core entity conservation rules. A BERT-based named entity recognition method is used to identify time entities from the original query text. These time entities refer to the time range specified in the search request and are key features to ensure the timeliness of search results, such as the last 3 days or a specific year and month. Core actions refer to the user's expected operation behavior and are core elements to define the type of search task, such as retrieval or extraction. Then, core entity conservation rules are constructed, meaning that the subsequently selected completion statements must completely retain these time entities and core actions and cannot be arbitrarily replaced or deleted to ensure that the completion does not deviate from the user's original search intent.

[0075] After that, each candidate statement in the first semantic equivalence candidate set and the original query text are respectively transformed into multi-dimensional semantic vectors, using the TF-IDF word frequency weighting method + basic word segmentation: First, use a conventional word segmentation tool (such as jieba) to segment each candidate statement in the first semantic equivalence candidate set and the original query text respectively, and remove stop words without actual semantics, such as "de", "he", etc.; then count all the segmented words to construct a unified vocabulary, including all the unique words that appear in all texts; subsequently calculate the TF-IDF values of each word in each text, where TF represents the proportion of the number of occurrences of a word in the current text, TF = the number of occurrences of a certain word in the current text ÷ the total number of words in the current text, IDF represents the scarcity degree of a word in all texts, IDF = log 10 (total number of texts ÷ number of texts containing the word), where the fewer times a certain word appears in all texts, the scarcer it is, and the larger the IDF value, which represents that the word has a higher distinguishability for texts and can more accurately distinguish the semantic differences of different texts. Then multiply the two to obtain the weight of the word; finally, arrange the TF-IDF values of all words in the corresponding vocabulary of each text in order, and the formed numerical sequence is the multi-dimensional semantic vector of the text.

[0076] Finally, taking the core entity conservation rule as a constraint, the cosine similarity calculation method is used to screen and complete the statements. The semantic similarity between the two is calculated by cosine similarity = (vector A · vector B) / (||vector A|| × ||vector B||), and the value range is [0, 1]. The closer the value is to 1, the more similar the semantics. The similarity threshold is set to 0.8, and the candidate statements that meet both the core entity conservation rule and the semantic similarity ≥ 0.8 are screened out, that is, the first semantic equivalence completion statements are obtained.

[0077] Through the above steps, combined with the semantic completion model trained with communication field data, the effect of generating accurate and semantically equivalent completion statements while strictly retaining the core elements of the user's original retrieval intention is achieved, providing high-quality input for the subsequent extraction of multiple sets of retrieval feature words.

[0078] Furthermore, step A400 in the method provided by the embodiments of the present application includes: A410: By performing entity linking between the first set of retrieval feature words and the nodes of the domain knowledge graph, the original semantic layer retrieval features are constructed.

[0079] A420: Taking the original semantic layer retrieval features as the root nodes, perform semantic association divergence at a preset extended level to obtain the first-level tree node set.

[0080] A430: Based on the first-level tree node set, construct the first initial-level tree.

[0081] A440: Locally call the first historical recurrence frequency set of the first level tree node set, and calculate the first feature word initial weight set based on the first historical recurrence frequency set.

[0082] A450: After mapping and loading the first feature word initial weight set onto the first initial hierarchical tree, perform weight threshold pruning based on the preset hierarchical decay rule to construct the first-level retrieval feature tree.

[0083] In this embodiment of the application, the domain knowledge graph refers to a special knowledge graph applicable to communication data retrieval scenarios. It includes nodes such as communication devices, data types, and time dimensions, as well as the relationships between nodes. Those skilled in the art can directly reuse existing open-source graphs in the industry.

[0084] In one embodiment, the first set of search feature words extracted from multiple semantically equivalent completion statements in the aforementioned steps is first identified as the vocabulary set that accurately represents the user's core search needs. These first set of search feature words are then matched and associated with nodes in the domain knowledge graph. Firstly, string similarity matching is used. Taking the calculation of the similarity between the base station and the communication infrastructure-base station node name in the domain knowledge graph as an example, an edit distance + substring matching and completion method can be employed, as follows: First, the case of both strings is standardized, and no extra characters are processed. Next, the minimum number of edits required to convert "base station" to "communication infrastructure - base station" is calculated. This requires inserting "communication infrastructure -" before "base station," totaling 7 insertion operations and 7 edits. Then, using the longer string "communication infrastructure - base station" (9 characters long) as a baseline, the basic similarity is calculated as 1 - (number of edits) / (length of the longer string) = 1 - 7 / 9 = 0.22. This similarity calculation completes the entity linking. Further verification is performed using node attributes such as aliases and categories. This confirms a unique correspondence between the first set of search feature words and nodes in the domain knowledge graph. Finally, the associated node combination is used as the original semantic layer search feature, giving the originally isolated search feature words a foundation of domain semantic association.

[0085] Next, using the original semantic layer retrieval features as the root node, the preset extension levels are executed by traversing the knowledge graph association relationships, such as the preset extension 2-layer semantic association divergence: based on the hierarchical relationship and attribute association relationship between nodes in the domain knowledge graph, the expansion is carried out layer by layer. For example, starting from the root node of the base station, the first layer diverges out the hierarchical relationship of its subordinate nodes 5G base station and 4G base station. The second layer further diverges out the attribute association nodes 5G macro base station and 5G micro base station from the 5G base station. Finally, all the diverged nodes are collected to form the first-level tree node set. This node set contains the root node and the retrieval feature words of semantic association at each level, realizing the semantic range expansion of the retrieval feature words.

[0086] Subsequently, based on the first-level tree node set, a recursive construction method is used to construct the first initial hierarchical tree: the original semantic layer retrieval features are set as the top-level root node of the tree, the 5G base stations and 4G base stations obtained from the first layer are set as child nodes of the root node, i.e., the second layer, and the 5G macro base stations and 5G micro base stations obtained from the second layer are set as child nodes of the 5G base stations, i.e., the third layer, forming a first initial hierarchical tree with a clear hierarchical structure. At this time, the tree only contains node information and does not reflect the retrieval importance of the nodes.

[0087] Next, the system locally calls the first historical recurrence frequency set of the first-level tree node set, which is the record of the number of times each node's feature words were used as search conditions in past retrieval processes stored locally by the system. For example, 5G base stations were used 150 times in the past 3 months, and 4G base stations were used 90 times. The initial weight set of the first feature words is calculated using the frequency proportion weighting method: first, the total historical usage frequency of all nodes in the first-level tree node set is counted, assuming it is 300 times, and then the usage frequency proportion of each node is calculated. For example, the weight of 5G base stations = 150 / 300 = 0.5, and the weight of 4G base stations = 90 / 300 = 0.3, etc. This weight value directly reflects the search popularity of the node feature words. The higher the popularity, the greater the weight, and the higher the priority matching in subsequent searches.

[0088] Finally, the initial weight set of the first feature words is mapped and loaded into the first initial hierarchical tree according to the node correspondence. Then, hierarchical attenuation threshold pruning is performed to prune the weights: the preset hierarchical attenuation rule is that for each level down, the node weight needs to be multiplied by an attenuation coefficient of 0.8 and the minimum retention threshold for the weight threshold is 0.2. For example, the initial weight of the third-level 5G macro base station is 0.5 × 0.8 = 0.4. Since it extends down one level from the second-level 5G base station, attenuation operation needs to be performed. If its weight is higher than the minimum retention threshold of 0.2, it is retained; while nodes with weights lower than 0.2 are pruned and removed. Finally, a first-level retrieval feature tree with a concise structure, reasonable node weights, and clear hierarchical semantics is constructed.

[0089] By following the above steps and combining knowledge graphs in the communications field with historical retrieval data, isolated search terms are transformed into hierarchical search feature trees with closely related semantics, weights adapted to search requirements, and a concise structure, thus providing efficient feature support for subsequent accurate retrieval.

[0090] Furthermore, step A500 in the method provided in this application embodiment includes: A510: Extract the weight set of the first node of the first-level retrieval feature tree.

[0091] A520: Calculate the depth of the first tree and the entropy of the first weight distribution based on the weight set of the first node.

[0092] A530: After weighted normalization of the first tree depth and the first weight distribution entropy, and outputting the first retrieval difficulty coefficient, the task is classified according to the first retrieval difficulty coefficient, and the first retrieval task level is output.

[0093] A540: Locate the first distributed retrieval unit based on the first retrieval task level.

[0094] In this embodiment, the first distributed retrieval unit is a retrieval unit that adopts a distributed retrieval node architecture, such as a Spark-based distributed computing node, and has independent data query and processing capabilities.

[0095] Optionally, firstly, for the constructed first-level retrieval feature tree, the first node weight set is extracted using a tree node traversal algorithm: by using depth-first traversal or breadth-first traversal, all nodes of the first-level retrieval feature tree are visited in sequence, including the root node and child nodes at each level. The initial weights of the pre-loaded feature words in each node are read. The weight values ​​are derived from the initial weight set of the first feature words calculated based on the historical recurrence frequency in the previous steps. After hierarchical decay pruning, the weight values ​​of all nodes are arranged into a set form according to the node hierarchy, thus obtaining the first node weight set.

[0096] Next, the depth of the first tree and the entropy of the first weight distribution are calculated based on the first node weight set. When calculating the depth of the first tree, the longest path node counting method is used. Starting from the root node, all child node paths are traversed, and the total number of nodes in the path with the most nodes is counted. Then, 1 is subtracted. The path length = number of nodes - 1, which is the depth of the first tree. This depth reflects the structural complexity of the retrieval feature tree. The greater the depth, the more levels need to be traversed during retrieval. When calculating the entropy of the first weight distribution, the information entropy calculation formula H = -Σ(pi × log2pi) is used, where pi is the proportion of the weight of a single node to the total weight of all nodes. First, the total weight of all weights in the first node weight set is calculated, and then the proportion pi of each weight is calculated in turn. Substituting these values ​​into the formula, the entropy of the first weight distribution is obtained. The larger the entropy value, the more uniform the distribution of node weights, and the more nodes need to be balanced during retrieval.

[0097] Then, a weighted normalization method is used to process the obtained first tree depth and first weight distribution entropy, outputting the first retrieval difficulty coefficient. The task is then classified based on this coefficient. Specifically: First, the first tree depth and first weight distribution entropy are normalized, mapping their values ​​to the [0,1] interval. For example, if the preset maximum tree depth is 5, and the first tree depth is 2, the normalized value is 2 / 5 = 0.4; if the preset maximum weight distribution entropy is 2, and the first weight distribution entropy is 1.58, the normalized value is 1.58 / 2 = 0.79. Next, based on the retrieval task's focus on structural complexity and weight uniformity, weights are assigned to the two normalized indicators: tree depth with a weight of 0.4, and weight distribution entropy with a weight of 0.6. The weighted sum is calculated by multiplying the indicator value by the weight, and this weighted sum is the first retrieval difficulty coefficient. Then, based on the preset difficulty level thresholds (0-0.3 for easy, 0.3-0.7 for medium, and 0.7-1 for complex), the range of the first search difficulty coefficient is determined, and the corresponding first search task level is output.

[0098] Finally, based on the first retrieval task level, a retrieval unit matching mechanism is used to locate the first distributed retrieval unit: a matching mapping relationship between retrieval task level and distributed retrieval unit performance is pre-established. For example, simple level matches basic performance retrieval units; medium level matches standard performance retrieval units; and complex level matches high performance retrieval units, where the computing power of the retrieval unit also increases accordingly. Then, based on the output first retrieval task level, retrieval units that meet the performance requirements are found, and their network addresses, resource identifiers, and other information are determined, thus completing the location of the first distributed retrieval unit.

[0099] By combining tree traversal, information entropy calculation, and weighted normalization, a distributed retrieval unit is accurately matched to the complexity of the hierarchical retrieval feature tree, ensuring efficient execution of the retrieval task.

[0100] Furthermore, step A600 in the method provided in this application embodiment includes: A610: The first distributed retrieval unit initiates an initial query based on the top-level node feature words of the first-level retrieval feature tree and outputs the first-level retrieval results.

[0101] A620: Based on the first-level retrieval results, the first distributed retrieval unit dynamically adds the second-level node feature words of the first-level retrieval feature tree to perform a secondary query and outputs the second-level retrieval results.

[0102] A630: Based on the second-level search results, the first distributed retrieval unit dynamically adds the third-level node feature words of the first-level retrieval feature tree and performs three queries.

[0103] A640: By analogy, traverse the first-level retrieval feature tree to iteratively superimpose the feature word conditions until the first global retrieval result set is output.

[0104] A650: Using the search permission strategy as the result desensitization constraint, perform desensitization and removal of unauthorized fields in the first global search result set, and output the first refined result set.

[0105] In one embodiment, the first distributed retrieval unit first locates the top-level node feature word of the first-level retrieval feature tree. This feature word is the core entry point of the tree and is determined by the top-level node in the first-level retrieval feature tree constructed in the preceding steps. For example, communication data represents the core data category to be retrieved. Subsequently, the retrieval unit calls the communication data retrieval engine, using the top-level node feature word as the sole query condition, and matches all raw data containing that feature word in the preset communication data repository. The dataset that meets the core category requirements is then selected as the first-level retrieval result.

[0106] After obtaining the first-level search results, the first distributed retrieval unit does not re-query the full data. Instead, it performs condition superposition based on the existing result set, extracting second-level node feature words from the first-level retrieval feature tree. These feature words are semantically related child nodes of the top-level nodes, such as voice communication data, further limiting the data type. These feature words are then dynamically added to the initial query rules as new query conditions. Next, the retrieval unit performs a secondary filtering within the first-level search results, retaining only datasets that simultaneously contain both top-level node feature words and second-level node feature words, and outputting the second-level search results.

[0107] Then, continuing the logic of the above secondary query steps, the first distributed retrieval unit extracts third-level node feature words from the first-level retrieval feature tree. These feature words are subdivided semantic nodes of the second-level nodes, such as voice communication data from the past 7 days, adding a time dimension constraint. Based on the second-level retrieval results, this feature word is dynamically added again as a query condition. Subsequently, the retrieval unit performs three filtering operations on the second-level retrieval results, eliminating data that does not meet the combination of time + type + core category conditions, and retaining the dataset that meets all the superimposed conditions.

[0108] Subsequently, the first distributed retrieval unit follows the hierarchical structure of the first-level retrieval feature tree. For example, if the tree contains three levels of nodes, it traverses to the third level and continues to perform subsequent query operations in a similar manner. At each level, it extracts the node feature words from the tree and dynamically adds them to existing query condition combinations, then performs filtering within the current retrieval result set. This iterative process continues until all valid levels of the tree have been traversed, meaning there are no more node feature words to add to the next level. The resulting dataset is the first global retrieval result set. This result set is a complete dataset that meets the feature word combination conditions of all levels of the first-level retrieval feature tree, achieving semantic retrieval condition coverage from core to subdivided levels. Furthermore, all filtering is completed iteratively within the same result set, avoiding the complexity of aggregating results from multiple sources.

[0109] Finally, the first distributed retrieval unit invokes the user ID-based associated retrieval permission strategy from step A140. For example, it uses the restriction that ordinary users are not authorized to access sensitive fields such as user ID numbers and communication content details as a result desensitization constraint to perform field-level desensitization processing on the first global retrieval result set. This involves replacing characters in unauthorized fields, masking sensitive information with appropriate symbols, or directly removing fields that exceed permissions, thus deleting or hiding all unauthorized fields in the result set that do not meet permission requirements.

[0110] The final output is the first refined result set, which fully meets the iterative filtering conditions of the first-level retrieval feature tree, satisfies the security access requirements of user permissions, and is the result of dynamic superposition and iterative narrowing of conditions within a single level tree, rather than the summary of multiple rounds of independent query results, ensuring the unity of data accuracy and compliance.

[0111] In summary, the semantic parsing-based intelligent retrieval method for communication data provided in this application has the following technical effects: This application, upon receiving the original query text submitted by a user, triggers a user ID-based associated retrieval permission strategy binding. Based on the user ID, a user profile is loaded to semantically complete the original query text, resulting in multiple semantically equivalent completed statements. Feature words of these statements are extracted and combined with a domain knowledge graph to construct multiple hierarchical retrieval feature trees. Multiple distributed retrieval units are located based on the node weight features of the trees. These units dynamically superimpose conditions on the feature trees from top to bottom, outputting multiple refined result sets. Finally, the metadata timestamps of the result sets are retrieved and sorted in descending order for integration. This accurately outputs a time-series retrieval sequence of communication data, making the intelligent retrieval results of communication data more precise and in line with user permissions and personalized needs. This achieves the technical effects of accurate semantic understanding of communication data, orderly and efficient retrieval, and precise location of communication data that matches user intent.

[0112] Example 2, as Figure 2As shown, based on the same inventive concept as in Embodiment 1 above, this application provides a semantic parsing-based intelligent communication data retrieval platform, the platform comprising: The associated retrieval permission policy execution module 1 is used to trigger the binding of associated retrieval permission policies based on user ID after receiving the original query text submitted by the user.

[0113] The semantic equivalence completion statement acquisition module 2 is used to load the user profile based on the user ID and perform semantic completion on the original query text to obtain multiple semantic equivalence completion statements.

[0114] The feature word acquisition module 3 is used to extract feature words from the multiple semantically equivalent completion statements to obtain multiple sets of search feature words.

[0115] The hierarchical retrieval feature tree construction module 4 constructs multiple hierarchical retrieval feature trees by semantically associating the multiple sets of retrieval feature words based on the domain knowledge graph.

[0116] The distributed retrieval unit positioning module 5 is used to perform hierarchical narrowing retrieval task matching based on the node weight features of the multiple hierarchical retrieval feature trees, and to locate multiple distributed retrieval units.

[0117] The refined result set acquisition module 6 is used to drive the multiple distributed retrieval units to perform top-down conditional superposition retrieval of the multiple hierarchical retrieval feature trees under the constraint of the retrieval permission policy, and output multiple refined result sets.

[0118] The time-series retrieval sequence acquisition module 7 is used to retrieve the metadata timestamps of the multiple refined result sets, sort and integrate them in descending order, and output the time-series retrieval sequence of communication data.

[0119] Furthermore, the associated retrieval permission policy execution module 1 is used to perform the following steps: Extract real-time request attributes and the user ID from the request context of the original query text, wherein the real-time request attributes include user identity / role and operating environment status; associate the user ID with the security audit database to retrieve historical unauthorized operation records and construct an operation risk tag; match the initial search permission according to the user identity / role, and then dynamically downgrade the initial search permission based on the operation risk tag, outputting the corrected search permission; perform real-time risk control to adapt the corrected search permission according to the operating environment status, and output the associated search permission strategy.

[0120] Furthermore, the semantic equivalence completion statement acquisition module 2 is used to perform the following steps: Using the user ID as the query key, the user profile database is queried to obtain real-time user profile features; the original query text and real-time user profile features are concatenated into a structured input, and multi-branch semantic completion is performed through a semantic completion model to generate multiple semantically equivalent completion statements corresponding to multiple personalized search intent vectors.

[0121] Furthermore, the semantic equivalence completion statement acquisition module 2 is used to perform the following steps: The real-time user profile features are converted into multi-dimensional dynamic feature vectors, wherein the multi-dimensional dynamic feature vectors have dynamic detached properties; after word segmentation from the original query text, intent anchors are inserted to generate a query intent sequence; based on the query intent sequence, the multi-dimensional dynamic feature vectors are updated by random perturbation of vector weights based on probability distribution, and the multiple personalized search intent vectors are constructed by concatenating the vectors; the original query text and the multiple personalized search intent vectors are used as structured inputs, and multi-branch semantic completion is performed through the semantic completion model to generate the multiple semantically equivalent completed sentences.

[0122] Furthermore, the semantic equivalence completion statement acquisition module 2 is used to perform the following steps: A first intent identifier is constructed based on a first personalized retrieval intent vector, and a first structured input matrix is ​​generated by concatenating the first intent identifier with the original query text. The semantic completion model is driven to perform bundle-width adjustable search decoding on the first structured input matrix, and output a first semantic equivalence candidate set. Time entities and core actions are extracted from the original query text, and core entity conservation rules are constructed. The semantic similarity between the first semantic equivalence candidate set and the original query text is calculated using the core entity conservation rules as constraints, so as to filter and obtain the first semantic equivalence completion statement.

[0123] Furthermore, the hierarchical retrieval feature tree construction module 4 is used to perform the following steps: By linking the first set of retrieval feature words with the nodes of the domain knowledge graph, an original semantic layer retrieval feature is constructed; using the original semantic layer retrieval feature as the root node, a preset extension level of semantic association divergence is performed to obtain a first-level tree node set; a first initial level tree is constructed based on the first level tree node set; the first historical recurrence frequency set of the first level tree node set is locally called, and the first feature word initial weight set is calculated based on the first historical recurrence frequency set; after mapping and loading the first feature word initial weight set onto the first initial level tree, weight threshold pruning is performed based on a preset level decay rule to construct a first-level retrieval feature tree.

[0124] Furthermore, the refining result set acquisition module 6 is used to perform the following steps: The first distributed retrieval unit initiates an initial query based on the top-level node feature words of the first-level retrieval feature tree and outputs the first-level retrieval results. Based on the first-level retrieval results, the first distributed retrieval unit dynamically adds the second-level node feature words of the first-level retrieval feature tree to perform a second query and outputs the second-level retrieval results. Based on the second-level retrieval results, the first distributed retrieval unit dynamically adds the third-level node feature words of the first-level retrieval feature tree to perform a third query. The first-level retrieval feature tree is traversed in a similar manner to iterate and superimpose feature word conditions until the first global retrieval result set is output. Using the retrieval permission strategy as the result desensitization constraint, the unauthorized fields in the first global retrieval result set are desensitized and removed, and the first refined result set is output.

[0125] Furthermore, the distributed retrieval unit positioning module 5 is used to perform the following steps: Extract the first node weight set of the first-level retrieval feature tree; calculate the first tree depth and the first weight distribution entropy based on the first node weight set; weight normalize the first tree depth and the first weight distribution entropy, output the first retrieval difficulty coefficient, classify the task according to the first retrieval difficulty coefficient, and output the first retrieval task level; match and locate the first distributed retrieval unit according to the first retrieval task level.

[0126] Furthermore, the semantic equivalence completion statement acquisition module 2 is used to perform the following steps: The real-time user profile features include historical search features, business preference features, and semantic habit features.

[0127] The semantic parsing-based intelligent retrieval platform for communication data provided in this embodiment of the invention can execute the semantic parsing-based intelligent retrieval method for communication data provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0128] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0129] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for intelligent retrieval of communication data based on semantic parsing, characterized in that, The method includes: Upon receiving the original query text submitted by the user, the associated retrieval permission policy binding based on the user ID is triggered; Based on the user ID, load the user profile and perform semantic completion on the original query text to obtain multiple semantically equivalent completed statements; By extracting feature words from the multiple semantically equivalent completion statements, multiple sets of retrieval feature words are obtained; Based on the domain knowledge graph, the semantic associations of the multiple sets of retrieval feature words are diverged to construct multiple hierarchical retrieval feature trees; Based on the node weight features of the multiple hierarchical retrieval feature trees, hierarchical narrowing retrieval task matching is performed to locate multiple distributed retrieval units; With the retrieval permission policy as a constraint, the multiple distributed retrieval units are driven to perform top-down conditional dynamic superposition retrieval of the multiple hierarchical retrieval feature trees, and output multiple refined result sets; The metadata timestamps of the multiple refined result sets are retrieved, sorted in descending order, and integrated to output a time-series retrieval sequence of communication data.

2. The intelligent communication data retrieval method based on semantic parsing as described in claim 1, characterized in that, After receiving the original query text submitted by the user, the method triggers the binding of the associated retrieval permission policy based on the user ID, which includes: Extract real-time request attributes and the user ID from the request context of the original query text, wherein the real-time request attributes include user identity role and operating environment status; The user ID is associated with the security audit database to retrieve historical unauthorized operation records, and operation risk tags are constructed. After matching the initial search permissions according to the user's identity and role, the dynamic permission downgrade correction of the initial search permissions is performed based on the operation risk tag, and the corrected search permissions are output. Based on the operating environment status, perform real-time risk control to adapt the modified search permissions, and output the associated search permission strategy.

3. The intelligent communication data retrieval method based on semantic parsing as described in claim 1, characterized in that, The method involves loading a user profile based on the user ID and performing semantic completion on the original query text to obtain multiple semantically equivalent completed statements. Using the user ID as the query key, query the user profile database to retrieve real-time user profile features; The original query text and real-time user profile features are concatenated into a structured input, and then a semantic completion model is used to perform multi-branch semantic completion to generate multiple semantically equivalent completion statements corresponding to multiple personalized search intent vectors.

4. The intelligent communication data retrieval method based on semantic parsing as described in claim 3, characterized in that, The method involves concatenating the original query text and real-time user profile features into a structured input, performing multi-branch semantic completion via a semantic completion model, and generating multiple semantically equivalent completed statements corresponding to multiple personalized search intent vectors. The real-time user profile features are converted into multi-dimensional dynamic feature vectors, wherein the multi-dimensional dynamic feature vectors have dynamic detached properties; After extracting and segmenting words from the original query text, intent anchor characters are inserted to generate a query intent sequence; Based on the query intent sequence, the multi-dimensional dynamic feature vector is updated by random perturbation of vector weights based on probability distribution, and then the multiple personalized search intent vectors are constructed by concatenating the vectors. The original query text and multiple personalized search intent vectors are used as structured inputs, and the semantic completion model performs multi-branch semantic completion to generate the multiple semantically equivalent completed statements.

5. The intelligent communication data retrieval method based on semantic parsing as described in claim 4, characterized in that, The method involves taking the original query text and multiple personalized search intent vectors as structured inputs, performing multi-branch semantic completion via the semantic completion model, and generating the multiple semantically equivalent completed statements. A first intent identifier is constructed based on a first personalized search intent vector, and a first structured input matrix is ​​generated by concatenating the first intent identifier with the original query text. The semantic completion model is driven to perform a beam-width adjustable search decoding on the first structured input matrix, and outputs a first semantic equivalence candidate set; Extract time entities and core actions from the original query text, and construct core entity conservation rules; Using the core entity conservation rule as a constraint, the semantic similarity between the first semantic equivalence candidate set and the original query text is calculated to filter and obtain the first semantic equivalence complete statement.

6. The intelligent communication data retrieval method based on semantic parsing as described in claim 1, characterized in that, The method involves semantic association divergence based on a domain knowledge graph to construct a multi-level retrieval feature tree, comprising: By linking the first set of retrieval feature words with the nodes of the domain knowledge graph, the original semantic layer retrieval features are constructed; Using the original semantic layer retrieval features as the root node, perform semantic association divergence at a preset extension level to obtain the first-level tree node set; Construct a first initial hierarchical tree based on the first hierarchical tree node set; The first historical recurrence frequency set of the first level tree node set is called locally, and the initial weight set of the first feature word is calculated based on the first historical recurrence frequency set; After mapping and loading the first initial weight set of the first feature words into the first initial hierarchical tree, weight threshold pruning is performed based on the preset hierarchical decay rule to construct the first-level retrieval feature tree.

7. The intelligent communication data retrieval method based on semantic parsing as described in claim 1, characterized in that, Using the aforementioned retrieval permission policy as a constraint, the method drives the multiple distributed retrieval units to perform top-down conditional dynamic superposition retrieval of the multiple hierarchical retrieval feature trees, outputting multiple refined result sets. The method includes: The first distributed retrieval unit initiates an initial query based on the top-level node feature words of the first-level retrieval feature tree and outputs the first-level retrieval results; Based on the first-level retrieval results, the first distributed retrieval unit dynamically adds the second-level node feature words of the first-level retrieval feature tree to perform a secondary query and outputs the second-level retrieval results; Based on the second-level search results, the first distributed retrieval unit dynamically adds the third-level node feature words of the first-level retrieval feature tree and performs three queries. By analogy, the first-level retrieval feature tree is traversed to iteratively superimpose feature word conditions until the first global retrieval result set is output; Using the search permission policy as the result desensitization constraint, the unauthorized fields in the first global search result set are desensitized and removed, and the first refined result set is output.

8. The intelligent retrieval method for communication data based on semantic parsing as described in claim 6, characterized in that, The method involves matching hierarchical narrowing retrieval tasks based on the node weight features of the multiple hierarchical retrieval feature trees to locate multiple distributed retrieval units. Extract the weight set of the first node of the first-level retrieval feature tree; Calculate the depth of the first tree and the entropy of the first weight distribution based on the weight set of the first node; After weighted normalization of the first tree depth and the first weight distribution entropy, and outputting the first retrieval difficulty coefficient, the task is classified according to the first retrieval difficulty coefficient, and the first retrieval task level is output. The first distributed retrieval unit is located based on the first retrieval task level.

9. The intelligent communication data retrieval method based on semantic parsing as described in claim 3, characterized in that, The real-time user profile features include historical search features, business preference features, and semantic habit features.

10. A semantic parsing-based intelligent communication data retrieval platform, characterized in that, The platform is used to implement the semantic parsing-based intelligent retrieval method for communication data according to any one of claims 1-9, and the platform includes: The associated retrieval permission policy execution module is used to trigger the binding of associated retrieval permission policies based on user ID after receiving the original query text submitted by the user. The semantic equivalence completion statement acquisition module is used to load the user profile based on the user ID and perform semantic completion on the original query text to obtain multiple semantic equivalence completion statements; The feature word acquisition module is used to extract feature words from the multiple semantically equivalent completion statements to obtain multiple sets of search feature words; The hierarchical retrieval feature tree construction module uses a domain knowledge graph to perform semantic association divergence of the multiple sets of retrieval feature words in order to construct multiple hierarchical retrieval feature trees; The distributed retrieval unit positioning module is used to perform hierarchical narrowing retrieval task matching based on the node weight features of the multiple hierarchical retrieval feature trees, and to locate multiple distributed retrieval units. The refined result set acquisition module is used to drive the multiple distributed retrieval units to perform top-down conditional dynamic superposition retrieval of the multiple hierarchical retrieval feature trees, constrained by the retrieval permission policy, and output multiple refined result sets. The time-series retrieval sequence acquisition module is used to retrieve the metadata timestamps of the multiple refined result sets, sort and integrate them in descending order, and output the time-series retrieval sequence of communication data.

Images