Query data processing method and device, equipment, readable storage medium and product

Abstract

This application relates to a query data processing method, apparatus, device, readable storage medium, and product. The method includes: acquiring charging query data sent in the current query dialogue through a target account and historical dialogue data of the target account; identifying the current service stage of the current query dialogue based on the historical dialogue data; constructing context anchors based on the current service stage and historical dialogue data; performing semantic extraction on the charging query data based on the context anchors to obtain original semantic elements; invoking a charging service knowledge graph; adjusting the basic prompt word template based on the context anchors to obtain target prompt words; and obtaining the target semantic parsing result of the charging query data through semantic inference processing based on the target prompt words and original semantic elements. The target semantic parsing result is used to determine the response information regarding the charging query data and sends the response information through a charging customer service account, thereby improving the accuracy of semantic understanding.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a query data processing method, apparatus, device, readable storage medium, and product. Background Technology

[0002] With the development of computer technology, the electric vehicle industry has also experienced rapid growth. Charging infrastructure coverage is constantly expanding, and user charging scenarios are becoming increasingly diverse. Consequently, problems such as charging start-up failures, charging interruptions, abnormal charges, and equipment malfunctions are also increasing. This has made intelligent charging customer service a core support for efficiently resolving user requests and improving service experience. Semantic understanding, as a core component of intelligent charging customer service, directly determines the accuracy of the system's recognition of user query intent and the effectiveness of its problem response, thus being crucial for ensuring customer service quality.

[0003] In related technologies, a common approach is based on intent classification and slot filling. This relies on predefined intent categories and fixed user expression templates, using a natural language understanding module to parse individual user inputs in isolation to achieve intent recognition and key information extraction. However, this approach struggles to accurately pinpoint the core message in scenarios involving ambiguous references or omitted expressions, resulting in low accuracy in semantic understanding. Summary of the Invention

[0004] Therefore, it is necessary to provide a query data processing method, apparatus, device, readable storage medium, and product that can improve the accuracy of semantic understanding in response to the above-mentioned technical problems.

[0005] Firstly, this application provides a method for processing query data, including:

[0006] Obtain charging query data sent through the target account in the current query conversation and the historical conversation data of the target account, wherein the current query conversation is a charging query conversation between the target account and the charging customer service account;

[0007] Based on the historical dialogue data, the current service stage of the current query dialogue is identified, and a context anchor point for the charging query data is constructed based on the current service stage and the historical dialogue data.

[0008] Based on the context anchor, semantic extraction is performed on the charging query data to obtain the original semantic elements. The charging service knowledge graph is then invoked, and the basic prompt word template corresponding to the current service stage is adjusted based on the context anchor to obtain the target prompt word.

[0009] Based on the target prompt words and the original semantic elements, the target semantic parsing result of the charging query data is obtained through semantic inference processing. The target semantic parsing result is used to determine the reply information about the charging query data and send the reply information through the charging customer service account.

[0010] Secondly, this application also provides a query data processing apparatus, comprising:

[0011] The data acquisition module is used to acquire charging query data sent by the target account in the current query conversation and the historical conversation data of the target account. The current query conversation is a charging query conversation between the target account and the charging customer service account.

[0012] The data recognition module is used to identify the current service stage of the current query dialogue based on the historical dialogue data, and to construct a context anchor point for the charging query data based on the current service stage and the historical dialogue data.

[0013] The template adjustment module is used to perform semantic extraction on the charging query data based on the context anchor point to obtain the original semantic elements, call the charging service knowledge graph, and adjust the basic prompt word template corresponding to the current service stage based on the context anchor point to obtain the target prompt word.

[0014] The semantic parsing module is used to obtain the target semantic parsing result of the charging query data through semantic inference processing based on the target prompt words and the original semantic elements. The target semantic parsing result is used to determine the reply information about the charging query data and send the reply information through the charging customer service account.

[0015] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0016] Obtain charging query data sent through the target account in the current query conversation and the historical conversation data of the target account, wherein the current query conversation is a charging query conversation between the target account and the charging customer service account;

[0017] Based on the historical dialogue data, the current service stage of the current query dialogue is identified, and a context anchor point for the charging query data is constructed based on the current service stage and the historical dialogue data.

[0018] Based on the context anchor, semantic extraction is performed on the charging query data to obtain the original semantic elements. The charging service knowledge graph is then invoked, and the basic prompt word template corresponding to the current service stage is adjusted based on the context anchor to obtain the target prompt word.

[0019] Based on the target prompt words and the original semantic elements, the target semantic parsing result of the charging query data is obtained through semantic inference processing. The target semantic parsing result is used to determine the reply information about the charging query data and send the reply information through the charging customer service account.

[0020] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0021] Obtain charging query data sent through the target account in the current query conversation and the historical conversation data of the target account, wherein the current query conversation is a charging query conversation between the target account and the charging customer service account;

[0022] Based on the historical dialogue data, the current service stage of the current query dialogue is identified, and a context anchor point for the charging query data is constructed based on the current service stage and the historical dialogue data.

[0023] Based on the context anchor, semantic extraction is performed on the charging query data to obtain the original semantic elements. The charging service knowledge graph is then invoked, and the basic prompt word template corresponding to the current service stage is adjusted based on the context anchor to obtain the target prompt word.

[0024] Based on the target prompt words and the original semantic elements, the target semantic parsing result of the charging query data is obtained through semantic inference processing. The target semantic parsing result is used to determine the reply information about the charging query data and send the reply information through the charging customer service account.

[0025] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0026] Obtain charging query data sent through the target account in the current query conversation and the historical conversation data of the target account, wherein the current query conversation is a charging query conversation between the target account and the charging customer service account;

[0027] Based on the historical dialogue data, the current service stage of the current query dialogue is identified, and a context anchor point for the charging query data is constructed based on the current service stage and the historical dialogue data.

[0028] Based on the context anchor, semantic extraction is performed on the charging query data to obtain the original semantic elements. The charging service knowledge graph is then invoked, and the basic prompt word template corresponding to the current service stage is adjusted based on the context anchor to obtain the target prompt word.

[0029] Based on the target prompt words and the original semantic elements, the target semantic parsing result of the charging query data is obtained through semantic inference processing. The target semantic parsing result is used to determine the reply information about the charging query data and send the reply information through the charging customer service account.

[0030] The aforementioned query data processing method, apparatus, device, readable storage medium, and product acquire charging query data sent by the target account in the current query dialogue and historical dialogue data of the target account. The current query dialogue is a charging query dialogue between the target account and the charging customer service account. Based on the historical dialogue data, the current service stage of the current query dialogue is identified, achieving a deep binding between the dialogue context and the charging service process, and clearly locating the service node of the current charging dialogue. Furthermore, based on the current service stage and historical dialogue data, and combining the dimensions of historical dialogue and the dimensions of the current actual service node, contextual anchors can be accurately identified from the charging query data. Based on the contextual anchors, the semantics of the charging query data are accurately extracted to obtain the original semantic elements. That is to say, the historical dialogue data with the dimension of historical dialogue can help extract the semantics of the charging query data of the current query dialogue, avoiding isolated parsing based on the current input. Thus, it can accurately locate the core demand when facing ambiguous references and omitted expressions. Then, the charging service knowledge graph is invoked, and the basic prompt word template corresponding to the current service stage is adjusted based on contextual anchors to obtain highly accurate target prompt words. Based on the target prompt words and original semantic elements, semantic inference processing is used to accurately obtain the target semantic parsing results of the charging query data. The target semantic parsing results are used to determine the response information regarding the charging query data and send the response information through the charging customer service account. This improves the accuracy of semantic understanding. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1This is an application environment diagram of a query data processing method in one embodiment;

[0033] Figure 2 This is a flowchart illustrating a data processing method in one embodiment;

[0034] Figure 3 This is a structural block diagram of a query data processing device in one embodiment;

[0035] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0037] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0038] The query data processing method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on another network server.

[0039] In some embodiments, after obtaining charging query data sent through the target account in the current query dialogue, terminal 102 transmits the charging query data to server 104. Server 104, after obtaining the charging query data from the target account, also obtains the target account's historical dialogue data. Based on the historical dialogue data, server 104 identifies the current service stage of the current query dialogue. Based on the current service stage and historical dialogue data, server 104 constructs context anchors for the charging query data. Based on the context anchors, server 104 performs semantic extraction on the charging query data to obtain original semantic elements. It then calls the charging service knowledge graph and, based on the context anchors, adjusts the basic prompt word template corresponding to the current service stage to obtain target prompt words. Based on the target prompt words and original semantic elements, server 104 performs semantic inference processing to obtain the target semantic parsing result of the charging query data. Based on the target semantic parsing result, server 104 determines the response information regarding the charging query data and sends the response information through the charging customer service account.

[0040] Terminal 102 refers to the terminal held by the target object corresponding to the target object account. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, etc. Server 104 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0041] In one exemplary embodiment, such as Figure 2 As shown, a query data processing method is provided, which can be applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps S202 to S208. Wherein:

[0042] Step S202: Obtain the charging query data sent by the target account in the current query conversation and the historical conversation data of the target account. The current query conversation is a charging query conversation between the target account and the charging customer service account.

[0043] The current query dialogue refers to the query dialogue initiated by the target account at this time, which is a charging query dialogue between the target account and the charging customer service account. Charging query data refers to the data sent by the target account regarding charging queries. In some embodiments, the charging query data can be in text or image format. In other embodiments, the charging query data is multimodal data, including text statements and image data. Historical dialogue data is a collection of interaction data generated by the target account within historical queries; it can also be a collection of interaction data generated by the target account within the current query dialogue. Historical dialogue data includes the target account's historical input information, historical feedback information, and historical interaction timestamps.

[0044] For example, after obtaining the charging query data sent through the target account in the current query conversation, the server retrieves the historical conversation data of the target account from the cache.

[0045] Step S204: Based on historical dialogue data, identify the current service stage of the current query dialogue, and construct context anchors for the charging query data based on the current service stage and historical dialogue data.

[0046] The current service stage refers to the service stage in which the current query dialogue is located. Service stages can include various core service stages such as pre-charging consultation, in-charging troubleshooting, post-charging settlement consultation, charging equipment reservation, and member service consultation. Context anchors are charging service entities with clear identifiers and attributes, and are structured semantic reference units formed by attaching meta-information such as their source, timeliness, and stage association to the charging service entity. Specifically, a charging service entity refers to an object with a clear semantic orientation in the charging service scenario, including charging station entities, charging pile equipment entities, charging service function entities, and fault code entities; source meta-information refers to the acquisition channel of the charging service entity (historical dialogue data or charging query data); timeliness meta-information refers to the valid time range of the charging service entity; and stage association meta-information refers to the degree of association between the charging service entity and the current service stage (represented by a value in the range of 0-1, with the value closer to 1 indicating a higher degree of association).

[0047] In some embodiments, after acquiring historical dialogue data, the server identifies the service stage based on the context sequence in the historical dialogue data to determine the current service stage of the current query dialogue. The context sequence in the historical dialogue data refers to an ordered data sequence sorted by interaction timestamps, containing intent-related information, system service response direction information, and interaction topic evolution information for each interaction node. Optionally, after identifying the context sequence in the historical dialogue data, the server extracts context features such as topic keywords, interaction frequency, and problem-solving progress indicators from the context sequence. Then, it calls a pre-built feature library of the entire charging service process, matching the extracted context sequence features with the feature library to determine the current service stage of the current query dialogue. The pre-built feature library of the entire charging service process is pre-constructed by a semantic understanding system and includes multiple core service stages such as pre-charging consultation, in-charging troubleshooting, post-charging settlement consultation, charging equipment reservation, and member service consultation. Each service stage corresponds to a unique set of context features. The components of the context feature set include a core theme keyword set, a typical object question type set, interaction progress identification rules, and related service module identifiers.

[0048] For example, the text statement in the charging query data is "Why can't this charging pile start?", and the image data in the charging query data is an image of "Fault Code E2" (obtained by the target object taking a picture of the charging pile's operation panel). Historical dialogue data includes: [timestamp t1] the target object's input text statement "What charging piles are available nearby?", [timestamp t2] the semantic understanding system's feedback text statement "There are 3 available charging stations within 3 kilometers of your current location: X Park Charging Station (2 available DC charging piles) and Y Shopping Mall Charging Station (1 available AC charging pile). Do you need detailed navigation information?", [timestamp t3] the target object's input text statement "Which charging interface does the X Park Charging Station support?", and [timestamp t4] the semantic understanding system's feedback text statement "The DC charging piles at X Park Charging Station support standard interfaces, compatible with most new energy passenger vehicles. Do you need to reserve a charging pile at this station?". Where t1 < t2 < t3 < t4 < t5, that is, t1 is earlier than t2, t2 is earlier than t3, t3 is earlier than t4, and t4 is earlier than t5.

[0049] The server then extracts the context sequence (sorted by timestamp) from historical dialogue data: [Target object inquires about nearby available charging stations → System provides feedback on available stations and asks if navigation is needed → Target object inquires about the charging station interface type at the target station → System provides feedback on the interface standard and asks if a reservation is needed]. Further extraction of the context sequence's features reveals the core keywords as "available charging stations," "charging station interface," and "reservation." Typical object question types include charging station queries and equipment parameter inquiries. The interaction progress is indicated by "Charging station reservation and preparation not completed."

[0050] The server matches the extracted contextual features with the feature library of the entire charging service process, and obtains the matching results. The matching results show that the contextual feature set of the pre-charging consultation stage has the highest matching degree with the extracted contextual features, reaching 92%. Therefore, the current service stage is identified as the "pre-charging consultation stage". Among them, the contextual feature set of the pre-charging consultation stage includes core topic keywords "available charging piles", "charging interface", "reservation", typical object question types "site query, equipment parameter consultation", and interaction progress indicator "not yet ready for use".

[0051] In some embodiments, the server first extracts charging service entities from historical dialogue data and charging query data, respectively. For each extracted charging service entity, the server calculates the similarity between the stage information of the charging service entity and the current service stage to obtain stage-related metadata. The server obtains the entity details information of the charging service entity and extracts the source metadata and timeliness metadata from the entity details information. Based on each charging service entity, the stage-related metadata, source metadata, and timeliness metadata of each charging service entity, the server integrates them according to a preset structured format (entity identifier - entity attribute - source metadata - timeliness metadata - stage-related metadata) to form a context anchor. The specific steps of the above-mentioned charging service entity extraction include: calling a predefined charging service entity dictionary; based on the predefined charging service entity dictionary, performing keyword matching and entity type labeling on historical dialogue data and charging query data, while filtering out redundant information without clear semantic reference, to obtain at least one charging service entity. A charging service entity refers to an entity that provides charging services.

[0052] The construction of context anchors is explained below: First, based on a predefined charging service entity dictionary, charging service entities are extracted from historical dialogue data and charging query data, respectively. The extracted charging service entities include: "X Park Charging Station" (charging station entity), "Charging Pile" (charging pile equipment entity), "Standard Interface" (equipment parameter entity), and "Fault Code E2" (fault code entity). Then, metadata is added to each charging service entity: For the charging service entity "X Park Charging Station": the source metadata is historical dialogue data, the timeliness metadata is valid in the current query dialogue (time range t1-current time), and the stage association metadata is obtained by calculating the similarity between the stage information of the charging service entity and the core features of the "pre-charging consultation stage" (site query, reservation). For example, the stage association metadata has a similarity of 0.95; For the charging service entity "Charging Pile": the source metadata is charging query data, the timeliness metadata is valid within the current query dialogue, and the stage association metadata has a similarity of 0.90.

[0053] Furthermore, the context anchors obtained by integrating them in a structured format include: 1. Entity identifier: S1, entity attribute: X Park charging station (location: within 3 kilometers of the target account's current location, pile type: DC pile); source metadata: historical dialogue data, timeliness: t1-current time, stage-related metadata: similarity 0.95; 2. Entity identifier: D1, entity attribute: charging pile (type: DC pile), source metadata: historical dialogue data + current original query data, timeliness source information: t1-current time, stage-related metadata: similarity 0.90.

[0054] Step S206: Based on context anchors, perform semantic extraction on the charging query data to obtain the original semantic elements. Call the charging service knowledge graph and adjust the basic prompt word template corresponding to the current service stage based on context anchors to obtain the target prompt words.

[0055] In this context, the original semantic elements are the most basic semantic units that constitute the original demand expression of the target object in the charging query data. In some embodiments, the original semantic elements include explicit target keywords and implicit target referential cues. Explicit keywords refer to words or phrases in the charging query data that directly reflect the demand of the target object and are directly identifiable. Implicit referential cues refer to words or phrases in the charging query data that indirectly point to a specific charging service entity and whose semantic reference can only be clarified by combining contextual anchors (such as pronouns, ellipsis, etc.).

[0056] In some embodiments, a semantic extraction model is invoked to perform semantic extraction based on contextual anchors and charging query data to obtain the original semantic elements.

[0057] In some embodiments, semantic extraction is performed on charging query data based on context anchors to obtain original semantic elements, including: identifying synonymous semantic boundaries in the charging query data based on context anchors to divide the charging query data into multiple structured semantic fragment groups; obtaining the business semantic domain corresponding to the current service stage; performing lexical matching under domain constraints on each structured semantic fragment group and the business semantic domain to obtain target explicit keywords matching the current service stage; constructing a textual reference mapping table based on each structured semantic fragment group and context anchors; and determining the original semantic elements based on the textual reference mapping table and the target explicit keywords.

[0058] Synchronous semantic boundary recognition refers to simultaneously performing semantic segmentation on text and image data, ensuring that the segmented text semantic units and image semantic units are semantically related. The segmentation process uses the attributes and meta-information of the charging service entity in the context anchor as a semantic reference. A structured semantic fragment group consists of a set of semantically related text structured semantic fragments (text semantic units) and image structured semantic fragments (image semantic units). A structured semantic fragment refers to the smallest semantic unit with a clear semantic boundary and independent semantic direction; each structured semantic fragment contains only one text clause or one image local region. A text clause is the smallest sentence unit in text data with a complete grammatical structure and independent semantics; an image local region is the smallest image region in image data containing a single core visual object. The business semantic domain refers to the semantic set defined by the scope of the charging service business corresponding to the current service stage. This business semantic domain is constructed based on the core theme keyword set and typical object question type set of the current service stage in a predefined feature library of the entire charging service process. The business semantic domain clearly defines the scope boundary of effective semantics under the current service stage.

[0059] Optionally, the server extracts semantic features of the core charging service entity from context anchors. Then, it performs word segmentation and syntactic analysis on the text statements in the charging query data, identifying subject-verb-object structures and semantic pauses. Based on the semantic matching results between semantic pauses and the core charging service entity, the server determines the text semantic boundary. Based on this boundary, the text statements are segmented to obtain text semantic units. The server performs object detection on the image data in the charging query data, identifying the core visual objects. Based on the matching results between the edge contours of the core visual objects and the visual features of the core charging service entity in the context anchors, the server determines the image semantic boundary. Based on this boundary, the image data is segmented to obtain image semantic units. For each text semantic unit, the server calculates the semantic similarity between the text semantic unit and each image semantic unit (using a semantic vector similarity-based calculation method). Image semantic units with semantic similarity greater than a semantic threshold are selected. Each selected image semantic unit is then associated and combined with the text semantic unit to form a structured semantic fragment group.

[0060] Optionally, after obtaining the business semantic domain defined by the current service stage, for each structured semantic fragment group, the server matches the text semantic units or image semantic units in the structured semantic fragment group with the effective semantic words in the business semantic domain, uses the matched effective semantic words as candidate explicit keywords, calculates the semantic relevance between the candidate explicit keywords and the charging service entity in the context anchor, and selects the target keyword from the candidate explicit keywords based on the semantic relevance of each candidate explicit keyword.

[0061] For example, based on the current service stage, a set of core theme keywords and a set of typical target object problem types are extracted from a predefined feature library of the entire charging service process to obtain the defined business semantic domain.

[0062] For example, for each candidate explicit keyword, the semantic relevance between the candidate explicit keyword and each charging service entity in the context anchor is calculated. Then, if at least one semantic relevance is greater than or equal to a relevance threshold (e.g., 0.6), the candidate explicit keyword is used as the target keyword. Alternatively, the semantic relevance between the candidate explicit keyword and the core charging service entity in the context anchor is calculated; if the semantic relevance is greater than or equal to the relevance threshold, the candidate explicit keyword is used as the target explicit keyword. Here, the core charging service entity can be a charging service entity derived from charging query data.

[0063] For example, the semantic relevance between candidate explicit keywords and each charging service entity can be calculated using the following formula (1):

[0064] (1)

[0065] in, Candidate explicit keywords With charging service entities Semantic relevance; Candidate explicit keywords semantic set: For charging service entities A semantic set; The number of elements in the intersection of two semantic sets This represents the number of elements in the union of two semantic sets.

[0066] It should be noted that the formula (1) is designed for the semantic matching requirements of keywords and entities in the charging service scenario. It focuses on the core semantic overlap, avoids interference from redundant features, and takes into account both universality and domain adaptability. It is applicable to text keyword matching and can also be extended to the association calculation of simple visual words and entities.

[0067] Optionally, the server constructs a mapping relationship between implicit referents in structured semantic segments and their antecedents in historical dialogues based on the historical dialogue referential context in the structured semantic segment group and the context anchor. This results in a textual referential mapping table. The historical dialogue referential context refers to the charging service entity and its associated historical interaction statements in the context anchor, whose source metadata is the historical dialogue data. The historical dialogue referential context includes the referential expressions, antecedents, and their relationships appearing in the historical dialogue data.

[0068] For example, the server extracts historical dialogue referential context from context anchors to construct a historical referential corpus; it traverses all text semantic units in the structured semantic fragment group to identify implicit referents; for each implicit referent, it extracts candidate antecedents from the historical referential corpus (candidate antecedents must be consistent with the implicit referent in terms of grammatical attributes and semantic categories); by calculating the semantic matching degree between the implicit referent and the candidate antecedent (calculated using the cosine similarity formula), it selects the candidate antecedent with the highest semantic matching degree. If the semantic matching degree of the selected candidate antecedent is greater than or equal to the semantic matching degree threshold, the selected candidate antecedent is used as the target antecedent; finally, the implicit referents, target antecedents, and related attribute information are organized into a textual referential mapping table. Implicit referents refer to words in a structured semantic fragment group that do not explicitly point to a specific charging service entity, but whose semantic reference can be inferred from the context. These mainly include pronouns and ellipsis. Antecedents refer to explicit charging service entities or specific expressions in historical dialogue data that have a semantic relationship with implicit referents. The textual reference mapping table refers to structured data that records the correspondence between implicit referents and antecedents in tabular form. Its core fields include implicit referents, antecedents, source of antecedents (specific timestamps in historical dialogues), and semantic reference credibility. Semantic reference credibility refers to the degree of semantic matching between implicit referents and their corresponding antecedents, represented by a value in the range of 0-1.

[0069] Optionally, the server performs cross-modal alignment of visual objects in image modal semantic segments and reference expressions in text modal semantic segments based on multimodal historical interaction information recorded in structured semantic segment groups, text reference mapping tables, and context anchors, to obtain cross-modal reference mapping relationships. Here, image modal semantic segments refer to segments in the structured semantic segment group whose modality is a local region of an image, and can also be considered as image semantic units based on the structured semantic segment group; text modal semantic segments refer to segments in the structured semantic segment group whose modality is a text clause, and can also be considered as text semantic units based on the structured semantic segment group.

[0070] For example, visual features (such as shape, color, texture, positional relationship, etc.) of visual objects are extracted from image modal semantic fragments, while the referential expressions in text modal semantic fragments are parsed as the charging service entity they point to (i.e., antecedents). Then, typical visual feature descriptions corresponding to the antecedents in historical interactions are extracted from the multimodal historical interaction information recorded by context anchors, and used as their standard visual feature templates. The feature matching degree is obtained by calculating the similarity between the visual features of the visual object and the standard visual feature template. Visual objects with feature matching degrees greater than the feature matching threshold are associated with their corresponding referential expressions, and the alignment confidence degree (alignment confidence degree = feature matching degree × semantic pointing confidence degree) is calculated by combining the semantic pointing confidence degree in the text referential mapping table, thus forming a cross-modal referential mapping relationship.

[0071] Among them, visual objects refer to the core visual elements contained in image modal semantic fragments, possessing clear visual features and semantic orientations; referential expressions refer to explicit pronouns or implicit referents contained in text modal semantic fragments; multimodal historical interaction information refers to the interaction information in the context anchors whose source metadata is historical dialogue data, containing both text and image modalities (if there is no image interaction in the historical dialogue data, it only contains text interaction information and corresponding visual feature descriptions); cross-modal referential mapping relationship refers to the structured data that records the semantic association between visual objects and referential expressions, including fields such as visual object identifier, referential expression content, associated entity, alignment confidence, etc. Alignment confidence refers to the degree of confidence that the visual object and the referential expression point to the same entity, represented by a value in the range of 0-1.

[0072] The feature matching degree mentioned above can be calculated based on the following formula (2):

[0073] (2)

[0074] in, For visual objects with antecedent Feature matching degree, To normalize the feature difference, the normalized feature difference is calculated using the formula (3) below. The normalized feature difference can eliminate the influence of the dimensions of different visual features, making the matching results more objective.

[0075] (3)

[0076] Where n is the visual feature dimension (such as shape, color, position, etc., for example, n can be set to 5); Let be the normalized value of the visual feature of visual object v in the i-th dimension (range 0-1). , is the normalized value of the antecedent a in the i-th dimension of visual feature description (range 0-1); express and The maximum value in the formula (3) can solve the problems of inconsistent visual feature dimensions and large differences in numerical range for the feature matching requirements of "visual object-antecedent" in cross-modal scenarios. It can also combine the visual feature characteristics of charging service entities (such as charging pile interface, fault code display, etc.) and avoid the interference of extreme values ​​on the matching results by normalizing the maximum value.

[0077] Optionally, the server merges the textual reference mapping table, cross-modal reference mapping relationship, and target explicit keywords to obtain the original semantic elements.

[0078] For example, the server performs deduplication on the target explicit keywords, retaining only the unique target explicit keywords; then, it extracts implicit referents and their corresponding antecedents from the textual reference mapping table, combining the two into textual implicit reference cues; next, it extracts the referent expression, its corresponding visual object, and associated entity from the cross-modal reference mapping relationship, combining the three into cross-modal implicit reference cues; then, the server determines whether there is semantic overlap (i.e., pointing to the same associated entity) between the textual implicit reference cues and the cross-modal implicit reference cues; if there is overlap, it retains the cues with higher semantic pointing credibility and alignment credibility, obtaining the organized implicit reference cues; if not, it combines the textual implicit reference cues and the cross-modal implicit reference cues to obtain the organized implicit reference cues, i.e., the target implicit reference cues; finally, it integrates the deduplicated target explicit keywords with the organized implicit reference cues (target implicit reference cues) to form the original semantic elements. Among them, the sorted implicit reference clues can be understood as the set of elements obtained after merging, which can only be clearly understood in conjunction with the context. They include the implicit reference items in the text reference mapping table and their corresponding antecedents, the reference expressions in the cross-modal reference mapping relationship and their corresponding visual objects and related entities.

[0079] In the above embodiments, by performing synonymous semantic boundary recognition on the charging query data, the charging query data can be divided into different groups of structured semantic fragments, ensuring that the structured semantic fragment groups obtained from the segmentation have semantic relevance, thus solving the semantic misalignment problem caused by asynchronous semantic segmentation of multimodal data. Based on the business semantic domain, word matching is performed on each group of structured semantic fragments to obtain target explicit keywords. Based on the constructed text reference mapping table and the target explicit keywords, complete original semantic elements can be obtained, providing accurate and complete semantic input for subsequent semantic inference and improving the reliability of the overall semantic understanding solution.

[0080] In some embodiments, the charging service knowledge graph is invoked, and the basic prompt word template corresponding to the current service stage is adjusted based on context anchors to obtain the target prompt word. This includes: determining multiple matching candidate intent nodes from the charging service knowledge graph based on context anchors, and selecting the target intent node from the multiple candidate intent nodes based on the degree of matching between each candidate intent node and the current service stage; filling the basic prompt word template corresponding to the current service stage based on the semantic attributes carried by the target intent node to obtain candidate prompt words; and determining the target prompt word based on the candidate prompt words and the stage boundary conditions corresponding to the current service stage.

[0081] The charging service knowledge graph includes a pre-defined intent node layer, which consists of multiple intent nodes. The intent node layer is based on the core needs of objects in the entire charging service process. Each intent node has a unique node identifier, a clear semantic description, a corresponding service stage adaptation range, and a set of associated entity types.

[0082] Optionally, the server extracts the entity type, core semantic features, and stage-related metadata of all charging service entities from the context anchor to form an entity association basic data set; then, it extracts all intent nodes from the intent node layer of the charging service knowledge graph, filters out intent nodes whose adaptation scope includes the current service stage, and obtains an initial intent node set, which includes multiple intent nodes. For each intent node in the initial intent node set, the server calculates the bidirectional semantic association degree between the intent node and each charging service entity in the entity association basic data set. The bidirectional semantic association degree includes two dimensions: entity type matching degree and semantic feature fit degree; then, based on the pre-set comprehensive threshold of the bidirectional semantic association degree, it filters out intent nodes whose bidirectional semantic association degree is greater than or equal to the comprehensive threshold, and uses the filtered intent nodes as candidate intent nodes. The calculation of the bidirectional semantic association degree can be referred to formula (4):

[0083] (4)

[0084] in, Intent node With charging service entities The bidirectional semantic correlation degree, with a value range of 0-1; This is the weighting coefficient, which can be 0.4; Intent node The set of related entity types and the charging service entity The matching degree of the entity type, if The entity type belongs to The set of related entity types, then The value is 1 if it is set to 1, otherwise the value is 0. Intent node Semantic description and charging service entity The degree of fit of the core semantic features is obtained by calculating the cosine similarity between the semantic vectors of the two, with a value ranging from 0 to 1.

[0085] It should be noted that in Formula (4), the logic of type matching as the basis and semantic fit as the core is clarified for the association filtering requirements of intent nodes-entities in the charging service knowledge graph. Considering that in the charging service scenario, the incorrect entity type will directly lead to semantic understanding deviation, TypMat is used to ensure basic effectiveness, and SemFit is used to improve the association accuracy.

[0086] Optionally, after determining multiple candidate intent nodes, the server extracts the core demand weight distribution of the current service stage from a preset feature library of charging service stages. This core demand weight distribution refers to the proportion of importance of different types of core demands in the current service stage. Taking the pre-charging consultation stage as an example, its core demands include equipment usage consultation (weight 0.35), site information consultation (weight 0.3), fault consultation (weight 0.25), and reservation service (weight 0.1). For each candidate intent node, the server obtains the highest core demand weight from the core demand weight distribution of the current service stage, and uses the product between the highest core demand weight and the candidate intent node as the comprehensive relevance score of the candidate intent node. A preset number of candidate intent nodes with the highest comprehensive relevance scores are selected as the target intent nodes.

[0087] Optionally, the server obtains the basic prompt word template corresponding to the current service stage, identifies the replaceable slots in the basic prompt word template, and determines the type identifier and adapted information format of each replaceable slot. For each target intent node, according to the matching rules between the intent node and the replaceable slots, based on the semantic attributes of each target intent node, the server determines the matching replaceable slots and fills the semantic attributes carried by the target intent node into the matching replaceable slots to obtain the intermediate prompt word corresponding to the target intent node. The semantic attributes carried by the intent node refer to the inherent feature information of the node used to define the direction of semantic parsing. Semantic attributes include core parsing dimensions (such as parameter query dimensions, fault cause dimensions, site location dimensions, etc.), required fields for associated entities (i.e., the set of entity types that the intent node must be associated with), and semantic relationship type requirements (such as adaptation relationship, belonging relationship, fault correspondence relationship, etc.). Replaceable slots refer to placeholders reserved in the basic prompt word template for filling in personalized information related to the intent node. Each replaceable slot has a clear slot type identifier (such as entity type slot, semantic relationship type slot, parsing dimension slot, etc.). The matching rules between intent nodes and replaceable slots indicate the slots matched by different semantic attributes. For example, it shows that the core parsing dimension corresponds to the parsing dimension slot, the required field for associated entities corresponds to the entity type slot, and the semantic relationship type requirement corresponds to the semantic relationship type slot. Optionally, after determining the intermediate prompt words, the server extracts all entity elements from the context anchors to obtain a context constraint set. This context constraint set includes entity type constraints (the entity types existing in the current query dialogue), entity attribute constraints (the specific attribute information of the entity, such as interface standards, fault code values, etc.), and stage association constraints (the stage association meta-information of the entity is greater than or equal to the similarity threshold, for example, the similarity threshold is 0.8). For each intermediate prompt word, the entity type, semantic relationship type, and parsing dimension in the intermediate prompt word are parsed to obtain prompt word verification information. The adaptability of the prompt word verification information with the context constraint set is verified to obtain the verification result. Based on the verification results of each intermediate prompt word, the intermediate prompt words that pass the verification are identified and used as candidate prompt words.

[0088] For example, the specific steps for verifying the compatibility of the prompt word verification information with the context constraint set include: based on the prompt word verification information and the context constraint set, performing entity type existence verification (whether the entity type in the prompt word verification information is in the context entity type constraint of the context constraint set), semantic relationship validity verification (whether the semantic relationship type in the prompt word verification information exists in the association relationship of the context entities in the context constraint set), and parsing dimension relevance verification (whether the parsing dimension of the prompt word verification information is related to the context entity attributes of the context constraint set); if the entity type in the prompt word verification information is in the context entity type constraint of the context constraint set, and the semantic relationship type in the prompt word verification information exists in the association relationship of the context entities in the context constraint set, and the parsing dimension of the prompt word verification information is related to the context entity attributes of the context constraint set, the verification result is determined to be passed.

[0089] Optionally, the server selects the target prompt word from at least one candidate prompt word based on the stage boundary conditions corresponding to the current service stage. The stage boundary conditions refer to a set of pre-defined constraint rules for the current service stage that define the scope of service actions. These constraints include action permission constraints (the types of service actions allowed in the current service stage), entity expiration constraints (only entities valid within the current query dialog are allowed to be associated), and demand association constraints (the service action chain must be directly related to the target object identifier input demand). The service action path refers to a set of standardized process steps recorded in the charging service knowledge graph as directed edges, corresponding to each intent node of the charging service. Each service action path contains several service action nodes arranged in logical order, and each service action node has attributes such as a unique identifier, action description, input parameter requirements, and output result definition.

[0090] In the above embodiments, target intent nodes are filtered from the charging service knowledge graph through context anchors and the current service stage to accurately identify the intent. Then, based on the semantic attributes and stage boundary conditions of the target intent nodes, target prompt words are accurately queried, ensuring that the semantic guidance direction of the target prompt words is consistent with the object's true intent, avoiding parsing deviations caused by generalized prompt words, and strengthening the semantic parsing boundary of context constraints to ensure the accuracy of subsequent semantic understanding.

[0091] In some embodiments, determining the target prompt word based on candidate prompt words and the stage boundary conditions corresponding to the current service stage includes: extracting multiple service action paths from the charging service knowledge graph; for each candidate prompt word, filtering service action paths that match the candidate prompt word from the multiple service action paths based on the degree of matching between the candidate prompt word and each service action path; and performing stage constraint filtering on the multiple candidate prompt words based on the service action paths matched by each candidate prompt word and the stage boundary conditions corresponding to the current service stage to obtain the target prompt word.

[0092] Optionally, the server extracts multiple service action paths from the charging service knowledge graph. Each service action path is associated with corresponding intent node identifiers, core parsing dimensions, associated entity types, and other index information. For each candidate prompt word, the server parses the prompt word to obtain the parsing dimension, associated entity type, and semantic relationship type. Based on the parsing dimension, associated entity type, and semantic relationship type, the server forms a prompt word feature vector for the candidate prompt word. The server calculates the semantic similarity between the prompt word feature vector and each service action path to obtain the matching degree between the candidate prompt word and each service action path. The service action path with the highest matching degree is taken as the matching service action path for the candidate prompt word. For each candidate prompt word, the server extracts the service action nodes in the matching service action path in logical order to form the service action chain corresponding to the candidate prompt word. The starting action of the service action chain is the first node of the matching service action path, and the ending action is the last node of the matching service action path. The intermediate actions are arranged sequentially according to the path order, and each candidate prompt word corresponds to a unique service action chain.

[0093] For example, for each candidate suggestion word, the semantic similarity calculation steps between the candidate suggestion word and each service action path can refer to the following formula (5):

[0094] (5)

[0095] in, The semantic similarity between candidate prompt word t and service action path p is , with a value ranging from 0 to 1; i=1,2,3 correspond to the three matching dimensions of parsing dimension, associated entity type, and semantic relationship type, respectively. It is the semantic similarity between candidate prompt word t and service action path p on matching dimension i; The semantic similarity between candidate suggestion word t and service action path p in the parsing dimension; The semantic similarity between candidate suggestion word t and service action path p in terms of associated entity type; The semantic similarity between candidate suggestion word t and service action path p in terms of semantic relationship type; The value range is from 0 to 1; Formula (5) is based on the normalization calculation of Euclidean distance, taking into account the differences of the three matching dimensions (parsing dimension, associated entity type, and semantic relationship type), and solves the path selection bias problem caused by single-dimensional matching.

[0096] Optionally, stage constraint filtering is performed based on the service action chain of each candidate prompt and the stage boundary conditions of the current service stage to obtain candidate prompts that satisfy the current stage boundary conditions. For example, each stage boundary condition can be decomposed into verifiable constraint rule entries, each rule entry containing a rule type, verification standard, and judgment method (pass / fail). Therefore, for each candidate prompt, the server verifies each service action node in the service action chain of that candidate prompt according to each constraint rule entry of the stage boundary conditions. If all service action nodes in the service action chain satisfy all constraint rule entries, the candidate prompt corresponding to that service action chain is a valid prompt; if at least one service action node does not satisfy at least one constraint rule entry, the candidate prompt is eliminated. All valid prompts are integrated to form a set of valid prompts that satisfy the current stage boundary conditions. For example, in some embodiments, candidate prompts that satisfy the current stage boundary conditions can be directly used as target candidate prompts, without the need for further verification of candidate prompts that satisfy the current stage boundary conditions described below.

[0097] Of course, in other embodiments, after determining the candidate prompts that meet the boundary conditions of the current stage, the method further includes: the server determining the intent node identifier of the target intent node corresponding to each candidate prompt in the set of valid prompts, and querying the index position (sorting position) of each intent node identifier in the intent priority sequence. The candidate prompts in the set of valid prompts are then sorted in ascending order of index position to form a sequence of valid prompts. The index order of this intent priority sequence refers to the position number of each intent node in the intent priority sequence (incrementing from 1), with smaller position numbers indicating higher priority. Each candidate prompt in the set of valid prompts is associated with a unique intent node; therefore, the sorting priority of the prompts can be determined by the index position of the intent node.

[0098] Optionally, the server verifies the obtained valid prompt word sequence in conjunction with the structural integrity constraints of the basic prompt word template. These structural integrity constraints refer to the core structural elements that the basic prompt word template must include, such as service stage identifiers, explicit descriptions of associated entities, clear definitions of semantic relationships, complete coverage of parsing dimensions, and clear demand guidance. These core structural elements ensure that the prompt words can comprehensively and accurately guide the semantic deduction process. For example, the server first determines the structural integrity constraint elements of the basic prompt word template, lists the requirements for each constraint element, integrates all candidate prompt words in the valid prompt word sequence in sorted order to form initial integrated prompt words, and then verifies whether the initial integrated prompt words meet all requirements by comparing them with the structural integrity constraint elements. If the initial integrated prompt words meet all structural integrity constraint requirements, all candidate prompt words in the valid prompt word sequence are directly used as target prompt words. If any elements are missing or unclear, they are supplemented and improved based on entity attributes in the context anchor and semantic relationships in the charging service knowledge graph until all structural integrity constraints are met, forming the final target prompt words.

[0099] In the above embodiments, candidate prompts are matched through service action paths for initial screening, and then secondary screening is performed using stage boundary conditions. This ensures that the final target prompts are complete and valid, thereby avoiding the problem of semantic ambiguity caused by missing target prompt information and further ensuring the accuracy and effectiveness of semantic understanding.

[0100] Step S208: Based on the target prompt words and original semantic elements, the target semantic parsing result of the charging query data is obtained through semantic inference processing. The target semantic parsing result is used to determine the reply information about the charging query data and send the reply information through the charging customer service account.

[0101] In some embodiments, the original semantic elements include target explicit keywords and target implicit referential clues. Based on the target prompts and the original semantic elements, the target semantic parsing result of the charging query data is obtained through semantic inference processing, including: matching the target explicit keywords with entity nodes in the charging service knowledge graph to obtain an initial entity matching group; constructing a complete entity group based on the target implicit referential clues and context anchors; merging the initial entity matching group and the complete entity group to obtain a fused entity group; and determining the target semantic parsing result of the charging query data based on the fused entity group and the target prompts.

[0102] Optionally, the server extracts all target explicit keywords from the original semantic elements. For each target explicit keyword, the server calculates the semantic fit between the target explicit keyword and the semantic description of each entity node in the entity node library, and filters out entity nodes with a semantic fit greater than a fit threshold. Based on the filtered entity nodes and the target explicit keyword, a keyword-entity node association combination is formed. The server integrates all association combinations to obtain an initial entity matching group. The fit threshold can be set to 0.8.

[0103] For example, semantic fit can be referred to the following formula (6):

[0104] (6)

[0105] in, In this embodiment, it refers to the semantic fit between the target explicit keyword k and the entity node n, with a value range of 0 to 1. It is the semantic vector of the target explicit keyword k; It is a semantic vector that describes the semantics of entity node n; semantic vector The length of the mold, It is a semantic vector The modulus. β is the attribute association weight factor, with a preset value of 0.3; AttrRel(k,n) is the association degree between the explicit keyword k and the core attribute of the entity node n, with a value range of 0-1; where, if the target explicit keyword k contains all the core attributes of the entity node n, the corresponding association degree is 1; if the target explicit keyword k contains some of the core attributes of the entity node n, the corresponding association degree is 0.5; if the target explicit keyword k does not contain the core attributes of the entity node n, the corresponding association degree is 0. It should be noted that this formula (6) captures the overall semantic association and highlights the matching value of key attributes by integrating semantic vector similarity and core attribute association, and adjusts the influence of attribute association through β weight, which can be dynamically adjusted according to the attribute importance of the charging service entity to adapt to the structured attribute characteristics of the charging service entity.

[0106] For example, the original semantic elements include the target explicit keywords "charging pile" and "fault code E2". The core entity nodes in the pre-built charging service knowledge graph entity node library include: N1: Node ID = D1, Semantic Description = DC charging pile, Core Attribute = {Type: DC, Interface Standard: Interface Standard 1, Function: Start charging, Applicable Models: Most new energy passenger vehicles}; N2: Node ID = F1, Semantic Description = Fault Code E2, Core Attribute = {Represents Fault: Start-up fault, Triggering Scenario: Pre-charging start-up stage, Associated Device: DC charging pile}; N3: Node ID = S1, Semantic Description = X Park charging station, Core Attribute = {Location: Within 3 kilometers of the target object's current location, Included Device: DC charging pile, Pile Type Distribution: 2 idle DC charging piles}; N4: Node ID = P1, Semantic Description = Interface Standard 1, Core Attribute = {Applicable Device: DC charging pile, Applicable Models: Most new energy passenger vehicles}; N5: Node ID = Ft1, Semantic Description = Start-up fault, Core Attribute = {Fault Type: Start-up class, Associated Fault Code: E2, Involved Device: Charging pile}.

[0107] For the target explicit keyword "charging pile", this "charging pile" and N1 (DC charging pile) are: , , (Including the core attribute "charging station type") This "charging station" is different from other nodes (N2-N5). All are less than 0.8.

[0108] For the target explicit keyword "fault code E2", the "fault code E2" and N2 (DC charging pile) are: , , (Including the core attribute "charging station type") This "charging station" is different from other nodes (N1, N3-N5). All are less than 0.8.

[0109] Therefore, it can be known that the initial entity matching group is: {(charging pile, D1-DC charging pile), (fault code E2, F1-fault code E2)}.

[0110] Optionally, the server performs referential resolution mapping based on the target implicit referential clues in the original semantic elements and the anchored context fragments in the context anchors to obtain an implicit referential mapping group; based on each mapping item in the implicit referential mapping group and combined with the charging service knowledge graph, the server performs referential entity completion to obtain a completed entity group.

[0111] For example, the reference type is classified and determined based on the target implicit reference clues in the original semantic elements, resulting in pure text reference groups and cross-modal reference groups. For example: Extract all target implicit referential cues from the original semantic elements to construct an implicit referential cue set, and label each target implicit referential cue in the implicit referential cue set with its source modality (text modality / image modality / multimodality) and constituent elements; for each target implicit referential cue in the implicit referential cue set, if the target implicit referential cue only contains text modality elements, then the target implicit referential cue is determined to belong to the pure textual referential type; if the target implicit referential cue simultaneously contains a text modality referential expression and an image modality visual object, and the two have a clear semantic relationship (i.e., the textual referential expression points to the image visual object), then the target implicit referential cue is determined to belong to the cross-modal referential type; group the target implicit referential cue belonging to the pure textual referential type in the implicit referential cue set to obtain the pure textual referential group, and group the target implicit referential cue belonging to the cross-modal referential type in the implicit referential cue set to obtain the cross-modal referential group.

[0112] Among them, the pure textual reference group refers to the set of implicit reference clues that are composed of a single textual modality and can be resolved without relying on information from other modalities. The core forms include pronouns in the text (such as "this" and "should"), ellipsis, abbreviations, etc.; the cross-modal reference group refers to the set of implicit reference clues that are semantically related to the reference expression of the textual modality and the visual object of the image modality and require the combination of multimodal information to complete the resolution of reference. The core form is the association pair of "textual reference expression - image visual object".

[0113] For example, candidate entities are extracted based on the plain text reference group and the anchored context fragments in the context anchors according to grammatical roles, resulting in a plain text reference candidate mapping group. Here, the anchored context fragment refers to the set of textual information contained in the context anchor that is related to the plain text reference clue, including the charging service entity mentioned in the historical dialogue, the entity's grammatical attributes (such as singular / plural, noun / verb attributes), the grammatical relationship between the entity and the text statement (such as subject / object), the entity's mention frequency and timeliness information, etc.; grammatical role alignment refers to matching the grammatical attributes (such as part of speech, singular / plural, grammatical function) of the plain text reference clue with the grammatical attributes and grammatical relationships of the charging service entity in the anchored context fragment, ensuring that the two are grammatically compatible.

[0114] In view of this, the specific steps for determining the candidate mapping group of plain text references are as follows:

[0115] First, analyze each target implicit reference clue in the plain text reference group and extract its grammatical attributes (including part of speech (pronoun / noun / abbreviation), singular / plural form, and grammatical function (subject / object / modifier)) to form a set of plain text reference grammatical features. Then, extract the anchored context fragments from the context anchors, filter out the charging service entities, extract the grammatical attributes of each entity (such as part of speech and singular / plural attributes corresponding to the entity type) and the grammatical role in the historical dialogue (such as appearing as subject / object), and construct an entity grammatical feature library.

[0116] Then, the pre-set grammatical role alignment rules are obtained, which indicate that: the part-of-speech matching of plain text reference clues and entities (e.g., pronouns correspond to noun entities), singular and plural consistency (e.g., singular pronouns correspond to singular entities), and grammatical function matching (e.g., attributive references correspond to noun entities). For each target implicit reference clue in the plain text reference group, entities that meet the grammatical role alignment rules are selected from the entity grammatical feature library based on the grammatical role alignment rules, and the selected entities are used as candidate entities; finally, the grammatical alignment confidence between the target implicit reference clue and the candidate entities is calculated using the following formula (7):

[0117] (7)

[0118] in, The implicit reference to the target, clue h, and candidate entities. The grammatical alignment reliability ranges from 0 to 1; j=1,2,3 correspond to the three dimensions of part-of-speech matching, singular / plural consistency, and grammatical function matching, respectively. Preset weights for each dimension =0.4、 =0.3、 =0.3; Let be the fit score for the j-th dimension, with 1.0 for a perfect fit and 0.0 for a partial fit. The final implementation decomposes the calculation by grammatical dimension, ensuring clear logic and clearly defining the contribution of each dimension (part of speech, singular / plural, grammatical function) to the alignment result, while also ensuring the rigor of the alignment. Then, candidate entities with grammatical alignment confidence scores greater than a confidence threshold are selected, and association mappings are performed based on the selected candidate entities and their target implicit referential clues. The association mappings of each target implicit referential clue in the plain text referential group are integrated to obtain the plain text referential candidate mapping group.

[0119] For example, modal anchor binding is performed based on cross-modal reference groups, anchored context fragments in context anchors, and a preset cross-modal reference resolution rule base to obtain cross-modal reference anchor entity groups. Here, modal anchor binding refers to semantically binding the "textual reference expression - image visual object" association pair in the cross-modal reference clue with the charging service entity in the context anchor, explicitly defining the specific entity that the association pair points to.

[0120] The following details the steps for determining cross-modal reference anchored entity groups:

[0121] Since each target implicit reference cue in the cross-modal reference group contains both the text modality's reference expression and the image modality's visual object, we parse each target implicit reference cue in the cross-modal reference group to obtain the corresponding "text reference expression - image visual object" association pair, where the text reference expression is the text modality's reference expression and the image visual object is the image modality's visual object. Then, for each association pair, the grammatical features (part of speech, singular / plural, semantic reference) of the text pointer and the visual features (shape, color, position, core identifiers, etc.) of the image visual object are extracted to form a cross-modal feature set; charging service entities are extracted from the anchored context fragments of the context anchors to build an entity library, where each entity contains semantic features (entity type, core attributes) and visual feature descriptions (if any); the cross-modal reference resolution rule library is called to perform binding according to the following rules: 1. Modal feature matching rule: the grammatical features of the text pointer are aligned with the grammatical attributes of the entity, and the visual features of the image visual object are matched with the visual feature descriptions of the entity; 2. Semantic association verification rule: the semantic similarity between the text pointer and the entity (using the residual...) The similarity of the chords is greater than or equal to the text similarity threshold to ensure that the semantics of the text representation are highly matched with the semantics of the knowledge graph entity, and to avoid cross-modal representation resolution errors due to semantic deviation; the semantic correlation between the image visual object and the entity is greater than or equal to the image similarity threshold to avoid invalid binding where the visual and semantics are completely disconnected; 3. Entity adaptation judgment rules: the semantic category of the entity is consistent with the semantic orientation of the cross-modal representation clue (the cross-modal representation clue refers to the target implicit representation clue in the cross-modal representation group); then, for each cross-modal representation clue, the entity that satisfies all the rules is selected as the anchor entity; then, the following formula (8) is used to calculate the semantic correlation between the image visual object corresponding to each cross-modal representation clue and each entity:

[0122] (8)

[0123] in, For image visual objects v and entities The semantic relevance is 0-1; m is the visual feature dimension, with a default value of m=4 (shape, color, position, core identifier). Let be the normalized value (0-1) of the k-th dimension visual feature of the image visual object v; For entities The normalized value (0-1) of the k-th dimension visual feature description. Formula (8) adopts an average weighted model, which is simple and efficient to calculate. It is suitable for the rapid integration of multi-dimensional visual features and can solve the problem of unclear semantic association of multi-modal information for cross-modal alignment requirements of image visual objects and text entities.

[0124] Then, entities with semantic relevance greater than the relevance threshold are selected, and the selected entities are associated with the cross-modal reference clue. Based on the association mapping of each cross-modal reference clue, that is, based on the association mapping of each target implicit reference clue in the cross-modal reference group, the cross-modal reference anchoring entity group is obtained.

[0125] Optionally, a unified referencing mapping is performed based on the plain text referencing candidate mapping group and the cross-modal referencing anchoring entity group to obtain an implicit referencing mapping group. For example, each association mapping in the plain text referencing candidate mapping group and each association mapping in the cross-modal referencing anchoring entity group are deduplicated and then integrated to obtain an implicit referencing mapping group.

[0126] Optionally, after determining the implicit referencing mapping group, based on each association mapping in the implicit referencing mapping group, the entities involved in the association mapping are extracted. For each entity, supplementary information such as associated entities, core attributes, and semantic relationships are extracted from the charging service knowledge graph. Based on the supplementary information of each entity, a complete entity group is constructed. For example, for each entity, supplementary information of the entity is queried from the charging service knowledge graph, including the complete attribute set (such as device type, interface standard, location, etc.), directly associated entities (entities directly connected through semantic relationships), and indirectly associated entities (entities associated through a single transit). Then, the queried supplementary information is filtered to obtain the filtered supplementary information: that is, the core attributes are retained to those related to the current service stage, the associated entities are filtered to have an association meta-information value ≥0.8, and the semantic relationships are retained to retain the core relationships directly related to the target entity (such as ownership relationship, adaptation relationship, fault correspondence relationship, etc.). Finally, the entity and the filtered supplementary information are integrated in a structured format (entity identifier - complete attributes - associated entities - semantic relationships) to form the complete entity of the entity. Based on the completion entities of all entities, a complete entity group is obtained.

[0127] Optionally, the server merges the initial entity matching group and the complete entity group to obtain a fused entity group. For example, the server performs semantic association and information integration between the entity nodes in the initial entity matching group and the complete entities in the complete entity group, removes information redundancy, supplements missing information, and forms a fused entity with a unified structure and complete information. For example, the process involves parsing each association combination in the initial entity matching group, extracting entity nodes (including entity identifiers, semantic descriptions, and core attributes), and constructing an initial entity set. Based on each completed entity in the parsed completed entity group, entity identifiers, complete attributes, associated entities, and semantic relationships are extracted to construct a completed entity set. An association mapping is established between the initial entity set and the completed entity set based on entity identifiers, matching entities with the same entity identifier. For successfully matched entities, the core information of the initial entity is integrated with the complete attributes, associated entities, and semantic relationships of the completed entity, while removing duplicate information (such as core attributes already included in the initial entity) and supplementing missing information (such as associated entities and semantic relationships of the completed entity). For entity nodes in the initial entity set that do not match a completed entity, their original information is retained, and basic association information (such as core semantic relationships) is supplemented from the charging service knowledge graph. All integrated entities are then organized in a unified structured format to obtain a fused entity group.

[0128] Optionally, the server determines the target semantic parsing result of the charging query data based on the fused entity group and the target prompt words.

[0129] In the above embodiments, an initial entity matching group is obtained by matching explicit keywords with entity nodes in the charging service knowledge graph; a complete entity group is constructed based on the target implicit referential clues and contextual anchors; the initial entity matching group and the complete entity group are merged to obtain a fused entity group; and the target semantic parsing result of the charging query data is determined based on the fused entity group and the target prompt words. The entire process achieves a precise transformation from the original semantic elements to the target semantic parsing result, and breaks the isolation of single entity information, forming a structured fused entity group containing entity attributes, relationships, and semantic orientations, providing comprehensive and complete semantic support for deep semantic inference; it also solves the semantic ambiguity problem caused by implicit referentials in charging service consultations, improving the accuracy of parsing the deep needs of the target object; and it breaks through the limitations of surface keyword parsing, achieving depth and completeness of semantic understanding, providing accurate and effective service response basis for intelligent charging customer service.

[0130] In some embodiments, determining the target semantic parsing result of charging query data based on the fused entity group and target prompt words includes: starting with each entity in the fused entity group and using the current service stage as a constraint, filtering relationship paths in the charging service knowledge graph to obtain multiple relationship paths; filtering at least one candidate relationship path from the multiple relationship paths based on the semantic coverage of the target prompt words with each relationship path; for each candidate relationship path, verifying the state consistency of the candidate relationship path based on the service state marker in the context anchor to obtain the corresponding verification result; and filtering the target relationship path from the at least one candidate relationship path based on the verification result of each candidate relationship path, semantically integrating the endpoint entity indicated by the target relationship path and the corresponding association relationship to obtain the target semantic parsing result.

[0131] Among them, the relation path refers to the ordered link formed in the charging service knowledge graph, which uses entities as nodes and semantic relations as directed edges, from the starting entity to other related entities. Each path contains the starting entity, several relation predicates, and the ending entity.

[0132] Optionally, the server extracts all entities from the fused entity group, identifies the node identifier of each entity, and uses all entities in the fused entity group as starting entities to construct a starting entity set. From a predefined feature library of the entire charging service process, it extracts a set of core relationship types for the current service stage (e.g., core relationship types for the pre-charging consultation stage: belonging relationship, adaptation relationship, fault correspondence relationship, representation relationship) as constraints for path selection. For each starting entity in the starting entity set, the charging service knowledge graph is traversed to determine multiple undetermined paths originating from that entity. Undetermined paths with a length less than a length threshold are considered relationship paths. Each undetermined path contains at least one directed edge, and each directed edge reflects that the relation predicate belongs to a core relationship type.

[0133] Optionally, for each relationship path, the target prompt words are parsed to extract the set of semantic focus words, and the semantic type (entity class / relation class / parsing dimension class) of each semantic focus word is determined. For each relationship path, the set of relational predicates and the set of core attributes of the endpoint entity contained in the relationship path are extracted. The semantic coverage of the set of semantic focus words with respect to the relationship path is calculated, where the semantic coverage refers to the proportion of the number of focus words that are semantically related to the associated path to the total number of focus words. Relationship paths with semantic coverage exceeding the semantic coverage threshold are selected as candidate relationship paths. It should be noted that verifying whether the semantic coverage exceeds the semantic coverage threshold can be understood as semantic coverage detection, which is used to detect whether the relational predicates or endpoint entity attributes in the relationship path contain semantic focus words, or are highly semantically related to the semantic focus words. Here, semantic focus words refer to the core words that are explicitly specified in the target prompt words and need to be parsed in detail, including core entity names, key relationship descriptions, and parsing dimension keywords (such as "fault correspondence relationship", "representation relationship", "fault cause", "handling solution").

[0134] Service status markers refer to the key service status information recorded in the context anchors under the current service stage, including entity operation status (e.g., charging pile idle / faulty), interaction progress status (e.g., not booked / booked), fault trigger status (e.g., fault triggered / no fault), etc. Each status marker contains a status type, associated entity, and status value. Status consistency verification refers to verifying whether the status information implicit in the focused path is consistent with the status value in the service status markers, and whether there is any logical conflict. Therefore, the specific steps of status consistency verification are as follows: Extract all service status markers from the context anchors, construct a service status marker set, and clarify the status type, associated entity, and status value of each marker; for each candidate relationship path, analyze the path's implicit status information (e.g., if the candidate relationship path contains "fault correspondence," then the path's implicit status information is "entity fault status = triggered"); match the path's implicit status information with the service status marker set to find the status marker corresponding to the path's associated entity; verify whether the path's implicit status is consistent with the status value of the corresponding status marker. If they are consistent, the verification result indicates that the verification passed; if there is no corresponding status marker or the status value conflicts, the verification result indicates that the verification failed.

[0135] Optionally, based on the verification results of each candidate relationship path, candidate relationship paths that have passed verification are selected. Based on the selected candidate relationship paths, a state-consistent path group is determined. Semantic integration is performed on the endpoint entities and their relationships of all state-consistent paths. Semantic integration refers to the analysis of core entities, key semantic relationships, and endpoint entity attributes in the candidate relationship paths within the state-consistent path group. Combined with the parsing requirements of the target prompt words, the core requirements of the object are identified, forming a structured semantic output, thus obtaining the target semantic parsing result. For example, the specific steps for determining the target semantic parsing result are as follows: traverse each candidate relation path in the state-consistent path group, extract the starting entity, ending entity, and all relation predicates in each candidate relation path to obtain the core semantic elements of each candidate relation path, and construct a core semantic element set based on each core semantic element; integrate each core semantic element in the core semantic element set, remove duplicate entities and relations, clarify the hierarchical relationship between entities (such as belonging relationship, causal relationship), and obtain the integrated core semantic element set; determine the entities with related relationships based on the integrated core semantic element set, and regard the entities with related relationships as related entities; based on the parsing dimension requirements of the target prompt words (such as fault cause, solution), supplement the relevant attribute information of the ending entity from the charging service knowledge graph (such as the set of causes of the fault entity, the set of solutions); combine the charging query data of the target object identifier (such as "Why can't the charging pile be started") to deduce the core demand of the target object identifier (i.e., the core problem that the target object identifier wants to solve); organize the core demand, related entities, the relationship of related entities, and the relevant attribute information of each ending entity in a preset structured format to form the target semantic parsing result.

[0136] In the above embodiments, by stage-adaptive path filtering, prompt-focused path matching, and state-consistent path verification, the final output is the target semantic parsing result of semantic integration. This achieves accurate transformation from fused entity groups to structured target semantic parsing results and accurate adaptation of service stages and semantic paths, strengthens the semantic guidance role of target prompt words, and ensures the consistency and integrity of semantic parsing results.

[0137] In the above-mentioned data processing method, by acquiring the charging query data sent by the target account in the current query dialogue and the target account's historical dialogue data, the current query dialogue is a charging query dialogue between the target account and the charging customer service account. Based on the historical dialogue data, the current service stage of the current query dialogue is identified, achieving a deep binding between the dialogue context and the charging service process, and clearly locating the service node of the current charging dialogue. Then, based on the current service stage and historical dialogue data, and combining the dimensions of historical dialogue and the dimensions of the current actual service node, the context anchors can be accurately identified from the charging query data. Based on the context anchors, the charging query data is accurately semantically extracted to obtain the original semantic elements. That is to say, the historical dialogue data with the dimension of historical dialogue can help extract the semantics of the charging query data of the current query dialogue, avoiding isolated parsing based on the current input. Thus, when faced with ambiguous references and omitted expressions, the core demand can be accurately located. Then, the charging service knowledge graph is invoked, and the basic prompt word template corresponding to the current service stage is adjusted based on contextual anchors to obtain highly accurate target prompt words. Based on the target prompt words and original semantic elements, semantic inference processing is used to accurately obtain the target semantic parsing results of the charging query data. The target semantic parsing results are used to determine the response information regarding the charging query data and send the response information through the charging customer service account. This improves the accuracy of semantic understanding.

[0138] In one specific embodiment, the semantic understanding unit deployed on the server performs the following steps:

[0139] First, the semantic understanding unit acquires the charging query data sent by the target account in the current query conversation and the historical conversation data of the target account. The current query conversation is a charging query conversation between the target account and the charging customer service account.

[0140] Secondly, the semantic understanding unit identifies the current service stage of the current query dialogue based on historical dialogue data, and constructs context anchors for the charging query data based on the current service stage and historical dialogue data.

[0141] Next, the semantic understanding unit performs synonym semantic boundary identification on the charging query data based on context anchors to divide the charging query data into multiple structured semantic fragment groups; obtains the business semantic domain corresponding to the current service stage; performs domain-constrained lexical matching on each structured semantic fragment group based on each structured semantic fragment group and the business semantic domain to obtain the target explicit keywords that match the current service stage; constructs a text reference mapping table based on each structured semantic fragment group and context anchors; and determines the original semantic elements based on the text reference mapping table and the target explicit keywords.

[0142] Then, the semantic understanding unit determines multiple matching candidate intent nodes from the charging service knowledge graph based on context anchors, and selects the target intent node from the multiple candidate intent nodes based on the degree of matching between each candidate intent node and the current service stage.

[0143] Furthermore, the semantic understanding unit fills in the basic prompt word template corresponding to the current service stage based on the semantic attributes carried by the target intent node to obtain candidate prompt words; the semantic understanding unit extracts multiple service action paths from the charging service knowledge graph; for each candidate prompt word, based on the degree of matching between the candidate prompt word and each service action path, it selects the service action path that matches the candidate prompt word from the multiple service action paths; based on the service action path matched by each candidate prompt word and the stage boundary conditions corresponding to the current service stage, it performs stage constraint filtering on the multiple candidate prompt words to obtain the target prompt word.

[0144] Finally, the semantic understanding unit extracts the target explicit keywords and target implicit referential clues from the original semantic elements. Based on the explicit keywords, it matches them with entity nodes in the charging service knowledge graph to obtain an initial entity matching group; based on the target implicit referential clues and context anchors, it constructs a complete entity group; merging the initial entity matching group and the complete entity group yields a fused entity group; using each entity in the fused entity group as a starting point and the current service stage as a constraint, it filters relationship paths in the charging service knowledge graph to obtain multiple relationship paths; based on the semantic coverage of the target prompt words with each relationship path, it selects at least one candidate relationship path from the multiple relationship paths; for each candidate relationship path, it verifies the consistency of the candidate relationship path with the service status markers in the context anchors, obtaining the corresponding verification result; based on the verification result of each candidate relationship path, it selects the target relationship path from at least one candidate relationship path, and semantically integrates the endpoint entity indicated by the target relationship path with the corresponding association, obtaining the target semantic parsing result. Based on the target semantic parsing result, the semantic understanding unit determines the reply information regarding the charging query data and sends the reply information through the charging customer service account.

[0145] In the above embodiments, by acquiring the charging query data sent by the target account in the current query dialogue and the target account's historical dialogue data, the current query dialogue is a charging query dialogue between the target account and the charging customer service account. Based on the historical dialogue data, the current service stage of the current query dialogue is identified, achieving a deep binding between the dialogue context and the charging service process, and clearly locating the service node of the current charging dialogue. Furthermore, based on the current service stage and historical dialogue data, and combining the dimensions of historical dialogue and the dimensions of the current actual service node, contextual anchors can be accurately identified from the charging query data. Based on the contextual anchors, the charging query data is accurately semantically extracted to obtain the original semantic elements. That is to say, the historical dialogue data with the dimension of historical dialogue can help extract the semantics of the charging query data of the current query dialogue, avoiding isolated parsing based on the current input. Thus, when faced with ambiguous references and omitted expressions, the core demand can be accurately located. Then, the charging service knowledge graph is invoked, and the basic prompt word template corresponding to the current service stage is adjusted based on contextual anchors to obtain highly accurate target prompt words. Based on the target prompt words and original semantic elements, semantic inference processing is used to accurately obtain the target semantic parsing results of the charging query data. The target semantic parsing results are used to determine the response information regarding the charging query data and send the response information through the charging customer service account. This improves the accuracy of semantic understanding.

[0146] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0147] Based on the same inventive concept, this application also provides a query data processing apparatus for implementing the query data processing method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more query data processing apparatus embodiments provided below can be found in the limitations of the query data processing method described above, and will not be repeated here.

[0148] In one exemplary embodiment, such as Figure 3 As shown, a query data processing device 300 is provided, including: a data acquisition module 302, a data recognition module 304, a template adjustment module 306, and a semantic parsing module 308, wherein:

[0149] Data acquisition module 302 is used to acquire charging query data sent by the target account in the current query conversation and historical conversation data of the target account. The current query conversation is a charging query conversation between the target account and the charging customer service account.

[0150] The data recognition module 304 is used to identify the current service stage of the current query dialogue based on historical dialogue data, and to construct a context anchor point for the charging query data based on the current service stage and historical dialogue data.

[0151] The template adjustment module 306 is used to perform semantic extraction on charging query data based on context anchors to obtain the original semantic elements, call the charging service knowledge graph, and adjust the basic prompt word template corresponding to the current service stage based on context anchors to obtain the target prompt words.

[0152] The semantic parsing module 308 is used to obtain the target semantic parsing result of the charging query data through semantic inference processing based on the target prompt words and original semantic elements. The target semantic parsing result is used to determine the reply information about the charging query data and send the reply information through the charging customer service account.

[0153] In some embodiments, the template adjustment module 306 is used to perform synonym semantic boundary recognition on the charging query data based on context anchors, so as to divide the charging query data into multiple structured semantic fragment groups; obtain the business semantic domain corresponding to the current service stage; perform lexical matching under domain constraints on each structured semantic fragment group based on each structured semantic fragment group and the business semantic domain to obtain the target explicit keywords that match the current service stage; construct a text reference mapping table based on each structured semantic fragment group and context anchors; and determine the original semantic elements based on the text reference mapping table and the target explicit keywords.

[0154] In some embodiments, the template adjustment module 306 is used to determine multiple matching candidate intent nodes from the charging service knowledge graph based on context anchors, and to filter out target intent nodes from the multiple candidate intent nodes based on the degree of matching between each candidate intent node and the current service stage; to fill the basic prompt word template corresponding to the current service stage based on the semantic attributes carried by the target intent node to obtain candidate prompt words; and to determine the target prompt word based on the candidate prompt words and the stage boundary conditions corresponding to the current service stage.

[0155] In some embodiments, the template adjustment module 306 is used to extract multiple service action paths from the charging service knowledge graph; for each candidate prompt word, based on the degree of matching between the candidate prompt word and each service action path, the service action path that matches the candidate prompt word is selected from the multiple service action paths; based on the service action path matched by each candidate prompt word and the stage boundary conditions corresponding to the current service stage, the multiple candidate prompt words are subjected to stage constraint filtering to obtain the target prompt word.

[0156] In some embodiments, the original semantic elements include target explicit keywords and target implicit referential clues. The semantic parsing module 308 is used to match the explicit keywords with entity nodes in the charging service knowledge graph to obtain an initial entity matching group; construct a complete entity group based on the target implicit referential clues and context anchors; merge the initial entity matching group and the complete entity group to obtain a fused entity group; and determine the target semantic parsing result of the charging query data based on the fused entity group and target prompt words.

[0157] In some embodiments, the semantic parsing module 308 is used to filter relationship paths in the charging service knowledge graph, starting with each entity in the fused entity group and using the current service stage as a constraint, to obtain multiple relationship paths; based on the semantic coverage of the target prompt words with each relationship path, at least one candidate relationship path is selected from the multiple relationship paths; for each candidate relationship path, the state consistency of the candidate relationship path is verified based on the service state marker in the context anchor, to obtain the corresponding verification result; based on the verification result of each candidate relationship path, a target relationship path is selected from at least one candidate relationship path, and the endpoint entity indicated by the target relationship path and the corresponding association are semantically integrated to obtain the target semantic parsing result.

[0158] Each module in the aforementioned query data processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0159] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 4As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media to run. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a data query processing method.

[0160] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0161] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0162] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0163] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0164] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0165] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0166] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0167] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for processing query data, characterized in that, The method includes: Obtain charging query data sent through the target account in the current query conversation and the historical conversation data of the target account, wherein the current query conversation is a charging query conversation between the target account and the charging customer service account; Based on the historical dialogue data, the current service stage of the current query dialogue is identified, and a context anchor point for the charging query data is constructed based on the current service stage and the historical dialogue data. Based on the context anchors, semantic extraction is performed on the charging query data to obtain the original semantic elements; Based on the context anchor, multiple candidate intent nodes are identified from the charging service knowledge graph, and the target intent node is selected from the multiple candidate intent nodes based on the degree of matching between each candidate intent node and the current service stage. Based on the semantic attributes carried by the target intent node, the basic prompt word template corresponding to the current service stage is filled to obtain candidate prompt words; Based on the candidate suggestion words and the stage boundary conditions corresponding to the current service stage, the target suggestion word is determined; Based on the target prompt words and the original semantic elements, the target semantic parsing result of the charging query data is obtained through semantic inference processing. The target semantic parsing result is used to determine the reply information about the charging query data and send the reply information through the charging customer service account.

2. The method according to claim 1, characterized in that, The semantic extraction of the charging query data based on the context anchor points yields the original semantic elements, including: Based on the context anchors, synonym semantic boundary recognition is performed on the charging query data to divide the charging query data into multiple structured semantic fragment groups; Obtain the business semantic domain corresponding to the current service stage; Based on each structured semantic fragment group and the business semantic domain, perform lexical matching under domain constraints on each structured semantic fragment group to obtain the target explicit keywords that match the current service stage; Based on each group of structured semantic fragments and the context anchor, a text reference mapping table is constructed; Based on the text reference mapping table and the target explicit keywords, the original semantic elements are determined.

3. The method according to claim 2, characterized in that, The step of determining the target prompt word based on the candidate prompt words and the stage boundary conditions corresponding to the current service stage includes: Multiple service action paths are extracted from the charging service knowledge graph; For each candidate suggestion word, based on the degree of matching between the candidate suggestion word and each service action path, the service action path that matches the candidate suggestion word is selected from multiple service action paths; Based on the service action path matched by each candidate prompt word and the stage boundary conditions corresponding to the current service stage, stage constraint filtering is performed on multiple candidate prompt words to obtain the target prompt word.

4. The method according to claim 1, characterized in that, The original semantic elements include target explicit keywords and target implicit referential clues. Based on the target prompts and the original semantic elements, the target semantic parsing result of the charging query data is obtained through semantic inference processing, including: The initial entity matching group is obtained by matching the explicit keywords with the entity nodes in the charging service knowledge graph. Based on the target implicit referential clues and contextual anchors, construct a complete entity group; The initial entity matching group and the completed entity group are merged to obtain the fused entity group; Based on the fused entity group and the target prompt words, the target semantic parsing result of the charging query data is determined.

5. The method according to claim 4, characterized in that, The step of determining the target semantic parsing result of the charging query data based on the fused entity group and the target prompt words includes: Starting with each entity in the fused entity group and using the current service stage as a constraint, multiple relationship paths are obtained by filtering the relationship paths in the charging service knowledge graph. Based on the semantic coverage of the target prompt words with each relation path, at least one candidate relation path is selected from multiple relation paths; For each candidate relationship path, a state consistency verification is performed on the candidate relationship path based on the service state flag in the context anchor, and the corresponding verification result is obtained. Based on the verification results of each candidate relationship path, a target relationship path is selected from at least one candidate relationship path. The endpoint entity indicated by the target relationship path and the corresponding association relationship are semantically integrated to obtain the target semantic parsing result.

6. A query data processing device, characterized in that, The device includes: The data acquisition module is used to acquire charging query data sent by the target account in the current query conversation and the historical conversation data of the target account. The current query conversation is a charging query conversation between the target account and the charging customer service account. The data recognition module is used to identify the current service stage of the current query dialogue based on the historical dialogue data, and to construct a context anchor point for the charging query data based on the current service stage and the historical dialogue data. The template adjustment module is used to perform semantic extraction on the charging query data based on the context anchor to obtain the original semantic elements; based on the context anchor, determine multiple matching candidate intent nodes from the charging service knowledge graph, and filter the target intent node from the multiple candidate intent nodes based on the matching degree of each candidate intent node with the current service stage; based on the semantic attributes carried by the target intent node, fill the basic prompt word template corresponding to the current service stage to obtain candidate prompt words; and determine the target prompt word based on the candidate prompt words and the stage boundary conditions corresponding to the current service stage. The semantic parsing module is used to obtain the target semantic parsing result of the charging query data through semantic inference processing based on the target prompt words and the original semantic elements. The target semantic parsing result is used to determine the reply information about the charging query data and send the reply information through the charging customer service account.

7. The apparatus according to claim 6, characterized in that, The template adjustment module is used to perform synonym semantic boundary recognition on the charging query data based on the context anchor point, so as to divide the charging query data into multiple structured semantic fragment groups; obtain the business semantic domain corresponding to the current service stage; and perform lexical matching under domain constraints on each structured semantic fragment group based on each structured semantic fragment group and the business semantic domain to obtain the target explicit keywords that match the current service stage. Based on each group of structured semantic fragments and the context anchor, a text reference mapping table is constructed; Based on the text reference mapping table and the target explicit keywords, the original semantic elements are determined.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images