Multi-system access control method and device based on dynamic intention, storage medium and program product

Abstract

The embodiment of the invention provides a multi-system access control method and device based on a dynamic intention, a storage medium and a program product. In the embodiment of the invention, the problem description information is received, a dynamic intention recognition model realized based on a large language model is utilized, the problem description information is converted into a mark sequence and is segmented through the guidance of a first cue word, and mixed intention recognition and mark fragment single intention recognition technologies are cooperatively applied to the sequence; generating a plurality of target sub-intentions fusing the two types of candidate sub-intentions and the probability distribution thereof; and further utilizing an access control model realized based on a large language model to guide and determine a target service system and an access path thereof through a second cue word, constructing a calling logic relationship, carrying out collaborative calling and result integration on the target service system in sequence, and finally generating and outputting target answer information. According to the method, the response efficiency of multi-system query is remarkably improved through an end-to-end model driving process, so that the user experience feeling is enhanced.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a multi-system access control method, device, storage medium, and program product based on dynamic intent. Background Art

[0002] With the rapid development of artificial intelligence (AI) technology, intelligent question-answering systems are widely used in customer service, finance, e-commerce, and other fields. For example, an e-commerce customer service system provides users with an intuitive online customer service interface. Users can directly enter questions on the online customer service interface, such as those seeking order status, logistics information, or returns and exchanges. After receiving the user's question, the system identifies the user's intent and, based on the identified intent, calls the corresponding business system to query and answer the question.

[0003] In actual applications, user questions are often more complex and require the coordinated operation of multiple business systems to be resolved. For example, when a user wants to inquire about the detailed composition of the payment amount for a specific order and the final price after using a coupon, this involves the order system, payment system, and promotion system. When dealing with such complex problems, existing technologies usually adopt a step-by-step guidance method. That is, the customer service system first asks the user to enter a specific order number, based on which it calls the order system to obtain basic order information and feedback to the user; then, the user needs to further inquire about the detailed composition of the payment amount, and the system will call the payment system to inquire about the payment details and answer the question; finally, if the user wants to know the final price after using the coupon, they need to ask again explicitly, and the system will call the promotion system to verify the coupon usage information and give the corresponding answer. The entire query process presented by existing technologies has low response efficiency and poor user experience. Summary of the Invention

[0004] The embodiments of the present application provide a multi-system access control method, device, storage medium and program product based on dynamic intent, which are used to improve the response efficiency of the query process and thereby enhance the user experience.

[0005] The embodiment of the present application provides a multi-system access control method based on dynamic intent, comprising: receiving problem description information, the problem description information is described in natural language, and the problem description information is related to at least one business system among multiple business systems; generating a first prompt word according to the problem description information, inputting the first prompt word into a dynamic intent recognition model implemented based on a large language model, and under the guidance of the first prompt word, converting the problem description information into a tag sequence, dividing the tag sequence into a plurality of tag segments with independent semantics; performing mixed intent recognition on the tag sequence to obtain a plurality of first candidate sub-intentions and their probability distributions; performing single intent recognition on the plurality of tag segments to obtain a plurality of second candidate sub-intentions and their probability distributions; and performing single intent recognition on the plurality of first candidate sub-intentions and their probability distributions according to the plurality of first candidate sub-intentions and their probability distributions and the plurality of second candidate sub-intentions and their probability distributions. The invention discloses a method for distributing the target sub-intentions, generating multiple target sub-intentions; generating a second prompt word according to the multiple target sub-intentions, inputting the second prompt word into the access control model implemented based on the large language model, and under the guidance of the second prompt word, determining the multiple target business systems and their access path information corresponding to the multiple target sub-intentions based on the existing sub-intentions and the functional characteristics of the multiple business systems; constructing the calling logic relationship between the multiple target business systems according to the dependency and mutual exclusion relationship between the multiple business systems; calling the multiple target business systems according to the calling logic relationship, access path information and multiple target sub-intentions between the multiple target business systems to obtain the access result information returned by the multiple target business systems; integrating the calling results returned by the multiple target business systems to generate the target answer information corresponding to the question description information, and outputting the target answer information.

[0006] An embodiment of the present application also provides an electronic device, including: a processor and a memory, the memory being used to store a computer program. When the computer program is executed by the processor, the processor is enabled to implement each step of the dynamic intent-based multi-system access control method provided in the embodiment of the present application.

[0007] An embodiment of the present application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor is enabled to implement each step of the dynamic intent-based multi-system access control method provided in the embodiment of the present application.

[0008] An embodiment of the present application also provides a computer program product, including a computer program / instruction. When the computer program / instruction is executed by a processor, the processor is enabled to implement the various steps of the dynamic intent-based multi-system access control method provided in the embodiment of the present application.

[0009] In an embodiment of the present application, by receiving the question description information, using the dynamic intent recognition model implemented based on the large language model, guided by the first prompt word, the question description information is converted into a tag sequence and segmented, and the hybrid intent recognition and tag fragment single intent recognition technologies are collaboratively applied to the sequence to generate multiple target sub-intentions that integrate two types of candidate sub-intentions and their probability distributions; and then using the access control model implemented based on the large language model, guided by the second prompt word, combined with the existing sub-intentions, business system functional characteristics, and inter-system dependencies and mutual exclusion relationships, the target business system and its access path are determined, and a call logic relationship is constructed, thereby collaboratively calling the target business system and integrating the call results, and finally generating and outputting the target answer information. This solution significantly improves the response efficiency of multi-system queries through an end-to-end model-driven process, thereby effectively enhancing the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0011] Figure 1a A flowchart of a multi-system access control method based on dynamic intent provided by an exemplary embodiment of the present application;

[0012] Figure 1b A schematic diagram of a model architecture of a multi-system access control method based on dynamic intent provided by an exemplary embodiment of the present application;

[0013] Figure 2 A schematic diagram of a model architecture of a multi-system access control method based on dynamic intent provided by another exemplary embodiment of the present application;

[0014] Figure 3 A schematic structural diagram of an electronic device provided as an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0015] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0016] It should be noted that, in the case of user information involved in the embodiments of the present application, 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, stored data, displayed data, etc.) involved in the embodiments of the present application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or large models) are in compliance with relevant laws and standards.

[0017] In response to the technical defects of existing intelligent question-answering systems that require multiple interactions, have low response efficiency, and have poor user experience when processing complex questions involving multiple business systems, an embodiment of the present application provides a solution. By receiving question description information and utilizing a dynamic intent recognition model implemented based on a large language model, guided by a first prompt word, the question description information is converted into a tag sequence and segmented, and the hybrid intent recognition and tag fragment single intent recognition technologies are collaboratively applied to the sequence to generate multiple target sub-intentions that integrate two types of candidate sub-intentions and their probability distributions; and then utilizing an access control model implemented based on a large language model, guided by a second prompt word, combined with existing sub-intentions, business system functional characteristics, and dependencies and mutual exclusions between systems, the target business system and its access path are determined, and a call logic relationship is constructed. Based on this, the target business system is collaboratively called and the call results are integrated, and finally the target answer information is generated and output. This solution significantly improves the response efficiency of multi-system queries through an end-to-end model-driven process, thereby effectively enhancing the user experience.

[0018] The following describes in detail the technical solutions provided by various embodiments of the present application in conjunction with the accompanying drawings.

[0019] Figure 1a A flowchart of a multi-system access control method based on dynamic intent provided by an exemplary embodiment of the present application. Figure 1b The model architecture diagram of the multi-system access control method based on dynamic intent provided by an exemplary embodiment of the present application is implemented as shown in FIG. Figure 1b As shown in , the model architecture includes: dynamic intent recognition model, access control model and multiple business systems (including business system 1, business system 2, etc.). Figure 1a As shown, the method includes:

[0020] S101: Receive problem description information, where the problem description information is described in a natural language and is related to at least one business system among a plurality of business systems;

[0021] S102: Generate a first prompt word based on the question description information, input the first prompt word into a dynamic intent recognition model implemented based on a large language model, and under the guidance of the first prompt word, convert the question description information into a token sequence, and segment the token sequence into multiple token segments with independent semantics; perform mixed intent recognition on the token sequence to obtain multiple first candidate sub-intents and their probability distributions; perform single intent recognition on the multiple token segments to obtain multiple second candidate sub-intents and their probability distributions; and generate multiple target sub-intents based on the multiple first candidate sub-intents and their probability distributions and the multiple second candidate sub-intents and their probability distributions.

[0022] S103: Generate a second prompt word based on the multiple target sub-intents, input the second prompt word into the access control model implemented based on the large language model, and under the guidance of the second prompt word, determine the multiple target business systems corresponding to the multiple target sub-intents and their access path information based on the existing sub-intents and the functional characteristics of the multiple business systems; establish a call logic relationship between the multiple target business systems based on the dependency and mutual exclusion relationships between the multiple business systems; and call the multiple target business systems according to the call logic relationship between the multiple target business systems, the access path information, and the multiple target sub-intents to obtain access result information returned by the multiple target business systems;

[0023] S104: Integrate the call results returned by multiple target business systems, generate target answer information corresponding to the question description information, and output the target answer information.

[0024] In an embodiment of the present application, the device responsible for executing each step of the dynamic intent-based multi-system access control method provided in the above embodiment can be implemented as an electronic device, and the dynamic intent-based multi-system access control can be implemented on a server device. A server device refers to a high-performance device for providing computing resources and services. The embodiment of the present application does not limit the specific form of the server device. For example, the server device can be a physical server, a virtual server, or a distributed server cluster.

[0025] In an embodiment of the present application, the problem description information is a query request described by the user in natural language through a terminal device. In an embodiment of the present application, the specific type of the terminal device is not limited. For example, the terminal device can be a mobile device (such as a smart phone, tablet computer), or a fixed device (such as a desktop computer, workstation). Among them, the problem description information is related to at least one business system among multiple business systems. A business system is a service module or data source that independently provides specific functions, and its types are diverse. For example, a business system can be an order system that manages product information and order status, a payment system that processes transaction amounts and payment details, a promotion system that verifies coupons and calculates discount prices, or a logistics system that tracks delivery progress, etc. The specific type of the business system is not limited in the embodiment of the present application. The fact that the problem description information is related to at least one business system means that the user's problem description information needs to obtain an answer by accessing one or more business systems, and each business system is only responsible for a specific function. Therefore, it is necessary to call multiple business systems to obtain access result information in order to solve complex problems. For example, the problem description information is "My order 456 was not shipped on time. Please compensate me with a 30 yuan coupon and inform me of the latest logistics time." Then the problem description information is related to the order system (confirming the status of order 456 and the shipping responsibility), the logistics system (querying the current delivery progress) and the promotion system (generating a 30 yuan compensation coupon).

[0026] In an embodiment of the present application, the first prompt word is generated based on the received problem description information. As an input instruction of the dynamic intent recognition model implemented based on the large language model, it not only contains the problem description information, but also includes guidance information that instructs the dynamic intent recognition model to perform intent recognition from the overall semantics and single semantics. Among them, the dynamic intent recognition model is an advanced model based on AI (Artificial Intelligence) with autonomous thinking, reasoning and output capabilities. For example, the dynamic intent recognition model can be a generative model based on AI, the core principle of which is to train a large amount of text data through deep learning technology, especially a large language model, so as to have a deep understanding and generation ability of language. After receiving the first prompt word, the dynamic intent recognition model can conduct an in-depth analysis of the problem description information based on the guidance information in the first prompt word, grasp the general intention of the problem as a whole, and also focus on the various details in the problem, identify the specific intention under a single semantics, and finally output multiple target sub-intentions, providing precise intent guidance for subsequent access control and business system calls.

[0027] Specifically, under the guidance of the first prompt word, the dynamic intent recognition model converts the problem description information into a tag sequence, and divides the tag sequence into multiple tag segments with independent semantics. Among them, the tag sequence is an ordered arrangement of a series of tags (such as tokens) obtained after processing the problem description information. Each tag (token) represents a basic element in the problem description information, which can be a basic language unit such as a word, subword, or character, depending on the processing method. The tag segment is a subsequence with complete independent semantics in the tag sequence, which is dynamically segmented by the dynamic intent recognition model through the recognition of context boundaries. Each tag segment corresponds to an independently executable business intent unit. For example, the problem description information is "Query the payment details of order 123 and the final price after using the coupon SUMMER20", and the generated tag sequence can be ["query", "order", "123", "of", "payment", "details", "and", "use", "coupon", "SUMMER20", "after", "of", "final", "price"]. For another example, based on the above tag sequence, the tag fragments may include ["query", "order", "123"] → "query order 123", ["payment", "details"] → "payment details", ["use", "coupon", "SUMMER20"] → "use coupon SUMMER20", ["final", "price"] → "final price".

[0028] Furthermore, the intent recognition model performs mixed intent recognition on the tag sequence to obtain multiple first candidate sub-intentions and their probability distributions; performs single intent recognition on multiple tag segments to obtain multiple second candidate sub-intentions and their probability distributions; and generates multiple target sub-intentions based on the multiple first candidate sub-intentions and their probability distributions and the multiple second candidate sub-intentions and their probability distributions. Mixed intent recognition refers to the dynamic intent recognition model analyzing the entire tag sequence, identifying multiple possible sub-intentions from the perspective of overall semantics, and assigning a probability value to each sub-intention to represent the possibility of the intent. These possible sub-intentions and their corresponding probabilities constitute the first candidate sub-intentions and their probability distributions. Single intent recognition refers to the dynamic intent recognition model dividing the tag sequence into multiple tag segments with independent semantics, performing intent recognition on each tag segment separately, and also assigning a probability to the possible intent of each tag segment to form the second candidate sub-intention and its probability distribution. Finally, the dynamic intent recognition model combines the first candidate sub-intention and its probability distribution with the second candidate sub-intention and its probability distribution to generate the final multiple target sub-intentions. This approach, which identifies multiple sub-intents based on both overall semantics for mixed intent and single semantics for single intent, more comprehensively considers both the overall semantics and details of the question description, improving the accuracy of intent identification. This provides a more accurate foundation for dynamic intent-based multi-system access control, making multi-system access control more accurate, improving query response efficiency, and enhancing the user experience.

[0029] For example, when a user enters a question description such as "Query the payment details for order 123 and the final price after using the SUMMER20 coupon," hybrid intent recognition can capture the implicit cross-system dependency of the "final price" objective—which requires collaborative calculations with the payment and promotion systems. This generates the first candidate sub-intent and its probability distribution, such as the output {"Query payment details": 0.92, "Calculate discount price": 0.88}. The probability value represents the dynamic intent recognition model's quantitative confidence in the likelihood of the sub-intent. Taking the aforementioned question as an example, the dynamic intent recognition model divides the token sequence into four token segments: ["Query order 123"] corresponds to the order system query intent, ["Payment details"] corresponds to the payment system detail retrieval intent, ["Use SUMMER20 coupon"] corresponds to the promotion system detail retrieval intent, and ["Final price"] implies the price calculation requirement. After analyzing each labeled segment independently, we generate a second candidate sub-intent and its probability distribution, such as {"order query": 0.95, "payment details": 0.97, "coupon redemption": 0.93, "price acquisition": 0.85}. The probability value reflects the confidence that each labeled segment expresses the sub-intent. The final target sub-intent might be {"order query": 0.95, "payment details query": 0.92, "coupon redemption": 0.93, "discount price calculation": 0.88}.

[0030] In an embodiment of the present application, the second prompt word is generated based on multiple target sub-intentions, which serve as input instructions for the access control model implemented based on the large language model. In addition to containing the target sub-intention, the second prompt word can also be accompanied by some other important information to help the access control model more accurately determine the target business system and its access path information. For example, the second prompt word can include the following aspects: priority information of multiple target sub-intentions, functional feature descriptions of multiple business systems, descriptions of dependencies between multiple business systems, and constraints on access paths. Among them, different target sub-intentions may have different priorities. The priority information of multiple target sub-intentions can help the access control model determine the call logic of multiple business systems in order of priority when processing multiple target sub-intentions; the functional feature descriptions of multiple business systems can provide functional features of business systems related to multiple target sub-intentions, so that the access control model can more accurately match the correspondence between target sub-intentions and business systems; the dependency descriptions between multiple business systems can clearly indicate the dependencies between business systems, such as some business systems need to be called first before they can provide data support for other business systems. These dependencies are crucial for building correct call logic; the constraints on access paths can guide the access control model to select appropriate access paths to ensure data security and compliance.

[0031] Among them, the access control model is an intelligent decision-making system based on a large language model. Its core principle is to realize dynamic analysis and access control decision-making of target sub-intentions through joint modeling of multimodal inputs (such as text, metadata, and permission rules), combined with rule engines and deep learning reasoning capabilities. For example, the access control model can be a multi-task learning model based on a large language model architecture. It has the ability to jointly reason about complex intentions and permission rules through training on massive business system interfaces, permission policies, and user behavior data. After receiving the second prompt word, the access control model can use multiple target sub-intentions and other accompanying information, combined with the functional characteristics of existing sub-intentions and multiple business systems, to determine the target business system and its access path information corresponding to each target sub-intention. At the same time, the access control model can construct the calling logic relationship between the target business systems based on the dependencies and mutual exclusion relationships between the business systems, thereby realizing efficient and accurate calling of the target business system, ensuring the accuracy and efficiency of multi-system access control.

[0032] Specifically, guided by the second prompt, the access control model identifies multiple target business systems and their access path information corresponding to multiple target sub-intents based on existing sub-intents and the functional characteristics of multiple business systems. Existing sub-intents refer to previously identified and recorded sub-intents. These can come from historical query records, a system-defined common intent library, or intents successfully identified in previous similar problem descriptions. These sub-intents provide the access control model with a reference and empirical experience, helping it more accurately understand the meaning and requirements of the current target sub-intent. Each business system has specific functions and operational scopes. The functional characteristics of multiple business systems provide detailed descriptions of these functions and operational scopes, helping the access control model understand the capabilities of each business system and thus match the target sub-intent with the appropriate business system. For example, an order system has functions such as querying order status and updating order information; an inventory system can query inventory quantities and record inbound and outbound operations. Access path information refers to the detailed information required to access a target business system, ensuring that the access control model can correctly construct a request and successfully call the target business system. These detailed information includes, but is not limited to, the access interface address, request method, required parameter format, authentication method, and security protocol.

[0033] Furthermore, the access control model constructs a logical calling relationship between multiple target business systems based on the dependencies and mutual exclusion relationships between multiple business systems. A dependency relationship between multiple business systems means that one business system relies on the output of another business system to properly perform its functions. For example, the payment system must obtain the payment serial number provided by the order system to query the refund progress. This data transfer constitutes a dependency relationship. A mutual exclusion relationship between multiple business systems means that business systems cannot be called in parallel due to resource competition or conflicting business rules. For example, the points system and the coupon system share a pool of user discount quotas. Calling them simultaneously will result in resource overallocation.

[0034] The call logic relationship represents the call sequence and input-output dependencies between multiple target business systems. The call sequence refers to the order in which the access control model calls multiple target business systems. For example, the order system is called first, followed by the payment system. Input-output dependencies refer to the output of one business system being used as input data for another. For example, the payment ID output by the order system is automatically injected into the payment system interface.

[0035] Furthermore, the access control model can call multiple target business systems according to the call logic relationship between the multiple target business systems, access path information, and multiple target sub-intentions to obtain access result information returned by the multiple target business systems. Among them, the access result information refers to the information returned by the target business system after being called, that is, the call result returned by the target business system. In the embodiments of the present application, the specific type of access result information is not limited. For example, the access result information can be a query result, an operation result, a data update result, etc.

[0036] In an embodiment of the present application, the call results returned by multiple target business systems are integrated to generate target answer information corresponding to the question description information, and the target answer information is output. The core operation of the integration may include standardization, logical association, and semantic reorganization of the call results returned by the multiple target business systems. Standardization processing includes unit unification (for example, the payment system returns 299.00CNY → converted to 299 yuan), time formatting (for example, the logistics timestamp 2024-06-29T14:30:00Z → converted to today's 14:30) and status code escaping (for example, converting the system's internal status code into user language, such as converting 200 to success); logical association processing includes data concatenation (for example, the order system outputs order number = 123 + the payment system outputs payment amount = 299 yuan → associated with order 123 payment amount: 299 yuan) and contradiction resolution (for example, if the payment system returns payment but the logistics system returns not shipped, the automatic supplementary statement: the payment has been deducted and the product is being prepared); semantic reorganization processing includes converting machine data structures into natural language (for example, {"discount calculation":{"original price":299,"discount":20,"actual payment":279}} → converted to the product's original price of 299 yuan, using a coupon to get an instant discount of 20 yuan, and the final actual payment of 279 yuan).

[0037] Figure 2 This is a schematic diagram of the model architecture of a multi-system access control method based on dynamic intent provided by another exemplary embodiment of the present application. Figure 2 As shown, an example of the internal implementation structure of the dynamic intent recognition model is shown. In this example, the dynamic intent recognition model includes an input layer A, an encoding layer B, an entity extraction layer C, and an intent recognition layer D.

[0038] Among them, the input layer A in the dynamic intent recognition model includes a word segmenter. Based on the internal implementation structure of the dynamic intent recognition model, a specific implementation method of converting the problem description information into a token sequence and dividing the token sequence into multiple token fragments with independent semantics includes: using the word segmenter in the input layer A to perform word segmentation processing on the problem description information, and performing punctuation normalization processing on the word segmentation results to obtain a token sequence; matching each word segmentation in the token sequence with the semantic segmentation symbol in the predefined rule template, using the word segmentation position of the semantic segmentation symbol in the match as the segmentation position, and semantically segmenting the token sequence according to the segmentation position to obtain multiple token fragments.

[0039] Among them, the word segmenter can decompose the problem description information into smaller units, which can be words, subwords, characters or other meaningful language fragments. In this embodiment, the specific type of the word segmenter is not limited. For example, the word segmenter can be a jieba word segmenter, or it can be a custom word segmenter (Tokenizer) based on a pre-trained model such as BERT. Word segmentation processing refers to an adaptive word segmentation technology based on contextual semantics and business scenarios. Its core lies in dynamically adjusting the segmentation strategy in combination with domain knowledge. For example, when the problem description information is "the payment details of order 123 and the price after the SUMMER20 coupon", the word segmenter can recognize "payment details" as a business atomic phrase instead of splitting it into "payment" and "details", while retaining "SUMMER20" as a complete coupon coding entity, generating a word segmentation result: ["order", "123", "of", "payment details", "and", "SUMMER20", "coupon", "after", "price"]. The word segmentation result needs to be normalized by punctuation, that is, full-width / half-width symbols are uniformly converted into standard separators (such as full-width commas "," are converted into half-width ","), forming a standardized tag sequence: ["order", "123", "of", "payment details", ","", "SUMMER20", "coupon", "after", "price", "?"].

[0040] Semantic segmentation relies on semantic segmentation symbols in predefined rule templates. Semantic segmentation symbols are essentially specialized identifiers that mark intent context boundaries. Their function is to identify logical breakpoints within natural language problem descriptions. Predefined rule templates are a set of predefined rules and patterns that guide how to semantically segment problem descriptions. Predefined rule templates include specific words, phrases, or symbols, such as "+," "by the way," and "and." These words, phrases, or symbols are defined as semantic segmentation symbols, indicating the location of semantic segmentation. A segmentation position refers to a location in a token sequence that can be segmented, as determined by the semantic segmentation symbol. For example, in the token sequence "Query the payment status of an order + payment amount," the position of "+" is the segmentation position. Semantic segmentation refers to the process of segmenting a token sequence into multiple semantically independent token segments based on the segmentation position. Each token segment expresses a complete or relatively independent meaning. For example, based on the segmentation position, the token sequence is segmented into two token segments: "Query the payment status of an order" and "payment amount." Each segment expresses an independent semantic intent.

[0041] In this embodiment, the dynamic intent recognition model is designed to solve the problem that the traditional intent recognition model is inaccurate when processing complex multi-intent queries. A new model architecture is proposed. The entire dynamic intent recognition model adopts a hybrid architecture design, which not only includes a large language model and the word segmenter in the above embodiment, but also includes some traditional neural network models. Figure 2 As shown, the encoding layer B in the dynamic intent recognition model includes a multi-layer encoding network using an attention mechanism, the entity extraction layer C includes an entity extraction model, and the intent recognition layer D includes an intent classifier based on a large language model. Among them, the intent classifier is implemented using a large language model. In the embodiments of the present application, the specific type of the large language model is not limited. For example, the large language model can be GPT-4, Tongyi Qianwen, or Kimi, etc.

[0042] Based on the internal implementation structure of the dynamic intent recognition model described in the above embodiment, a specific implementation method of performing mixed intent recognition on a tag sequence to obtain multiple first candidate sub-intentions and their probability distributions; performing single intent recognition on multiple tag fragments to obtain multiple second candidate sub-intentions and their probability distributions includes: inputting the tag sequence and the multiple tag fragments into the encoding layer B respectively and using a multi-layer encoding network with an attention mechanism for encoding processing to obtain a first feature vector and multiple second feature vectors with context information; inputting the first feature vector and the multiple second feature vectors into the entity extraction model in the entity extraction layer C respectively to identify and label business entities to obtain a third feature vector and multiple fourth feature vectors with business entity labeling information; inputting the third feature vector and the multiple fourth feature vectors into the intent classifier in the intent recognition layer D respectively, performing mixed intent recognition on the third feature vector to obtain multiple first candidate sub-intentions, performing single intent recognition on the multiple fourth feature vectors respectively to obtain multiple second candidate sub-intentions, and applying an activation function to calculate the probability distribution of each of the multiple first candidate sub-intentions and the multiple second candidate sub-intentions.

[0043] Among them, the multi-layer encoding network with an attention mechanism is a deep learning network structure. The attention mechanism enables the multi-layer encoding network to focus on important parts when processing a tag sequence, thereby better capturing contextual information. For example, a multi-layer encoding network that includes an attention mechanism can be implemented using a Transformer network model. The multi-layer encoding network can be composed of multiple encoding layers, each of which includes an attention mechanism to gradually extract higher-level features. After the tag sequence is encoded by the encoding layer, a first feature vector is obtained. It not only contains the semantic information of each tag itself, but also incorporates the contextual information of the surrounding tags. The first feature vector can more comprehensively represent the semantic role and relationship of the tags in the sequence. After multiple tag segments are encoded by the encoding layer, a second feature vector is obtained. Similar to the first feature vector, the second feature vector is generated for each tag segment. Each second feature vector corresponds to a tag segment and also contains the contextual information of the tag segment.

[0044] Based on the obtained first feature vector and multiple second feature vectors, they can be input into the entity extraction model in the entity extraction layer C to identify and label business entities. Among them, the entity extraction model can be a machine learning model based on sequence labeling, which is used to identify business entities in a tag sequence or multiple tag segments, and label the identified business entities on the feature vector. For example, the entity extraction model can be implemented using a BiLSTM-CRF entity extraction submodel. Using this entity extraction model, business entities containing incorrect formats can be corrected and extracted. Business entities refer to key information units related to specific business operations in the problem description information. For example, user ID, order number, mobile phone number, etc. Based on the first feature vector, the vector at the location of the identified business entity is embedded with a business entity label to form a feature representation with enhanced semantics as a third feature vector with business entity labeling information. Similarly, the second feature vector of each tag segment is embedded with the business entity label information it contains as multiple fourth feature vectors with business entity labeling information. For example, the input tag sequence: ["query", "order", "A123", "of", "logistics"], the third feature vector after entity labeling: [query, order,<order_id> , of, logistics] (where<order_id> label embedding vector).

[0045] Optionally, in addition to labeling business entities for the first feature vector and multiple second feature vectors, the entity extraction model can also be expanded to label parts of speech (such as verb / noun), sentiment polarity (such as positive / negative) or syntactic dependencies (such as subject, predicate, and object) to enrich the semantic information of the feature vector.

[0046] The intent classifier is a neural network-based classification model (its architecture can include fully connected layers, activation functions, and the Sentence-BERT model). It predicts the intent category based on the third eigenvector and multiple fourth eigenvectors, including mixed intent recognition and single intent recognition. Simultaneously, the intent classifier applies the activation function to calculate the probability distribution of multiple first candidate sub-intents and multiple second candidate sub-intents.

[0047] Mixed intent recognition on the third eigenvector and application of an activation function to calculate the probability distribution of multiple first candidate sub-intents can be achieved by the following steps: 1. Input the third eigenvector into a fully connected layer, outputting a K-dimensional raw score vector (K = the total number of sub-intent categories); 2. Calculating the independent probability of each sub-intent using the Sigmoid activation function. For example, if the question description is "Query the order status and logistics of ID-10086," the third eigenvector corresponding to the question description is input into the fully connected layer of the intent classifier, outputting the raw score vector: [Order Query: 2.2, Logistics Query: 1.5, Cancel: -1.0]. The multiple first candidate sub-intents are: "Order Query," "Logistics Query," and "Cancel." Applying the Sigmoid activation function calculates the probability of each of the multiple first candidate sub-intents as [0.90, 0.82, 0.27]. Finally, the multiple first candidate sub-intents and their probability distributions are obtained as [Order Query: 0.90, Logistics Query: 0.82, Cancel: 0.27]. Among them, the activation function Sigmoid is used for mixed intent recognition and independently calculates the probability of each intent.

[0048] Among them, single intent recognition of multiple fourth eigenvectors and application of activation function to calculate the probability distribution of multiple second candidate sub-intentions can be achieved through the following steps: 1. Input each eigenvector (labeled fragment with entity annotation) in the multiple fourth eigenvectors into the fully connected layer, and calculate the original score vector for each labeled fragment; 2. Calculate the probability distribution of all intent categories of each labeled fragment through the activation function Softmax. For example, labeled segment 1: "Order of ID-10086", labeled segment 2: "Logistics", the feature vectors of multiple fourth feature vectors corresponding to labeled segment 1 and labeled segment 2 are input into the fully connected layer, and the original score vector is calculated for labeled segment 1: [Order query: 2.85, Logistics query: 0.32, Cancellation: -1.75], and the original score vector is calculated for labeled segment 2: [Logistics query: 3.10, Order query: -0.45, Cancellation: -2.80]; by applying the activation function Softmax, the probability of all intent categories of labeled segment 1 is calculated as [0.91, 0.07, 0.02], and the probability of all intent categories of labeled segment 2 is calculated as [0.97, 0.03, 0.00]; finally, multiple second candidate sub-intents and their probability distributions are obtained as labeled segment 1: [Order query: 0.91, Logistics query: 0.07, Cancellation: 0.02], labeled segment 2: [Logistics query: 0.97, Order query: 0.03]. Among them, the activation function Softmax is used for single intent recognition, forcing the sum of probabilities to be 1.

[0049] In an optional embodiment, based on the multiple first candidate sub-intentions and their probability distributions and the multiple second candidate sub-intentions and their probability distributions obtained in the above embodiment, multiple target sub-intentions are generated in the intention recognition layer D of the dynamic intention recognition model, including: selecting multiple first valid sub-intentions that meet the first probability distribution conditions from the multiple first candidate sub-intentions according to the probability distribution of the multiple first candidate sub-intentions; selecting multiple second valid sub-intentions that meet the second probability distribution conditions from the multiple second candidate sub-intentions according to the probability distribution of the multiple second candidate sub-intentions; obtaining the intersection of the first valid sub-intention and the second valid sub-intention as the third valid sub-intention; calculating the semantic similarity of any two third valid sub-intentions, taking the two third valid sub-intentions whose semantic similarity is less than the first semantic similarity threshold as independent target sub-intentions, and merging the two third valid sub-intentions whose semantic similarity is greater than the second semantic similarity threshold into one target sub-intention; the first semantic similarity threshold is less than the second semantic similarity threshold.

[0050] Among them, the first probability distribution condition refers to the rule for screening valid intents from the mixed intent recognition results. The second probability distribution condition refers to the rule for screening valid intents from the single intent recognition results. In the embodiment of the present application, the specific form of the first probability distribution condition or the second probability distribution condition is not limited. For example, the first probability distribution condition or the second probability distribution condition can be probability threshold filtering (i.e., selecting sub-intentions with probability > threshold); it can also be a probability difference constraint (i.e., requiring the highest and second highest probability difference > preset sub-intentions); it can also be Top-K selection (i.e., taking the top K intentions with the highest probability). For example, multiple first candidate sub-intentions and their probability distributions are [order query: 0.90, logistics query: 0.82, cancellation: 0.27]. Assume that the first probability distribution condition is probability ≥ 0.6, then multiple first valid sub-intentions include order query (0.90) and logistics query (0.82). For another example, Marked Segment 1: [Order Query: 0.91, Logistics Query: 0.07, Cancellation: 0.02], Marked Segment 2: [Logistics Query: 0.97, Order Query: 0.03], assuming that the second probability distribution condition is the highest probability ≥ 0.9, then the multiple second valid sub-intentions include Marked Segment 1 → Order Query (0.91) and Marked Segment 2 → Logistics Query (0.97). The third valid sub-intention refers to the intersection of the first valid sub-intention and the second valid sub-intention, representing a high-confidence intention that has been double-verified. For example, multiple first valid sub-intentions include {order query, logistics query}, and multiple second valid sub-intentions include {order query, logistics query, payment amount query}, then multiple third valid sub-intentions include {order query, logistics query}.

[0051] Furthermore, based on the third effective sub-intention obtained in the above embodiment, multiple target sub-intentions are generated, and by calculating the semantic similarity of any two third effective sub-intentions, the two third effective sub-intentions whose semantic similarity is less than the first semantic similarity threshold are taken as independent target sub-intentions, and the two third effective sub-intentions whose semantic similarity is greater than the second semantic similarity threshold are merged into one target sub-intention. Wherein, semantic similarity refers to the degree of proximity between two sub-intentions in the semantic space, which is measured by vector cosine similarity (range [-1,1]). When the semantic similarity of two third effective sub-intentions is lower than the first semantic similarity threshold, the two third effective sub-intentions are regarded as semantically unrelated intentions, indicating that the two are essentially independent and should be retained as separate targets and need to be processed independently, which can be called independent intentions; when the semantic similarity of two third effective sub-intentions is higher than the second semantic similarity threshold, the two third effective sub-intentions are regarded as semantically equivalent intentions, indicating that the two are semantically equivalent or highly overlapping and need to be merged into a unified target, which can be called a merged intention. When the semantic similarity of two third-valid sub-intents falls between the first and second semantic similarity thresholds, the two third-valid sub-intents are considered semantically similar intents, and the original intent remains unprocessed and can be called the original intent. This design avoids both semantic loss caused by excessive merging and intent fragmentation caused by excessive splitting. Multiple target sub-intents are a deduplicated collection of multiple independent intents, multiple merged intents, and multiple original intents, preserving semantic independence while eliminating redundancy.

[0052] Optionally, the semantic similarity between any two third valid sub-intents can be calculated using the Sentence-BERT model. Specifically, the steps for calculating semantic similarity using Sentence-BERT include: Step 1: Convert the third valid sub-intent into normalized text; Step 2: Encode the normalized text into a 768-dimensional semantic vector using a pre-trained Sentence-BERT model; Step 3: Calculate the cosine similarity of the two vectors. For example, the problem description information is "Check the order status and logistics trajectory of ID-10086", and the third valid sub-intention includes intent A: order status query and intent B: logistics trajectory query; calculate the semantic similarity between intent A and intent B: step 1, normalize intent A: order status query to query order status, and normalize intent B logistics trajectory query → to query logistics trajectory; step 2, use the pre-trained Sentence-BERT model to encode query order status and query logistics trajectory as "emb_A = model.encode("Query order status")" and "emb_B = model.encode("Query logistics trajectory")"; step 3, calculate sim_score = cosine_sim(emb_A, emb_B) = 0.68, set the first semantic similarity threshold = 0.3, the second semantic similarity threshold = 0.8, the similarity 0.68∈(0.3, 0.8), and keep the two independent sub-intents (query order status and query logistics trajectory) as the target sub-intent.

[0053] In an optional embodiment, the access control model is associated with a knowledge base, and the knowledge base E maintains the correspondence between existing sub-intentions and business systems, access path information required to access business systems, and functional characteristics of multiple business systems. This knowledge base is essentially different from the empirical knowledge base of traditional machine learning models: it does not store general language knowledge or training parameters, but is specifically used to maintain dynamic business information directly related to access control of multiple business systems. The knowledge base is maintained through an interface independent of the model, and does not need to be repeatedly input with each request. It supports real-time hot updates to ensure that the model itself does not need to be reconstructed when business rules change. Among them, the relevant description of the correspondence between existing sub-intentions and business systems, and access path information required to access business systems can be referred to the above embodiments and will not be repeated here.

[0054] In an optional embodiment, the access control model is a newly proposed model architecture to solve the problems of low response efficiency and poor user experience of traditional models when processing complex multi-intent queries. Figure 2As shown, an example of the internal implementation structure of the access control model is shown. In this example, the access control model includes: a relationship matching layer F, a semantic similarity calculation layer G, a generation layer H, and a business system call layer I. Among them, the access control model includes a routing matching model implemented based on a large language model in the relationship matching layer F and the semantic similarity calculation layer G. Based on the internal implementation structure of the access control model, a specific implementation method for determining multiple target business systems corresponding to multiple target sub-intentions and their access path information based on the functional characteristics of existing sub-intentions and multiple business systems includes: inputting multiple target sub-intentions into the routing matching model in the relationship matching layer F, and for each target sub-intention, matching the target sub-intention in the corresponding relationship maintained in the knowledge base.

[0055] The route matching model is a hybrid decision-making model based on a large language model, combining rule matching with semantic computing. It can accurately locate business systems and access paths based on target sub-intents. This route matching model innovatively combines two routing methods: first, matching known intents with a static knowledge base, and second, dynamically generating access paths for unknown intents. This dual-mode mechanism ensures rapid response to high-frequency intents while ensuring 100% routable intents, fundamentally addressing the inability of traditional solutions to handle unforeseen intents.

[0056] Optionally, if there is already a sub-intent in the match, the business system and its access path information corresponding to the existing sub-intent in the match will be used as the target business system and its access path information corresponding to the target sub-intent.

[0057] For example, if multiple target sub-intents (such as "order status query") are input into the routing matching model, each target sub-intent is matched against the corresponding relationships maintained in the knowledge base. The matching sub-intent "order status query" corresponds to the business system "order system" and the access path information is " / api / orders / status?order_id={id}". The routing matching model outputs "Target business system: order system; access path information: / api / orders / status?order_id={id}".

[0058] Optionally, if there is already a sub-intent that has not been matched, then in the semantic similarity calculation layer G, the semantic similarity between the target sub-intent and the functional features of multiple business systems is calculated through the routing matching model, and the business system with the highest semantic similarity to the target sub-intent is used as the target business system corresponding to the target sub-intent; according to the access address, access interface and interface parameters of the target business system, the access path information required to access the target business system is dynamically generated; and the target sub-intent is used as an existing sub-intent, and the target sub-intent, the target business system corresponding to the target sub-intent and its access path information are updated to the corresponding relationship. Among them, the routing matching model can include a Sentence-BERT model, which is used to calculate the semantic similarity between the target sub-intent and the functional features of multiple business systems, further improving the accuracy and efficiency of matching. Among them, the specific description of the Sentence-BERT model can refer to the aforementioned embodiment and will not be repeated here.

[0059] Among them, the access address, access interface, and interface parameters of the target business system in knowledge base E together constitute the business system call protocol specification. The access address of the target business system is the network location identifier of the target business system, which is essentially the basic URL or IP address of the system, for example, https: / / order-system.example.com:8080. The access interface of the target business system is the API path for the specific functional operations of the target business system, which is essentially the resource path based on the access address, for example, / api / v2 / orders / status. The interface parameters of the target business system are the request parameter specifications required for the target business system call, including path parameters, query parameters, and request body parameters, for example, path parameters: / orders / {order_id}, query parameters: ? user_id=10086&type=VIP, request body parameters: JSON / XML format data. The specific steps for dynamically generating the access path information required to access the target business system based on the access address, access interface, and interface parameters of the target business system can be: 1) obtaining the access address (e.g., https: / / customs.example.com), access interface (e.g., / api / tariff / estimate), and interface parameters (e.g., {"goods_value":"{{goods value}}","category":"{{goods category}}","dest_country":"{{destination country}}"}) of the target business system from the knowledge base E; 2) injecting the business entities extracted by intent recognition into the interface parameters (e.g., {"goods value":"2000","goods category":"camera","destination country":"CN"}); and 3) assembling the complete access path information (e.g., POST https: / / customs.example.com / api / tariff / estimate). After generation, the target sub-intent and its path are immediately added to the knowledge base as an existing sub-intent and a new record, enabling the system to self-evolve.

[0060] It should be noted that in the embodiments of the present application, this method of combining static query and dynamic generation to determine multiple target business systems and their access path information corresponding to multiple target sub-intents not only improves matching efficiency, but also ensures that the corresponding target business systems and access path information can be found. At the same time, through the update mechanism of the access control model, the existing sub-intents and their corresponding relationships with business systems and access path information are continuously enriched and increased, so that subsequent access processes can match existing sub-intents with a greater probability, further improving efficiency.

[0061] In an alternative embodiment, based on Figure 2The relationship generator in the generation layer H of the access control model shown here constructs a call logic relationship between multiple target business systems based on the dependencies and mutual exclusion relationships between multiple business systems. This includes: inputting the identification information of multiple target business systems and multiple target sub-intentions into the relationship generator in the generation layer H of the access control model, analyzing the association relationships between the multiple target sub-intentions; if the multiple target sub-intentions are in a parallel relationship, a parallel call logic relationship is formed between the multiple target business systems; if the multiple target sub-intentions are in a coupled relationship, based on the identification information of the multiple target business systems and the dependencies and mutual exclusion relationships between the multiple business systems, a coupled call logic relationship is formed between the multiple target business systems. The coupled call logic relationship represents the call sequence and input-output dependency relationships between the multiple target business systems. The identification information of the multiple target business systems is used to uniquely identify specific information of each target business system. This can include information such as the business system name, ID, and address, which is used to identify and distinguish different business systems in the relationship generator.

[0062] The association relationship between multiple target sub-intentions refers to the degree of mutual connection and dependency between multiple target sub-intentions, which is used to determine whether they need to be processed in parallel or whether there is a dependency relationship. This relationship can be parallel or coupled.

[0063] A parallel relationship means that there is no direct dependency between multiple target sub-intentions, and they can be executed independently. In this case, a parallel calling logic relationship is formed between multiple target business systems, and it is determined that there is no collaborative relationship between multiple target business systems, that is, it is considered that these target business systems can be called at the same time, and there is no specific execution order. Under a parallel calling logic relationship, the work of each target business system is independent and can be executed in parallel. For example, assuming that the target sub-intentions are "query order quantity" and "query inventory quantity", there is no dependency relationship between these two sub-intentions. After analysis, the relationship generator determines that they are in a parallel relationship, so the order management system and inventory management system can be called at the same time, forming a parallel calling logic relationship.

[0064] A coupling relationship refers to the existence of a direct dependency between multiple target sub-intents, where the output of one sub-intent may be the input of another. In this case, a coupled calling logic relationship is formed between multiple target business systems, meaning that these target business systems need to be called in a certain order. Under a coupled calling logic relationship, the calling order and input-output dependencies of the target business systems are clearly specified. For example, suppose the target sub-intents are "query order quantity" and "generate sales report", where "generate sales report" needs to depend on the result of "query order quantity". After analysis, the relationship generator determines that they are coupled, so it first calls the order management system to obtain the order quantity, and then calls the sales statistics system to generate the sales report, forming a coupled calling logic relationship.

[0065] In an optional embodiment, based on the dependency and mutual exclusion relationships between multiple business systems, a DAG (directed acyclic graph) can be used to construct the call logic relationship between multiple target business systems. Specifically, an empty DAG is first initialized, and then the identification information of multiple target business systems is added as nodes to the DAG; the connection method between the nodes is determined based on the dependency and mutual exclusion relationships between the target business systems. If a dependency relationship exists, a directed edge is added to represent the call order; if a mutual exclusion relationship is encountered, a specific scheduling strategy is used to ensure mutual exclusion execution. The loops in the DAG are then checked and processed to eliminate possible circular dependencies. Finally, the call order of the target business systems is determined based on the topological sorting of the DAG. This method can not only clearly display the complex relationships between multiple target business systems and ensure the correctness of the call order, but also realize parallel calls of unrelated target business systems, thereby improving overall efficiency.

[0066] In an alternative embodiment, based on Figure 2 The business call engine in the business system call layer I in the access control model shown calls multiple target business systems according to the call logic relationship, access path information and multiple target sub-intentions between the multiple target business systems to obtain access result information returned by the multiple target business systems, including: inputting the coupled call logic relationship, access path information and multiple target sub-intentions into the business call engine in the business system call layer I in the access control model, and calling the multiple target business systems according to the call order in the coupled call logic relationship; during the calling process, the corresponding target sub-intention is used as the input information of the called target business system, and it is determined whether to use the output of the previous target business system as the input information of the called target business system according to the input-output relationship, so that the called target business system outputs access result information adapted to the corresponding target sub-intention.

[0067] The business call engine is a key component in the access control model, located in Business System Call Layer I. It is responsible for actually executing the call operation on the target business system based on the call logic relationship, access path information, and target sub-intent to obtain the required access result information. In the aforementioned embodiment, the call logic relationship between multiple target business systems was established based on the dependencies and mutual exclusion relationships between the multiple business systems. This step is also performed by the business call engine to ensure the correctness of the call sequence and input and output dependencies.

[0068] Among them, the adapted access result information refers to the output result of the called target business system matching the requirements of the corresponding target sub-intent. The business call engine can determine the specific data to be obtained or the specific operations to be performed in the target business system based on the business entities marked in the target sub-intent. During the call process, the business call engine uses the corresponding target sub-intent as the input information of the called target business system. If there is an input-output dependency, the engine will also use the output result of the previous target business system as the input information of the called target business system. This ensures that the call of each target business system is based on the correct context and data.

[0069] For example, suppose the target sub-intents are "Query Order Status" and "Query Logistics Status." "Query Order Status" must be executed first, and its output (order status) may include a logistics tracking number, which is used in the subsequent "Query Logistics Status" operation. Based on the coupled call logic, the business call engine determines that "Query Order Status" should be called before "Query Logistics Status." The access path for querying order status is https: / / orders.example.com / status, and the access path for querying logistics status is https: / / logistics.example.com / tracking. The target sub-intent "Query Order Status" is annotated with order number 123456, and the target sub-intent "Query Logistics Status" is annotated with logistics tracking number LK789012. The call process is as follows: first, the order management system is called, order number 123456 is entered, and the order status (assuming it is "Shipped") and logistics tracking number LK789012 are obtained. Then, the logistics system is called, logistics tracking number LK789012 is entered, and the logistics status (assuming it is "In Transit, Estimated Delivery within 3 Days") is checked. This approach enables more efficient processing of calls to multiple target business systems, ensuring that each call is based on the correct context and data, thereby improving overall response speed and accuracy. The execution of the business call engine enhances the system's flexibility and scalability, enabling more effective management and execution of complex business processes and improving user experience.

[0070] like Figure 2The figure shows an example of the internal implementation structure of the information generation model. In this example, the information generation model includes: a data input layer J, a rule guidance layer K, a semantic control layer L and an information output layer M. Based on the internal implementation structure of the information generation model, a method for integrating the call results returned by multiple target business systems, generating target answer information corresponding to the question description information, and outputting the target answer information includes the following specific implementation steps: generating a third prompt word based on the call results returned by the multiple target business systems, inputting the third prompt word into the information generation model implemented based on the large language model, and under the guidance of the third prompt word, inputting the call results returned by the multiple target business systems into the data input layer J for standardization processing, and applying conflict decision to the results of the standardization processing to obtain a conflict-free structured data set; inputting the conflict-free structured data set into the rule guidance layer K, performing semantic reorganization operations on the conflict-free structured data set, and generating natural language with business logic constraints and standardization as initial answer information; inputting the initial answer information into the semantic control layer L to adapt the initial answer information to user preferences, so as to bind the user's historical interaction features to the initial answer information, and associating the initial answer information bound to the user's historical interaction features with the key semantic elements in the question description information, so as to output the target answer information in the information output layer M.

[0071] Among them, the third prompt word is generated based on the call results returned by multiple target business systems. As the input instruction of the information generation model implemented based on the large language model, it not only includes the call results returned by multiple target business systems, but also includes guiding information such as standardized templates and conflict decision rules. Among them, the information generation model is an advanced model based on AI, which has the ability of data integration, semantic reorganization and personalized adaptation. For example, the information generation model can be a generative model based on the large language model. Its core principle is to train a large amount of text data through deep learning technology, so as to have a deep understanding and generation ability of language. After receiving the third prompt word, the information generation model can standardize and resolve conflicts of the call results according to the guiding information in the third prompt word, generate a conflict-free structured data set, further semantically reorganize and adapt the conflict-free structured data set to user preferences, and finally output the target answer information corresponding to the question description information.

[0072] Specifically, under the guidance of the third prompt word, the call results returned by multiple target business systems are input into the data input layer J. Standardization processing is based on standardized templates, and structural conversion is implemented for heterogeneous call results returned by multiple business systems. The standardized template in the third prompt word contains basic fields (such as intent label, task ID, processing time) and business fields (such as pay_status and pay_order_id in payment scenarios, and dispatch_status and dispatch_address in dispatch scenarios). Standardization processing converts call results in different formats into a unified format, such as unifying the timestamp into a specified format, and translating the status codes of different business systems into a unified user language description. This process ensures that all data enters the subsequent processing stage with a consistent structure, laying the foundation for generating accurate target answer information. For example, (1) the fields of the payment system JSON response {"status":"success","order":"OD123"} are directly extracted; (2) the physical system response "the product has been shipped from the Shanghai warehouse" can be vectorized through BERT and matched to the logistics template field dispatch_status="in_transit", thus achieving semantic alignment of unstructured data.

[0073] Furthermore, in the data input layer J, the results of the standardization processing are applied to conflict decision-making to obtain a conflict-free structured data set. Conflict decision-making refers to the process of resolving data contradictions in the prompt words according to preset rules. For example, conflict decision-making can be defined as query>payment>refund>fulfillment>settlement>service card>stored-value card>dispatch>others. At the same time, the call result with a newer timestamp has a higher priority. When the call results are contradictory, such as the payment interface returns "paid" and the dispatch interface returns "unpaid pending dispatch", the system will select the trusted data source (here is the payment interface) according to priority, and record the conflict log for subsequent model optimization. Among them, a conflict-free structured data set is a collection with a unified format and data consistency. For example, a conflict-free structured data set can contain information such as order quantity and total sales, and all information is consistent, providing a reliable basis for the subsequent generation of target answer information.

[0074] At the rule-guided layer K, semantic reorganization transforms the conflict-free dataset into natural language that conforms to business logic, generating an initial answer. Semantic reorganization can be achieved using a pre-trained syntax tree. For example, given the conflict-free input dataset of {"order_id":"OD123","ship_time":"2023-06-3010:30","tracking_no":"SF123","delivery_eta":"July 2nd"}, semantic reorganization produces the following initial answer: "Your order OD123 was shipped at 10:30 on June 30th, with tracking number SF123, and is expected to arrive on July 2nd."

[0075] At the semantic control layer L, the information generation model can adapt the initial answer information to user preferences, that is, bind the user's historical interaction features to the initial answer information. For example, if the user prefers concise answers, the generated answer will be more concise and clear. This adaptation is based on the user's past interaction data, such as common vocabulary, answer style, etc., so that the answer is more in line with the user's expression habits and improves the user experience. Furthermore, the information generation model can associate the initial answer information bound to the user's historical interaction features with the key semantic elements in the question description information. Among them, the key semantic elements are the core content in the question description information, such as "a certain product", "a certain region", etc. By analyzing the question description information, extracting these key elements, and associating them with the initial answer information, it is ensured that the answer is closely related to the question description information.

[0076] Finally, the information output layer M generates a target answer corresponding to the question description. For example, if the question description is "How much did order OD123 cost, and where is it?", the output target answer might be "The payment amount for order OD123 is 299 yuan, the current logistics status is in transit, and delivery is expected within 3 days." This target answer integrates the call results of multiple target business systems. After standardization, conflict resolution, semantic reorganization, and user preference adaptation, it is both accurate and consistent with user expression habits, effectively improving the user experience.

[0077] Optionally, after generating the initial answer information, the information generation model can also perform dirty data processing on the initial answer information to ensure data quality and answer reliability. Dirty data processing includes filling missing values ​​and detecting outliers in the initial answer information. Missing value filling refers to the process of completing missing portions of the initial answer information. Missing value filling can be implemented by invoking a rules engine; machine learning models trained on historical data can also be used to predict and fill missing values. Outlier detection refers to the process of identifying abnormal information in the initial answer information that does not conform to expected patterns or regularities. Outlier detection can be generated based on the business entity extraction result interface input parameters and apply the 3σ rule for judgment. The 3σ rule states that if the value of information in the initial answer information exceeds the mean plus three standard deviations (μ+3σ) or falls below the mean minus three standard deviations (μ-3σ), the data point is considered an outlier and can be eliminated.

[0078] For example, the payment amount of an order is missing in the initial answer information. In this case, the missing payment status can be marked as "pending payment" by default through the rule engine, and a machine learning model trained based on historical data can be used to predict the payment amount of the order. For another example, the payment amount of an order in the initial answer information is "-50 yuan", which is obviously not in line with the actual situation. Assuming that the mean payment amount of all normal orders is 100 yuan and the standard deviation is 20 yuan, the normal range should be (100-3×20, 100+3×20), or (40, 160) yuan. The payment amount of "-50 yuan" exceeds this range, so it is identified as an outlier and eliminated, and at the same time triggers an intent clarification message such as "an error occurred" to prompt the user.

[0079] It should be noted that in the above embodiments, a large language model refers to a model whose model parameters meet the set parameter quantity requirements, and there is no limitation on the parameter quantity requirements. Different definitions can be used in different scenarios and fields. For example, in some scenarios or fields, a large language model refers to a model with a parameter scale of tens of billions, hundreds of billions, or even trillions. The definition of the model parameter quantity is only an example.

[0080] The detailed implementation and beneficial effects of each step in the method of this embodiment have been described in detail in the aforementioned embodiments and will not be elaborated here.

[0081] In addition, some of the processes described in the above embodiments and the accompanying drawings include multiple operations that appear in a specific order, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0082] Figure 3 This is a schematic diagram of an electronic device structure provided by an exemplary embodiment of the present application. Figure 3 As shown, the electronic device includes: a memory 34 and a processor 35.

[0083] The memory 34 is used to store computer programs and can be configured to store various other data to support operations on the electronic device. Examples of such data include instructions for any application or method operating on the electronic device, problem description information, tag sequence, access result information, etc.

[0084] The memory 34 may be implemented by any type of volatile or non-volatile memory device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0085] The processor 35 is coupled to the memory 34 and is used to execute the computer program in the memory 34, so as to: receive problem description information, the problem description information is described in natural language, and the problem description information is related to at least one business system among multiple business systems; generate a first prompt word according to the problem description information, input the first prompt word into a dynamic intent recognition model based on a large language model, and under the guidance of the first prompt word, convert the problem description information into a tag sequence, and divide the tag sequence into multiple tag segments with independent semantics; perform mixed intent recognition on the tag sequence to obtain multiple first candidate sub-intentions and their probability distributions; perform single intent recognition on the multiple tag segments to obtain multiple second candidate sub-intentions and their probability distributions; and perform mixed intent recognition on the multiple tag segments to obtain multiple second candidate sub-intentions and their probability distributions according to the multiple first candidate sub-intentions and their probability distributions and the multiple second candidate sub-intentions and their probability distributions. Probability distribution, generate multiple target sub-intentions; generate a second prompt word according to the multiple target sub-intentions, input the second prompt word into the access control model implemented based on the large language model, and under the guidance of the second prompt word, based on the existing sub-intentions and the functional characteristics of multiple business systems, determine the multiple target business systems and their access path information corresponding to the multiple target sub-intentions; according to the dependency and mutual exclusion relationship between the multiple business systems, construct the calling logic relationship between the multiple target business systems; according to the calling logic relationship, access path information and multiple target sub-intentions between the multiple target business systems, call the multiple target business systems to obtain the access result information returned by the multiple target business systems; integrate the calling results returned by the multiple target business systems, generate the target answer information corresponding to the question description information, and output the target answer information.

[0086] In an optional embodiment, the processor 35 includes a word segmenter in the dynamic intent recognition model, converts the problem description information into a token sequence, and divides the token sequence into multiple token segments with independent semantics, including: using the word segmenter to segment the problem description information, and performing punctuation normalization on the segmentation results to obtain a token sequence; matching each word in the token sequence with a semantic segmentation symbol in a predefined rule template, using the word segmentation position of the semantic segmentation symbol in the match as the segmentation position, and semantically segmenting the token sequence according to the segmentation position to obtain multiple token segments.

[0087] In an optional embodiment, the dynamic intent recognition model includes a multi-layer encoding network using an attention mechanism, an entity extraction model, and an intent classifier implemented based on a large language model. The processor 35 performs mixed intent recognition on the tag sequence to obtain multiple first candidate sub-intentions and their probability distributions; and performs single intent recognition on multiple tag fragments to obtain multiple second candidate sub-intentions and their probability distributions, including: inputting the tag sequence and the multiple tag fragments into the multi-layer encoding network using the attention mechanism for encoding processing to obtain a first feature vector and multiple second feature vectors with context information; inputting the first feature vector and the multiple second feature vectors into the entity extraction model for business entity identification and labeling to obtain a third feature vector and multiple fourth feature vectors with business entity labeling information; inputting the third feature vector and the multiple fourth feature vectors into the intent classifier, performing mixed intent recognition on the third feature vector to obtain multiple first candidate sub-intentions, performing single intent recognition on the multiple fourth feature vectors to obtain multiple second candidate sub-intentions, and applying an activation function to calculate the probability distribution of each of the multiple first candidate sub-intentions and the multiple second candidate sub-intentions.

[0088] In an optional embodiment, the processor 35 generates multiple target sub-intentions based on multiple first candidate sub-intentions and their probability distributions and multiple second candidate sub-intentions and their probability distributions, including: selecting multiple first valid sub-intentions that meet the first probability distribution conditions from the multiple first candidate sub-intentions based on the probability distribution of the multiple first candidate sub-intentions; selecting multiple second valid sub-intentions that meet the second probability distribution conditions from the multiple second candidate sub-intentions based on the probability distribution of the multiple second candidate sub-intentions; obtaining the intersection of the first valid sub-intention and the second valid sub-intention as the third valid sub-intention; calculating the semantic similarity of any two third valid sub-intentions, taking the two third valid sub-intentions whose semantic similarity is less than the first semantic similarity threshold as independent target sub-intentions, and merging the two third valid sub-intentions whose semantic similarity is greater than the second semantic similarity threshold into one target sub-intention; the first semantic similarity threshold is less than the second semantic similarity threshold.

[0089] In an optional embodiment, the access control model is associated with a knowledge base, which maintains the correspondence between existing sub-intentions and business systems, access path information required to access business systems, and functional characteristics of multiple business systems; the access control model includes a routing matching model implemented based on a large language model, and based on the existing sub-intentions and functional characteristics of multiple business systems, the processor 35 determines multiple target business systems and their access path information corresponding to multiple target sub-intentions, including: inputting multiple target sub-intentions into the routing matching model, and for each target sub-intention, matching the target sub-intention in the correspondence maintained in the knowledge base; if there is already a sub-intention in the match, the matching The business systems and access path information corresponding to the existing sub-intentions in the target business system and its access path information are respectively used as the target business systems and access path information corresponding to the target sub-intention; if there is no match in the existing sub-intention, calculate the semantic similarity between the target sub-intention and the functional characteristics of multiple business systems, and use the business system with the highest semantic similarity to the target sub-intention as the target business system corresponding to the target sub-intention; dynamically generate the access path information required to access the target business system based on the access address, access interface and interface parameters of the target business system; and use the target sub-intention as the existing sub-intention, and update the target sub-intention, the target business system corresponding to the target sub-intention and its access path information to the corresponding relationship.

[0090] In an optional embodiment, the access control model also includes a relationship generator; the processor 35 constructs a calling logical relationship between multiple target business systems based on the dependency and mutual exclusion relationship between the multiple business systems, including: inputting the identification information of the multiple target business systems and the multiple target sub-intentions into the relationship generator, and analyzing the association relationship between the multiple target sub-intentions; if the multiple target sub-intentions are in a parallel relationship, a parallel calling logical relationship is formed between the multiple target business systems; if the multiple target sub-intentions are in a coupled relationship, based on the identification information of the multiple target business systems and the dependency and mutual exclusion relationship between the multiple business systems, a coupled calling logical relationship is formed between the multiple target business systems, and the coupled calling logical relationship represents the calling sequence and input-output dependency relationship between the multiple target business systems.

[0091] In an optional embodiment, the access control model includes a business call engine; the processor 35 calls multiple target business systems according to the call logic relationship, access path information and multiple target sub-intentions between the multiple target business systems to obtain access result information returned by the multiple target business systems, including: inputting the coupled call logic relationship, access path information and multiple target sub-intentions into the business call engine, and calling the multiple target business systems according to the call order in the coupled call logic relationship; during the call process, the corresponding target sub-intention is used as the input information of the called target business system, and it is determined based on the input-output relationship whether the output of the previous target business system is also used as the input information of the called target business system, so that the called target business system outputs access result information adapted to the corresponding target sub-intention.

[0092] In an optional embodiment, the processor 35 integrates the call results returned by multiple target business systems, generates target answer information corresponding to the question description information, and outputs the target answer information, including: generating a third prompt word based on the call results returned by the multiple target business systems, inputting the third prompt word into an information generation model implemented based on a large language model, and under the guidance of the third prompt word, standardizing the call results returned by the multiple target business systems, and applying conflict decision to the results of the standardization to obtain a conflict-free structured data set; and performing a semantic reorganization operation on the conflict-free structured data set to generate a standardized natural language with business logic constraints as the initial answer information; performing user preference adaptation on the initial answer information to bind the user's historical interaction features to the initial answer information, and associating the initial answer information bound to the user's historical interaction features with the key semantic elements in the question description information to output the target answer information.

[0093] Further, if Figure 3 As shown, the electronic device also includes: a communication component 36, a display 37, a power component 38, an audio component 39 and other components. Figure 3 Only some components are shown schematically, which does not mean that the electronic device only includes Figure 3 Components shown.

[0094] Accordingly, an embodiment of the present application further provides a computer-readable storage medium storing a computer program, which, when executed, can implement the steps that can be performed by the electronic device in the above method embodiment.

[0095] above Figure 3The communication component is configured to facilitate wired or wireless communication between the device where the communication component is located and other devices. The device where the communication component is located can access a wireless network based on a communication standard, such as WiFi, 2G, 3G, 4G / LTE, 5G and other mobile communication networks, or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0096] above Figure 3 The display in the embodiment includes a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touches, slides, and gestures on the touch panel. The touch sensor may not only sense the boundaries of a touch or slide action, but also detect the duration and pressure associated with the touch or slide operation.

[0097] above Figure 3 The power supply component in a device provides power to various components of the device in which the power supply component is located. The power supply component may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device in which the power supply component is located.

[0098] above Figure 3 The audio component in the device may be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC), and when the device where the audio component is located is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is configured to receive an external audio signal. The received audio signal may be further stored in a memory or sent via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0099] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0100] The present application is described with reference to the flow chart and / or block diagram of the method, device (system), and computer program product according to the embodiment of the present application. It should be understood that each flow process and / or box in the flow chart and / or block diagram and the combination of the flow process and / or box in the flow chart and / or block diagram can be realized by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processing machine or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for realizing the function specified in one flow chart flow or multiple flows and / or one box or multiple boxes of the block diagram.

[0101] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0102] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0103] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0104] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0105] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0106] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0107] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A multi-system access control method based on dynamic intent, characterized in that: include: receiving problem description information, where the problem description information is described in a natural language and is related to at least one business system among the plurality of business systems; Generate a first prompt word based on the question description information, input the first prompt word into a dynamic intent recognition model implemented based on a large language model, convert the question description information into a token sequence under the guidance of the first prompt word, and segment the token sequence into multiple token segments with independent semantics; perform mixed intent recognition on the token sequence to obtain multiple first candidate sub-intents and their probability distributions; Performing single intent recognition on each of the multiple marked segments to obtain multiple second candidate sub-intents and their probability distributions; generating multiple target sub-intents based on the multiple first candidate sub-intents and their probability distributions and the multiple second candidate sub-intents and their probability distributions; Generating a second prompt word based on the multiple target sub-intents, inputting the second prompt word into an access control model implemented based on a large language model, and under the guidance of the second prompt word, determining multiple target business systems corresponding to the multiple target sub-intents and their access path information based on the existing sub-intents and the functional characteristics of the multiple business systems; and constructing a call logic relationship between the multiple target business systems based on the dependencies and mutual exclusion relationships between the multiple business systems. Calling the multiple target business systems according to the calling logical relationship between the multiple target business systems, the access path information, and the multiple target sub-intentions to obtain access result information returned by the multiple target business systems; The call results returned by the multiple target business systems are integrated to generate target answer information corresponding to the problem description information, and the target answer information is output.

2. The method according to claim 1, characterized in that The dynamic intent recognition model includes a word segmenter that converts the question description information into a token sequence and segments the token sequence into multiple token segments with independent semantics, including: Using the word segmenter to perform word segmentation processing on the problem description information, and performing punctuation normalization processing on the word segmentation result to obtain a token sequence; Each word segment in the tag sequence is matched with a semantic segmentation symbol in a predefined rule template, the word segmentation position of the semantic segmentation symbol in the match is used as a segmentation position, and the tag sequence is semantically segmented according to the segmentation position to obtain the multiple tag fragments.

3. The method according to claim 1, characterized in that The dynamic intent recognition model includes a multi-layer encoding network using an attention mechanism, an entity extraction model, and an intent classifier implemented based on a large language model, and performs hybrid intent recognition on the tag sequence to obtain multiple first candidate sub-intents and their probability distributions; Performing single intent recognition on the multiple marked segments to obtain multiple second candidate sub-intents and their probability distributions, including: Inputting the tag sequence and the multiple tag segments into a multi-layer encoding network using an attention mechanism for encoding processing to obtain a first feature vector and multiple second feature vectors with context information; Inputting the first feature vector and the plurality of second feature vectors into the entity extraction model respectively to identify and label business entities, so as to obtain a third feature vector and a plurality of fourth feature vectors with business entity labeling information; The third eigenvector and multiple fourth eigenvectors are respectively input into the intent classifier, mixed intent recognition is performed on the third eigenvector to obtain multiple first candidate sub-intentions, single intent recognition is performed on the multiple fourth eigenvectors to obtain multiple second candidate sub-intentions, and an activation function is applied to calculate the probability distribution of the multiple first candidate sub-intentions and the multiple second candidate sub-intentions.

4. The method according to any one of claims 1 to 3, characterized in that Generate multiple target sub-intents according to the multiple first candidate sub-intents and their probability distributions and the multiple second candidate sub-intents and their probability distributions, including: According to the probability distribution of the plurality of first candidate sub-intentions, selecting a plurality of first valid sub-intentions that meet a first probability distribution condition from the plurality of first candidate sub-intentions; Selecting, according to the probability distribution of the plurality of second candidate sub-intentions, a plurality of second valid sub-intentions that meet a second probability distribution condition from the plurality of second candidate sub-intentions; Obtain the intersection of the first valid sub-intention and the second valid sub-intention as the third valid sub-intention; calculate the semantic similarity of any two third valid sub-intentions, take the two third valid sub-intentions whose semantic similarity is less than the first semantic similarity threshold as independent target sub-intentions, and merge the two third valid sub-intentions whose semantic similarity is greater than the second semantic similarity threshold into one target sub-intention; the first semantic similarity threshold is less than the second semantic similarity threshold.

5. The method according to any one of claims 1 to 3, characterized in that The access control model is associated with a knowledge base, wherein the knowledge base maintains correspondences between existing sub-intentions and business systems, access path information required to access the business systems, and functional characteristics of the multiple business systems; The access control model includes a routing matching model implemented based on a large language model, and determines multiple target business systems corresponding to the multiple target sub-intents and their access path information based on the existing sub-intents and the functional characteristics of the multiple business systems, including: Inputting the multiple target sub-intents into the routing matching model, and for each target sub-intent, matching the target sub-intent in the corresponding relationship maintained in the knowledge base; If there is a sub-intention in the match, the business system and its access path information corresponding to the sub-intention in the match will be used as the target business system and its access path information corresponding to the target sub-intention; If there is already a sub-intention that is not matched, calculate the semantic similarity between the target sub-intention and the functional characteristics of the multiple business systems, and use the business system with the highest semantic similarity to the target sub-intention as the target business system corresponding to the target sub-intention; dynamically generate the access path information required to access the target business system based on the access address, access interface and interface parameters of the target business system; and use the target sub-intention as an existing sub-intention, and update the target sub-intention, the target business system corresponding to the target sub-intention and its access path information into the corresponding relationship.

6. The method according to any one of claims 1 to 3, characterized in that The access control model also includes a relationship generator; Constructing a call logic relationship between the multiple target business systems based on the dependencies and mutual exclusion relationships between the multiple business systems, including: Inputting the identification information of the multiple target business systems and the multiple target sub-intents into the relationship generator, and analyzing the association relationship between the multiple target sub-intents; If the multiple target sub-intentions are in a parallel relationship, a parallel calling logical relationship is formed between the multiple target business systems; if the multiple target sub-intentions are in a coupled relationship, based on the identification information of the multiple target business systems and the dependencies and mutual exclusion relationships between the multiple business systems, a coupled calling logical relationship is formed between the multiple target business systems, and the coupled calling logical relationship represents the calling sequence and input-output dependency relationship between the multiple target business systems.

7. The method according to claim 6, characterized in that The access control model includes a service call engine; Calling the multiple target business systems according to the calling logical relationship between the multiple target business systems, the access path information, and the multiple target sub-intents to obtain access result information returned by the multiple target business systems, including: Inputting the coupled call logic relationship, access path information and the multiple target sub-intents into the business call engine, and calling the multiple target business systems according to the calling order in the coupled call logic relationship; During the calling process, the corresponding target sub-intention is used as the input information of the called target business system, and based on the input-output relationship, it is determined whether the output of the previous target business system is also used as the input information of the called target business system, so that the called target business system outputs access result information that is adapted to the corresponding target sub-intention.

8. The method according to claim 1, characterized in that Integrating the call results returned by the multiple target business systems, generating target answer information corresponding to the problem description information, and outputting the target answer information, including: A third prompt word is generated based on the call results returned by the multiple target business systems, and the third prompt word is input into an information generation model implemented based on a large language model. Under the guidance of the third prompt word, the call results returned by the multiple target business systems are standardized, and the results of the standardized processing are applied to conflict decision-making to obtain a conflict-free structured data set; and a semantic reorganization operation is performed on the conflict-free structured data set to generate a standardized natural language with business logic constraints as initial answer information; the initial answer information is adapted to user preferences to bind user historical interaction features to the initial answer information, and the initial answer information bound to the user historical interaction features is associated with key semantic elements in the question description information to output the target answer information.

9. An electronic device, characterized in that: include: A processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is enabled to implement the steps in the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is enabled to implement the steps of the method according to any one of claims 1 to 8.

11. A computer program product, characterized in that The method comprises a computer program / instruction, which, when executed by a processor, enables the processor to implement the steps of the method according to any one of claims 1 to 8.

Images