Cross-domain agent accurate scheduling method and system

Abstract

The invention relates to the field of artificial intelligence, in particular to a cross-domain agent accurate scheduling method and system. The method comprises the following steps: analyzing an intelligent agent capability description document, extracting a label, coding and storing; performing intention analysis on the user input to generate an intention vector; performing semantic matching on the intention vector and the tag vector to screen candidate agents; calculating a comprehensive scheduling score based on the real-time performance index and the historical calling record; and a target agent is selected and scheduling is executed. According to the invention, the accuracy and efficiency of agent scheduling are improved, and the adaptability of the system to the user demand is enhanced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for precise scheduling of cross-domain intelligent agents. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent agents have been widely applied in various fields, including intelligent customer service, intelligent assistants, and autonomous driving. The diversity of intelligent agents provides users with a wealth of service choices, but it also presents the challenge of accurately scheduling the most suitable agent for each user's needs. An intelligent agent scheduling system needs to select the most appropriate agent from among many to execute tasks based on the user's specific requirements, thereby improving user experience and system efficiency. Traditional intelligent agent scheduling is mainly based on simple keyword matching or preset rules. However, with the continuous enrichment of intelligent agent types and capabilities, this approach can no longer meet the needs of precise scheduling.

[0003] Existing technologies for agent scheduling suffer from the following shortcomings: First, existing scheduling methods lack multi-dimensional representations of agent capabilities, often focusing only on single or limited-dimensional capability descriptions. This leads to an inability to comprehensively assess agent suitability, particularly in cross-domain scenarios where scheduling accuracy significantly decreases. Second, traditional scheduling methods typically employ simple keyword matching or fixed rules, lacking deep semantic understanding of user input. This prevents accurate capture of potential user needs and contextual information, resulting in discrepancies between scheduling results and actual user expectations. Third, existing scheduling methods rarely consider agent real-time performance and user historical preferences, relying solely on static capability descriptions for matching. This ignores differences in agent performance during actual operation and personalized user needs, making it difficult to achieve optimal scheduling in dynamic environments. Summary of the Invention

[0004] This invention provides a method and system for precise scheduling of cross-domain intelligent agents, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for precise scheduling of cross-domain intelligent agents, comprising:

[0006] For the connected intelligent agent, by parsing the intelligent agent's capability description document, multiple dimensions of tags are extracted, and the tags are encoded and stored in a tag library to obtain a standardized multi-dimensional capability tag vector;

[0007] The system receives user input, performs intent parsing on the input, and extracts core requirement information, related scenario information, and data dependency information to generate an intent vector.

[0008] The intent vector is semantically matched with the multi-dimensional capability tag vectors in the tag library, the similarity between the vectors is calculated, and a set of candidate agents whose similarity meets a preset threshold is selected.

[0009] For each agent in the candidate agent set, its real-time performance metrics and historical call records are obtained. A performance score is calculated based on the real-time performance metrics, a user preference score is calculated based on the historical call records, and a comprehensive scheduling score is calculated based on the similarity, performance score, and user preference score.

[0010] Based on the comprehensive scheduling score, a target agent is selected from the set of candidate agents, a scheduling instruction is sent to the target agent, and the execution result is determined.

[0011] For the connected agent, by parsing the agent's capability description document, multiple dimensions of tags are extracted, and these tags are encoded and stored in a tag library to obtain a standardized multi-dimensional capability tag vector, including:

[0012] Obtain the capability description document of the intelligent agent, perform text preprocessing on the capability description document, remove redundant characters and segment sentences to obtain a set of standardized text fragments;

[0013] Semantic encoding is performed on each text segment in the standardized text segment set, and the text segment is converted into a semantic feature vector;

[0014] Based on the semantic feature vector and the pre-stored tag vectors in the tag library, the corresponding tags are automatically labeled to the agent to obtain preliminary labeling results;

[0015] The preliminary annotation results are validated for dimensions to determine whether they cover multiple dimensions simultaneously. If any dimensions are missing, the description information corresponding to the missing dimensions is extracted from the capability description document, and the corresponding dimension labels are supplemented based on the description information.

[0016] After dimensional verification, the labels are encoded according to preset encoding rules to generate unique identifiers. The unique identifiers are then associated with the semantic feature vectors of the corresponding labels and stored in the label library to obtain the multi-dimensional capability label vectors of the agent.

[0017] The preliminary annotation results are validated for dimensions to determine whether they cover multiple dimensions simultaneously. If any dimensions are missing, the description information corresponding to the missing dimensions is extracted from the capability description document, and the corresponding dimension labels are supplemented based on the description information, including:

[0018] Iterate through all labels in the preliminary annotation results, and based on the dimension attribute identifier of each label, count the dimension types covered by the preliminary annotation results to obtain the set of covered dimensions;

[0019] The covered dimension set is compared with the preset complete dimension set to identify the missing dimensions that have not been covered, and the dimension type of the missing dimensions is recorded.

[0020] For each dimension type of the missing dimension, a set of feature keywords corresponding to the dimension type is obtained from a preset dimension keyword library;

[0021] Retrieve text fragments containing keywords from the feature keyword set from the capability description document, extract the text fragments as descriptive information corresponding to the missing dimension, and perform semantic encoding on the descriptive information corresponding to the missing dimension to obtain the semantic feature vector of the missing dimension;

[0022] Based on the semantic feature vector of the missing dimension, the label attribute features corresponding to the missing dimension are identified through semantic parsing, and candidate labels belonging to the missing dimension type are selected from the label library according to the label attribute features. The candidate labels are added to the preliminary annotation results to obtain the dimension-complete annotation results.

[0023] Receive user input, perform intent parsing on the user input, extract core requirement information, related scenario information, and data dependency information to generate an intent vector, including:

[0024] Receive user input, perform text normalization processing on the user input, remove invalid characters and perform sentence segmentation to obtain standardized input text;

[0025] The standardized input text is parsed to identify the subject-verb-object components and modification relationships, thereby obtaining a grammatical structure tree.

[0026] Based on the grammar structure tree, the core phrases expressing the action target in the standardized input text are located, and the core phrases are extracted as core requirement information;

[0027] Based on the grammar structure tree, the scene description phrases that describe the application background and the dependency description phrases that describe the input and output constraints in the standardized input text are identified and extracted as related scene information and data dependency information, respectively.

[0028] The core requirement information and the associated scenario information are vector-concatenated according to a preset fusion strategy to generate an initial intent representation.

[0029] The data dependency information is added as a constraint to the initial intent representation to obtain the intent vector.

[0030] The intent vector is semantically matched with multi-dimensional capability tag vectors in the tag library, the similarity between the vectors is calculated, and a set of candidate agents whose similarity meets a preset threshold is selected, including:

[0031] The intent vector is decomposed to obtain demand dimension features and scenario dimension features;

[0032] For each agent's multi-dimensional capability label vector, extract the capability requirement dimension component corresponding to the requirement dimension feature, and the capability scene dimension component corresponding to the scene dimension feature from the multi-dimensional capability label vector;

[0033] Calculate the demand similarity between the demand dimension features and the capability demand dimension components, and the scenario similarity between the scenario dimension features and the capability scenario dimension components, respectively.

[0034] Based on preset dimensional weight coefficients, the demand similarity and the scene similarity are weighted and fused to obtain the comprehensive similarity between the intent vector and the multi-dimensional capability tag vector;

[0035] The overall similarity is compared with a preset threshold to select agents whose overall similarity meets the preset threshold, and the selected agents are added to the candidate agent set.

[0036] The agents in the candidate agent set are arranged in order according to the comprehensive similarity, resulting in a candidate agent set sorted by matching degree.

[0037] For each agent in the candidate agent set, its real-time performance metrics and historical call records are obtained. A performance score is calculated based on the real-time performance metrics, a user preference score is calculated based on the historical call records, and a comprehensive scheduling score is calculated based on the similarity, performance score, and user preference score, including:

[0038] For each agent in the candidate agent set, obtain the agent's real-time performance metrics;

[0039] Based on preset performance index weights, the response latency and task success rate in the real-time performance index are weighted and summed to obtain the performance score of the agent;

[0040] Retrieve the agent's historical call records from the historical call database, and count the number of calls and the number of successful calls related to the current user in the historical call records;

[0041] Based on the number of calls and the number of successful calls, the call frequency and success frequency of the agent in the current user's historical usage are calculated. Based on the call frequency and the success frequency, a comprehensive evaluation is performed according to a preset preference calculation rule to obtain the user preference score of the agent for the current user.

[0042] Based on preset scheduling weight coefficients, the similarity, performance score, and user preference score are weighted and fused to obtain the comprehensive scheduling score of the agent.

[0043] Based on the number of calls and the number of successful calls, the call frequency and success frequency of the agent in the current user's historical usage are calculated. Based on the call frequency and success frequency, a comprehensive evaluation is performed according to preset preference calculation rules to obtain the user preference score of the agent for the current user, including:

[0044] Retrieve the total number of calls made by the current user within a preset time window, as recorded in the historical call database, and use this as the user's call baseline.

[0045] The call frequency of the intelligent agent is obtained by calculating the ratio of the number of times the agent is called to the base number of user calls.

[0046] The success frequency of the agent is obtained by calculating the ratio of the number of successful calls to the total number of calls made by the agent.

[0047] The call frequency is subjected to time decay processing. Based on the time interval between each call in the historical call record and the current time, the contribution value of the corresponding call record in the call frequency is decayed and weighted to obtain the time-weighted call frequency.

[0048] The success frequency is processed by task type differentiation. Based on the task type processed in each successful call in the historical call record, the task type corresponding to the core requirement input by the current user is identified, and the successful call records matching the task type in the historical call record are extracted. The task-related success frequency is recalculated based on the successful call records.

[0049] According to the preset preference calculation rules, the time-weighted call frequency and the task-related success frequency are weighted and fused to obtain the user preference score of the agent for the current user.

[0050] A second aspect of the present invention provides a cross-domain intelligent agent precise scheduling system, comprising:

[0051] The first unit is used to extract multi-dimensional tags from the capability description document of the connected intelligent agent by parsing the intelligent agent's capability description document, and then encode and store the tags in a tag library to obtain a standardized multi-dimensional capability tag vector;

[0052] The second unit is used to receive user input, perform intent parsing on the user input, extract core requirement information, related scenario information, and data dependency information to generate an intent vector;

[0053] The third unit is used to perform semantic matching between the intent vector and the multi-dimensional capability tag vectors in the tag library, calculate the similarity between the vectors, and filter out a set of candidate agents whose similarity meets a preset threshold;

[0054] The fourth unit is used to obtain the real-time performance indicators and historical call records for each agent in the candidate agent set, calculate a performance score based on the real-time performance indicators, calculate a user preference score based on the historical call records, and calculate a comprehensive scheduling score based on the similarity, performance score, and user preference score.

[0055] The fifth unit is used to select a target agent from the candidate agent set based on the comprehensive scheduling score, send a scheduling instruction to the target agent, and determine the execution result.

[0056] A third aspect of the present invention provides an electronic device, comprising:

[0057] processor;

[0058] Memory used to store processor-executable instructions;

[0059] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0060] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0061] This invention provides a cross-domain intelligent agent precise scheduling method. By parsing the intelligent agent capability description document and extracting multi-dimensional tags, it encodes these tags into standardized capability tag vectors, achieving a refined expression of intelligent agent capabilities. This lays the foundation for subsequent precise matching and avoids the problems of single and coarse-grained capability expression in traditional methods. Through deep intent parsing of user input, it not only extracts core requirement information but also focuses on scenario information and data dependency information, generating a comprehensive intent vector. This makes intelligent agent matching more consistent with the user's actual needs and improves matching accuracy. This invention uses semantic matching rather than simple keyword matching. By calculating the similarity between the intent vector and the capability tag vector, it can understand the semantic relationship between user needs and intelligent agent capabilities, effectively overcoming the limitations of traditional keyword matching. In the intelligent agent selection process, this invention considers not only capability matching degree but also real-time performance indicators and historical call records. A multi-dimensional scoring mechanism calculates a comprehensive scheduling score, enabling intelligent agent scheduling to simultaneously consider functional fit, system performance, and user preferences, significantly improving the overall scheduling efficiency. This invention achieves efficient integration and precise allocation of cross-domain intelligent agent resources, improves the success rate of intelligent agent invocation and user satisfaction, and provides a more flexible and reliable intelligent agent scheduling solution for complex and ever-changing application scenarios. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating the cross-domain intelligent agent precise scheduling method according to an embodiment of the present invention.

[0063] Figure 2 This is a schematic diagram illustrating the process of determining the intent vector according to an embodiment of the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0066] Figure 1 This is a flowchart illustrating the cross-domain intelligent agent precise scheduling method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0067] For the connected intelligent agent, by parsing the intelligent agent's capability description document, multiple dimensions of tags are extracted, and the tags are encoded and stored in a tag library to obtain a standardized multi-dimensional capability tag vector;

[0068] The system receives user input, performs intent parsing on the input, and extracts core requirement information, related scenario information, and data dependency information to generate an intent vector.

[0069] The intent vector is semantically matched with the multi-dimensional capability tag vectors in the tag library, the similarity between the vectors is calculated, and a set of candidate agents whose similarity meets a preset threshold is selected.

[0070] For each agent in the candidate agent set, its real-time performance metrics and historical call records are obtained. A performance score is calculated based on the real-time performance metrics, a user preference score is calculated based on the historical call records, and a comprehensive scheduling score is calculated based on the similarity, performance score, and user preference score.

[0071] Based on the comprehensive scheduling score, a target agent is selected from the set of candidate agents, a scheduling instruction is sent to the target agent, and the execution result is determined.

[0072] In one optional implementation, for the accessed agent, by parsing the agent's capability description document, multiple-dimensional tags are extracted, and the tags are encoded and stored in a tag library to obtain a standardized multi-dimensional capability tag vector, including:

[0073] Obtain the capability description document of the intelligent agent, perform text preprocessing on the capability description document, remove redundant characters and segment sentences to obtain a set of standardized text fragments;

[0074] Semantic encoding is performed on each text segment in the standardized text segment set, and the text segment is converted into a semantic feature vector;

[0075] Based on the semantic feature vector and the pre-stored tag vectors in the tag library, the corresponding tags are automatically labeled to the agent to obtain preliminary labeling results;

[0076] The preliminary annotation results are validated for dimensions to determine whether they cover multiple dimensions simultaneously. If any dimensions are missing, the description information corresponding to the missing dimensions is extracted from the capability description document, and the corresponding dimension labels are supplemented based on the description information.

[0077] After dimensional verification, the labels are encoded according to preset encoding rules to generate unique identifiers. The unique identifiers are then associated with the semantic feature vectors of the corresponding labels and stored in the label library to obtain the multi-dimensional capability label vectors of the agent.

[0078] Capability description documents are typically provided by the agent developers and contain information describing the agent's functions, characteristics, and applicable scenarios. For example, for a customer service agent, its capability description document might include phrases such as "able to handle user inquiries," "supports multi-turn dialogues," and "able to recognize emotions and respond accordingly."

[0079] The acquired capability description documents undergo text preprocessing to remove redundant characters and segment sentences, resulting in a standardized set of text fragments. Specifically, regular expressions are used to match and remove special symbols, extra spaces, and other redundant characters. Natural language processing techniques are employed for sentence boundary detection, dividing long texts into semantically complete sentences or phrases. For example, the sentence "The agent can: 1) answer questions; 2) provide suggestions; 3) execute simple instructions" is processed into a standardized set of text fragments such as "[The agent can answer questions, the agent can provide suggestions, the agent can execute simple instructions]".

[0080] Next, each text segment in the standardized text segment set is semantically encoded, transforming it into a semantic feature vector. This step uses a pre-trained language model (such as BERT, GPT, etc.) to extract the semantic features of the text, mapping the text to a high-dimensional vector space. For example, for the text segment "The agent can answer questions," the semantic encoding model transforms it into a 768-dimensional vector [0.12, -0.32, 0.45, ...], which captures the semantic information of the text.

[0081] Based on semantic feature vectors and pre-stored tag vectors in a tag library, corresponding tags are automatically assigned to the agent, yielding preliminary annotation results. The tag library pre-stores tags for different capability dimensions and their corresponding semantic vectors, such as "question-answering ability," "multilingual support," and "sentiment analysis." By calculating the cosine similarity between the semantic vector of a text fragment and the tag vector, when the similarity exceeds a preset threshold (e.g., 0.75), the corresponding tag is assigned to the agent. For the text fragment "the agent can answer questions," the tag "question-answering ability" will be matched.

[0082] The initial annotation results undergo dimensional validation to determine if they simultaneously cover multiple predefined capability dimensions. These dimensions may include basic capabilities, interaction methods, domain expertise, performance metrics, etc. If any dimensions are missing, the corresponding descriptive information is extracted from the capability description document. For example, if the annotation results are found to lack a domain expertise dimension, domain-related keywords or sentences, such as terms from finance, healthcare, and education, are searched in the original document to extract relevant descriptions for supplementary annotation.

[0083] Based on the extracted missing dimension descriptions, corresponding dimension tags are added. These descriptions are also converted into semantic vectors and matched against the corresponding dimension tag vectors in the tag library to find the closest tag for addition. If the match is insufficient, the description is added as a new tag to the tag library. For example, if the description "focused on medical and health consultation" is extracted from a document, the tag for the medical field can be added.

[0084] After dimensional verification, the tags are encoded according to a preset encoding rule to generate a unique identifier. The encoding rule can be in the form of "dimensional code-tag category code-serial number" to ensure that each tag has a unique identifier. For example, "basic ability-question and answer-001" represents the first tag in the question and answer category under the basic ability dimension. In addition, to facilitate subsequent retrieval and comparison, the unique identifier is associated with the semantic feature vector of the corresponding tag and stored in the tag library.

[0085] Finally, the multi-dimensional capability tag vector of the agent is generated. This vector is structured data composed of multiple dimensional tags, which can be represented as [tag_1, tag_2, ..., tag_n], where each tag contains a unique identifier and a corresponding semantic vector. This standardized multi-dimensional capability tag vector can accurately describe the capability characteristics of the agent, providing a foundation for the management, retrieval, and matching of the agent.

[0086] The above methods enable the automated extraction of key information from the agent's capability description document, generating standardized multi-dimensional capability tag vectors. This provides a unified formal representation of the agent's capabilities, facilitating subsequent agent management and application. This representation method not only improves the understandability of agent capabilities but also lays the foundation for capability comparison and collaboration between agents.

[0087] In one optional implementation, the preliminary annotation results are subjected to dimensional verification to determine whether the preliminary annotation results simultaneously cover multiple dimensions. If there are missing dimensions, the description information corresponding to the missing dimensions is extracted from the capability description document, and the corresponding dimension labels are supplemented based on the description information, including:

[0088] Iterate through all labels in the preliminary annotation results, and based on the dimension attribute identifier of each label, count the dimension types covered by the preliminary annotation results to obtain the set of covered dimensions;

[0089] The covered dimension set is compared with the preset complete dimension set to identify the missing dimensions that have not been covered, and the dimension type of the missing dimensions is recorded.

[0090] For each dimension type of the missing dimension, a set of feature keywords corresponding to the dimension type is obtained from a preset dimension keyword library;

[0091] Retrieve text fragments containing keywords from the feature keyword set from the capability description document, extract the text fragments as descriptive information corresponding to the missing dimension, and perform semantic encoding on the descriptive information corresponding to the missing dimension to obtain the semantic feature vector of the missing dimension;

[0092] Based on the semantic feature vector of the missing dimension, the label attribute features corresponding to the missing dimension are identified through semantic parsing, and candidate labels belonging to the missing dimension type are selected from the label library according to the label attribute features. The candidate labels are added to the preliminary annotation results to obtain the dimension-complete annotation results.

[0093] The preliminary annotation results are validated for dimensions to determine whether they cover multiple dimensions simultaneously. If there are missing dimensions, the description information corresponding to the missing dimensions is extracted from the capability description document, and the corresponding dimension labels are supplemented based on the description information.

[0094] First, iterate through all tags in the initial annotation results. Based on the dimensional attribute identifier of each tag, count the dimensional types covered by the initial annotation results to obtain the set of covered dimensions. During this process, retrieve all tags contained in the initial annotation results; each tag is associated with an attribute identifier indicating its corresponding dimension. For example, in a skill / ability annotation scenario, tags include different dimensions such as programming languages, frameworks / tools, and domain knowledge, and each tag has a corresponding dimensional attribute identifier. By reading these dimensional attribute identifiers, construct a set of covered dimensions, such as {technical skills, industry knowledge}.

[0095] Next, the covered dimension set is compared with a preset complete dimension set to identify the missing dimensions that are not covered, and the dimension type of the missing dimensions is recorded. The preset complete dimension set is predefined according to the business domain; for example, the talent competency assessment domain may include dimensions such as {technical skills, industry knowledge, soft skills, management ability, and qualification certification}. The missing dimension set is obtained through set difference operations, for example, {soft skills, management ability, and qualification certification}.

[0096] For each dimension type of the missing dimension, a set of feature keywords corresponding to the dimension type is obtained from a pre-built dimension keyword library. The dimension keyword library is pre-constructed and stores the corresponding feature keywords for different dimension types. For example, the soft skills dimension is associated with the keyword set {communication, teamwork, leadership, problem-solving, time management}, while the management ability dimension is associated with the keyword set {project management, team management, resource allocation, risk control}.

[0097] For each missing dimension, the corresponding set of feature keywords is traversed, and keyword matching is performed in the ability description document. When a sentence or paragraph containing the feature keywords is found, the text fragment is extracted. For example, for the soft skills dimension, the text fragment "good at teamwork, with good communication and problem-solving skills" is extracted. To obtain more complete semantic information, a context expansion strategy can be adopted, extracting not only the sentences containing the keywords but also the sentences before and after them, forming more complete semantic units.

[0098] Semantic encoding is performed on the descriptive information corresponding to the missing dimensions to obtain semantic feature vectors for the missing dimensions. This involves using pre-trained language models (such as BERT, GPT, etc.) to encode the extracted text fragments, converting the text into high-dimensional semantic feature vectors. For example, the [CLS] tag output of the BERT model can be used as the semantic representation of the entire text fragment, resulting in a fixed-dimensional (e.g., 768-dimensional) feature vector that captures the semantic information of the text.

[0099] In the process of identifying label attribute features corresponding to missing dimensions through semantic parsing based on the semantic feature vectors of missing dimensions, and then selecting candidate labels belonging to the missing dimension type from the label library based on the label attribute features, a label library for specific dimension types is first constructed, with each label having a pre-calculated semantic feature vector. Then, the semantic similarity (e.g., cosine similarity) between the semantic feature vector of the missing dimension and the label vectors in the label library is calculated. Based on the similarity ranking, the labels with the highest similarity are selected as candidate labels. To enhance accuracy, a similarity threshold can be set; only labels with similarity exceeding the threshold are included in the candidate set.

[0100] Finally, candidate labels are added to the preliminary annotation results to obtain a complete annotation result. During the addition process, a confidence score (e.g., calculated based on semantic similarity) needs to be attached to each new label to indicate the reliability of the label matching. For any redundancy in labels, duplicate detection is performed to ensure that only one label with different expressions but similar semantics is retained. The source of the labels can also be recorded to distinguish between the initial annotations and the supplementary annotations, facilitating subsequent verification and optimization.

[0101] By employing the aforementioned dimensional verification and supplementation process, we can effectively identify and address missing dimensions in the annotation results, ensuring more comprehensive dimensional coverage and improving annotation quality and subsequent application effectiveness. This method is particularly suitable for text annotation scenarios requiring multi-dimensional analysis, such as talent capability assessment, product function classification, and customer needs analysis.

[0102] In one optional implementation, user input is received, and intent parsing is performed on the user input to extract core requirement information, related scenario information, and data dependency information to generate an intent vector, including:

[0103] Receive user input, perform text normalization processing on the user input, remove invalid characters and perform sentence segmentation to obtain standardized input text;

[0104] The standardized input text is parsed to identify the subject-verb-object components and modification relationships, thereby obtaining a grammatical structure tree.

[0105] Based on the grammar structure tree, the core phrases expressing the action target in the standardized input text are located, and the core phrases are extracted as core requirement information;

[0106] Based on the grammar structure tree, the scene description phrases that describe the application background and the dependency description phrases that describe the input and output constraints in the standardized input text are identified and extracted as related scene information and data dependency information, respectively.

[0107] The core requirement information and the associated scenario information are vector-concatenated according to a preset fusion strategy to generate an initial intent representation.

[0108] The data dependency information is added as a constraint to the initial intent representation to obtain the intent vector.

[0109] Figure 2 This is a schematic diagram illustrating the process of determining the intent vector according to an embodiment of the present invention. Figure 2 As shown, in one implementation, user input is first received. This input is a piece of natural language text, such as "Please check the weather in Beijing today and remind me if I need to bring an umbrella." For this type of input, intent parsing is required to accurately understand the user's needs.

[0110] The user input is processed for text normalization. First, invalid characters such as special characters, redundant spaces, and repeated punctuation marks are identified and removed. For example, "Help me check the weather in Beijing today!!! and remind me if I need to bring an umbrella..." is processed into "Help me check the weather in Beijing today! and remind me if I need to bring an umbrella." Next, the sentence is segmented based on punctuation marks and conjunctions. For example, the above input is segmented into two clauses: "Help me check the weather in Beijing today" and "Remind me if I need to bring an umbrella," forming standardized input text.

[0111] The standardized input text is parsed using grammatical structure analysis. Dependency parsing is employed to identify subject-verb-object components and modification relationships within the text. Taking the first clause as an example, grammatical analysis reveals that "help me" is a verb-object structure, "query" is the core verb, and "today's weather in Beijing" is the object. "Today" and "Beijing's" modify the time and location of "weather conditions," respectively. By constructing a grammatical structure tree, the dependency relationships between each phrase are clearly presented.

[0112] Based on the grammatical structure tree, we locate the core phrases that express the action goal, that is, identify the operation that the user really wants to perform. In the example above, the core verb is "query" and the core object is "weather conditions". Therefore, "query the weather" is extracted as the core requirement information. For the second clause, the core verb is "remind" and the core content is "need to bring an umbrella". This is extracted as another core requirement, "remind to bring an umbrella".

[0113] Next, scene description phrases that describe the application background are identified, such as the time word "today" and the location word "Beijing," and extracted as related scene information. At the same time, dependency description phrases that describe input and output constraints are identified, such as "whether" indicating that a boolean type output is required, and extracted as data dependency information.

[0114] The core requirement information and related scenario information are concatenated into vectors according to a preset fusion strategy to generate an initial intent representation. For the requirement of "checking the weather," it is fused with the scenario information "today" and "Beijing" to form a vector representation [checking the weather, today, Beijing]. The fusion strategy can be a simple vector concatenation or a weighted fusion based on an attention mechanism, assigning different weights to the requirements according to the importance of different scenario information.

[0115] Finally, data dependency information is added as a constraint to the initial intent representation to obtain the complete intent vector. In the example, "Do I need to bring an umbrella?" is added as a data dependency condition, and the final intent vector [Query weather, Today, Beijing, Output: Do I need to bring an umbrella?] is generated.

[0116] In practical applications, pre-trained language models such as BERT or RoBERTa can be used for text encoding to improve the accuracy of semantic understanding. For example, BERT can be used to encode standardized input text to obtain the contextual representation of each word, and then relevant information can be aggregated through an attention mechanism to identify core needs, scene information, and data dependencies.

[0117] For complex multi-turn dialogue scenarios, contextual information also needs to be considered. For example, if a user asks "What about tomorrow?" after checking the weather, the information "check the weather" and "Beijing" from the previous dialogue needs to be added to the current intent parsing as context to correctly identify it as "check the weather in Beijing tomorrow".

[0118] For ambiguous statements, a multi-candidate intent generation strategy can be adopted. For example, "I want to check my data usage" can mean "check mobile data usage" or "check website data usage". Multiple candidate intent vectors can be generated and sorted based on context or user history habits, and the most likely intent can be selected to respond.

[0119] In addition, domain knowledge bases can be established to enhance the understanding of professional terminology and specific scenarios. For example, in the financial field, professional terms such as "bull market" and "bear market" can be correctly interpreted in their specific meanings within the financial context through a domain knowledge base.

[0120] After the intent vector is generated, a similarity matching algorithm can be applied to match it with a predefined intent template to achieve more accurate service distribution. For example, the parsed intent vector can be compared with service templates such as "check the weather" and "set reminders" to select the most matching service for response.

[0121] Through the above methods, we can accurately extract the core intent from the user's natural language input, and combine it with contextual information and data dependencies to generate a structured intent vector, providing a foundation for subsequent intelligent services.

[0122] In one optional implementation, the intent vector is semantically matched with multi-dimensional capability tag vectors in a tag library, the similarity between the vectors is calculated, and a set of candidate agents whose similarity meets a preset threshold is selected, including:

[0123] The intent vector is decomposed to obtain demand dimension features and scenario dimension features;

[0124] For each agent's multi-dimensional capability label vector, extract the capability requirement dimension component corresponding to the requirement dimension feature, and the capability scene dimension component corresponding to the scene dimension feature from the multi-dimensional capability label vector;

[0125] Calculate the demand similarity between the demand dimension features and the capability demand dimension components, and the scenario similarity between the scenario dimension features and the capability scenario dimension components, respectively.

[0126] Based on preset dimensional weight coefficients, the demand similarity and the scene similarity are weighted and fused to obtain the comprehensive similarity between the intent vector and the multi-dimensional capability tag vector;

[0127] The overall similarity is compared with a preset threshold to select agents whose overall similarity meets the preset threshold, and the selected agents are added to the candidate agent set.

[0128] The agents in the candidate agent set are arranged in order according to the comprehensive similarity, resulting in a candidate agent set sorted by matching degree.

[0129] In this embodiment, the query intent input by the user is first received, the query intent is converted into an intent vector, and then the intent vector is semantically matched with the multi-dimensional capability tag vectors in the tag library. The similarity between the vectors is calculated, and a set of candidate agents whose similarity meets a preset threshold is selected.

[0130] First, the user's query intent is vectorized to generate an intent vector. The intent vector can be obtained by encoding a pre-trained language model such as BERT or GPT, and the dimension is usually 768 or 1024.

[0131] Next, the intent vector is decomposed to obtain demand dimension features and scenario dimension features. Demand dimension features mainly reflect the user's specific needs, such as translating documents or writing code; scenario dimension features reflect the application environment of the user's operation, such as business scenarios or academic research. The decomposition process can employ an attention mechanism. For a preset demand dimension vocabulary and scenario dimension vocabulary, the attention score between the intent vector and the keywords of each dimension is calculated, and the feature vectors corresponding to high scores are extracted. For example, when a user inputs "Help me translate an English business contract," the decomposition yields "translation" as the demand dimension feature and "business contract" as the scenario dimension feature.

[0132] For each agent's multi-dimensional capability label vector, components corresponding to the requirement dimension and scenario dimension features are extracted. The multi-dimensional capability label vector is pre-defined for each agent and includes its performance level in different capability dimensions. For example, if a translation agent's capability labels include "translation capability: 0.9", "business document processing: 0.8", and "academic paper processing: 0.5", then the capability requirement dimension component corresponding to the user requirement "translation" is extracted as 0.9, and the capability scenario dimension component corresponding to the "business contract" scenario is extracted as 0.8.

[0133] Calculate the demand similarity between demand dimension features and capability demand dimension components, and the scenario similarity between scenario dimension features and capability scenario dimension components. The similarity calculation can use the cosine similarity method, which is the dot product of two vectors divided by the product of their norms. For example, the closer the cosine value of the user demand "translation" vector and the agent's "translation capability" vector is to 1, the higher the matching degree.

[0134] Based on preset dimensional weight coefficients, the similarity between demand and scenario is weighted and fused to obtain the comprehensive similarity between the intent vector and the multi-dimensional capability tag vector. The dimensional weight coefficients can be adjusted according to the actual application scenario. For example, in some scenarios where demand matching is more important, the demand dimension weight can be set to 0.7 and the scenario dimension weight to 0.3. The comprehensive similarity calculation formula is: Comprehensive Similarity = Demand Weight × Demand Similarity + Scenario Weight × Scenario Similarity.

[0135] The calculated overall similarity is compared with a preset threshold. Agents whose overall similarity meets the preset threshold are selected and added to the candidate agent set. The preset threshold can be set according to actual application needs, such as 0.7 or 0.8, which means that only agents with similarity exceeding this value are considered qualified candidates.

[0136] The agents in the candidate agent set are sorted in order based on the comprehensive similarity, resulting in a candidate agent set ranked according to the degree of matching. The sorting is in descending order, with the agent with the highest similarity at the top, and is provided to the user for selection.

[0137] In practical applications, assuming a user inputs "Help me translate an English business contract and format it," this intent is first converted into a vector representation. Then, translation and formatting are decomposed as demand-dimensional features, and English and business contract as scenario-dimensional features. Multiple agents exist, such as professional translation agent A (translation ability: 0.95, formatting ability: 0.6, business document processing: 0.9), and general assistant agent B (translation ability: 0.7, formatting ability: 0.8, business document processing: 0.7), etc. After calculating the similarity of each dimension, if the weight of the demand dimension is set to 0.6 and the weight of the scenario dimension is set to 0.4, then the overall similarity of agent A is 0.6×[(0.95+0.6) / 2]+0.4×0.9=0.825, and the overall similarity of agent B is 0.6×[(0.7+0.8) / 2]+0.4×0.7=0.73. If the preset threshold is 0.7, both agents meet the conditions and are added to the candidate set. They are then sorted in descending order of comprehensive similarity, and finally, a list of candidate agents ranked A and B is recommended to the user.

[0138] Through the above methods, user intent and agent capabilities can be accurately matched, improving the accuracy of agent recommendations and user satisfaction. At the same time, the multi-dimensional matching strategy allows for a more comprehensive consideration of users' actual needs, providing more personalized agent services.

[0139] In one optional implementation, for each agent in the candidate agent set, its real-time performance metrics and historical call records are obtained; a performance score is calculated based on the real-time performance metrics; a user preference score is calculated based on the historical call records; and a comprehensive scheduling score is calculated based on the similarity, performance score, and user preference score, including:

[0140] For each agent in the candidate agent set, obtain the agent's real-time performance metrics;

[0141] Based on preset performance index weights, the response latency and task success rate in the real-time performance index are weighted and summed to obtain the performance score of the agent;

[0142] Retrieve the agent's historical call records from the historical call database, and count the number of calls and the number of successful calls related to the current user in the historical call records;

[0143] Based on the number of calls and the number of successful calls, the call frequency and success frequency of the agent in the current user's historical usage are calculated. Based on the call frequency and the success frequency, a comprehensive evaluation is performed according to a preset preference calculation rule to obtain the user preference score of the agent for the current user.

[0144] Based on preset scheduling weight coefficients, the similarity, performance score, and user preference score are weighted and fused to obtain the comprehensive scheduling score of the agent.

[0145] In one embodiment, for each agent in the candidate agent set, its real-time performance metrics and historical call records are first obtained. Real-time performance metrics mainly include response latency and task success rate, while historical call records contain historical data on how the agent was invoked by different users. Based on this information, a comprehensive scheduling score is calculated for each agent to select the most suitable agent for the user request.

[0146] When acquiring real-time performance metrics, the system connects to the agent monitoring module to collect the agent's operational status data in real time. Real-time performance metrics typically include: response latency (in milliseconds, representing the time required for the agent to process a request) and task success rate (in percentage form, representing the proportion of tasks successfully completed by the agent). For example, for agent A, the acquired real-time performance metrics are: response latency 200 milliseconds, and task success rate 95%.

[0147] For the acquired real-time performance metrics, preset performance metric weights are applied for calculation. Assuming the preset response latency weight is 0.4 and the task success rate weight is 0.6, the calculation formula is: Performance Score = Response Latency Score × 0.4 + Task Success Rate × 0.6. The response latency score needs to be normalized, which can be done as follows: Response Latency Score = 1 - (Current Latency / Maximum Acceptable Latency), ensuring that the shorter the latency, the higher the score. For the above agent A, if the maximum acceptable latency is 1000 milliseconds, its response latency score is 1 - (200 / 1000) = 0.8, and the final performance score is 0.8 × 0.4 + 0.95 × 0.6 = 0.89.

[0148] Retrieve historical call records of the current agent from the historical call database, focusing on interaction data related to the current user. Historical call records typically include fields such as call timestamp, calling user ID, and call result (success / failure). By filtering records with the current user ID, calculate the total number of times the agent was called by the current user and the number of successful calls. For example, agent A was called 50 times by the current user, with 45 of those successful completions.

[0149] Based on the above statistical data, the call frequency and success frequency are calculated. Call frequency represents how often the agent is called by the current user, and can be calculated by dividing the number of calls by the total number of calls within a time window (e.g., 30 days). Success frequency is the ratio of successful calls to the total number of calls. For agent A, its success frequency is 45 / 50 = 0.9. According to the preset preference calculation rules, the user preference score can be expressed as: User Preference Score = Call Frequency Weight × Call Frequency + Success Frequency Weight × Success Frequency. Assuming the call frequency weight is 0.3 and the success frequency weight is 0.7, if agent A's call frequency is 0.25 (i.e., accounting for 25% of all agents), then its user preference score is 0.25 × 0.3 + 0.9 × 0.7 = 0.705.

[0150] After obtaining the similarity score, performance score, and user preference score, a weighted fusion is performed according to preset scheduling weight coefficients to calculate the comprehensive scheduling score. Assuming the similarity weight is 0.4, the performance score weight is 0.3, and the user preference score weight is 0.3, the formula for calculating the comprehensive scheduling score of the agent is: Comprehensive Scheduling Score = Similarity × 0.4 + Performance Score × 0.3 + User Preference Score × 0.3. If the similarity of agent A is 0.85, then its comprehensive scheduling score is 0.85 × 0.4 + 0.89 × 0.3 + 0.705 × 0.3 = 0.82.

[0151] In practical applications, the weighting coefficients of each item can be adjusted according to different scenarios. For example, for application scenarios with high real-time requirements, the weight of response latency in the performance score can be increased; for application scenarios that emphasize personalized services, the weight of user preference score in the overall scheduling score can be increased.

[0152] By using this weighted fusion approach, the matching degree between the agent and the user request, the running performance, and the user's historical preferences are comprehensively considered. This ensures that the agent most suitable for handling the current user request is selected. From the set of candidate agents, the agent with the highest comprehensive scheduling score is selected to handle the user request, which can effectively improve the system response speed, task success rate, and user satisfaction.

[0153] When the overall scheduling scores of multiple agents are very close, a random factor or load balancing can be introduced to prevent some high-scoring agents from being over-utilized. This dynamic adjustment mechanism ensures stability and reliability while providing users with the best service experience.

[0154] In one optional implementation, the call frequency and success frequency of the agent in the current user's historical usage are calculated based on the number of calls and the number of successful calls. Based on the call frequency and the success frequency, a comprehensive evaluation is performed according to a preset preference calculation rule to obtain the user preference score of the agent for the current user, including:

[0155] Retrieve the total number of calls made by the current user within a preset time window, as recorded in the historical call database, and use this as the user's call baseline.

[0156] The call frequency of the intelligent agent is obtained by calculating the ratio of the number of times the agent is called to the base number of user calls.

[0157] The success frequency of the agent is obtained by calculating the ratio of the number of successful calls to the total number of calls made by the agent.

[0158] The call frequency is subjected to time decay processing. Based on the time interval between each call in the historical call record and the current time, the contribution value of the corresponding call record in the call frequency is decayed and weighted to obtain the time-weighted call frequency.

[0159] The success frequency is processed by task type differentiation. Based on the task type processed in each successful call in the historical call record, the task type corresponding to the core requirement input by the current user is identified, and the successful call records matching the task type in the historical call record are extracted. The task-related success frequency is recalculated based on the successful call records.

[0160] According to the preset preference calculation rules, the time-weighted call frequency and the task-related success frequency are weighted and fused to obtain the user preference score of the agent for the current user.

[0161] In implementing this invention, the user's historical call data is first acquired, including the number of calls and successful calls to different agents. This data is stored in a historical call database, recording the user's usage of each agent at different points in time.

[0162] Based on the number of calls and the number of successful calls obtained, the call frequency and success frequency of the agent in the current user's historical usage are calculated. This calculation process includes multiple steps to ensure the accuracy and practicality of the evaluation.

[0163] First, the total number of calls made by the current user within a preset time window is retrieved from the historical call database. This number serves as the user's call base. The preset time window can be the most recent week, month, or three months, and can be flexibly set according to the actual application scenario. For example, if a user has made a total of 500 calls to the agent service in the most recent month, then the user's call base is 500.

[0164] Next, the ratio of the number of calls to a specific agent to the base number of user calls is calculated to obtain the call frequency of that agent. If a user calls agent A a total of 100 times within a preset time window, then the call frequency of agent A is 100 / 500=0.2, which means that 20% of the user's agent call requests are for agent A.

[0165] Then, the ratio of the number of successful calls to the total number of calls is calculated to obtain the success frequency of the agent. If agent A successfully completes the task 80 times out of 100 calls, its success frequency is 80 / 100=0.8, which means that the user has an 80% success rate when calling agent A.

[0166] To more accurately reflect users' latest preferences, time-weighted call frequency is applied. Based on the time interval between each call in the historical call record and the current time, the contribution value of the corresponding call record in the call frequency is weighted and attenuated to obtain the time-weighted call frequency.

[0167] Time decay processing can employ an exponential decay function: for a call record with a time interval of t, its weight can be set as e^(-λt), where λ is the decay coefficient. Newer call records receive higher weights, and older call records receive lower weights. For example, if a call record from one week ago has a weight of 0.7, two weeks ago has 0.5, and one month ago has 0.3, then the corresponding weight for each record can be calculated based on its time, and then the weighted sum can be obtained to obtain the time-weighted call frequency.

[0168] To improve the relevance of the evaluation, the frequency of success is differentiated by task type. Based on the task type processed during each successful call in the historical call log, the task type corresponding to the core needs of the current user input is identified. Task types can include information retrieval, content creation, Q&A consultation, data analysis, etc.

[0169] Natural language processing (NLP) techniques are used to analyze the user's current input, extract core needs, and match them to corresponding task types. Then, successful call records matching the task type are filtered from historical call logs, and the task-related success frequency is recalculated based on these records.

[0170] For example, if the user's current task is to write an article, the system identifies the current task type as content creation and extracts all successful calls related to content creation from the historical call records. Assuming agent A has 30 successful calls in content creation tasks out of a total of 35 attempts, its success rate for this task type is approximately 30 / 35 ≈ 0.86.

[0171] Finally, based on preset preference calculation rules, the time-weighted call frequency and task-related success frequency are weighted and fused to obtain the agent's user preference score for the current user. The preference calculation rules can be:

[0172] User preference score = α × time-weighted call frequency + β × task-related success frequency

[0173] Here, α and β are weighting coefficients that can be adjusted according to the actual application scenario, and their sum is 1. For example, setting α=0.4 and β=0.6 will place more emphasis on the contribution of task-related success frequency to the preference score.

[0174] If agent A's time-weighted call frequency is 0.25 and its task-related success frequency is 0.86, then its user preference score is 0.4 × 0.25 + 0.6 × 0.86 = 0.616. This score reflects the degree of matching and preference of agent A for the current user on a specific task type.

[0175] In practical applications, thresholds can be set to normalize user preference scores, making the scores of different agents more comparable. For example, the preference score range of all candidate agents can be mapped to 0-100, making it easier for users to understand and make agent recommendations.

[0176] Through the above calculation process, the user's historical usage habits, the agent's task completion ability, and time factors can be comprehensively considered to recommend the most suitable agent to the user, thereby improving user experience and task completion efficiency.

[0177] Furthermore, to improve computational efficiency, incremental computation can be used, updating only the frequency data of the relevant agents when a new call record is generated, rather than recalculating all data. Old historical records can also be periodically cleaned up to maintain the database's efficient operation.

[0178] The cross-domain intelligent agent precise scheduling system of this invention includes:

[0179] The first unit is used to extract multi-dimensional tags from the capability description document of the connected intelligent agent by parsing the intelligent agent's capability description document, and then encode and store the tags in a tag library to obtain a standardized multi-dimensional capability tag vector;

[0180] The second unit is used to receive user input, perform intent parsing on the user input, extract core requirement information, related scenario information, and data dependency information to generate an intent vector;

[0181] The third unit is used to perform semantic matching between the intent vector and the multi-dimensional capability tag vectors in the tag library, calculate the similarity between the vectors, and filter out a set of candidate agents whose similarity meets a preset threshold;

[0182] The fourth unit is used to obtain the real-time performance indicators and historical call records for each agent in the candidate agent set, calculate a performance score based on the real-time performance indicators, calculate a user preference score based on the historical call records, and calculate a comprehensive scheduling score based on the similarity, performance score, and user preference score.

[0183] The fifth unit is used to select a target agent from the candidate agent set based on the comprehensive scheduling score, send a scheduling instruction to the target agent, and determine the execution result.

[0184] A third aspect of the present invention provides an electronic device, comprising:

[0185] processor;

[0186] Memory used to store processor-executable instructions;

[0187] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0188] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0189] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for precise scheduling of cross-domain intelligent agents, characterized in that, include: For the connected intelligent agent, by parsing the intelligent agent's capability description document, multiple dimensions of tags are extracted, and the tags are encoded and stored in a tag library to obtain a standardized multi-dimensional capability tag vector; The system receives user input, performs intent parsing on the input, and extracts core requirement information, related scenario information, and data dependency information to generate an intent vector. The intent vector is semantically matched with the multi-dimensional capability tag vectors in the tag library, the similarity between the vectors is calculated, and a set of candidate agents whose similarity meets a preset threshold is selected. For each agent in the candidate agent set, its real-time performance metrics and historical call records are obtained. A performance score is calculated based on the real-time performance metrics, a user preference score is calculated based on the historical call records, and a comprehensive scheduling score is calculated based on the similarity, performance score, and user preference score. Based on the comprehensive scheduling score, a target agent is selected from the set of candidate agents, a scheduling instruction is sent to the target agent, and the execution result is determined.

2. The method according to claim 1, characterized in that, For the connected agent, by parsing the agent's capability description document, multiple dimensions of tags are extracted, and these tags are encoded and stored in a tag library to obtain a standardized multi-dimensional capability tag vector, including: Obtain the capability description document of the intelligent agent, perform text preprocessing on the capability description document, remove redundant characters and segment sentences to obtain a set of standardized text fragments; Semantic encoding is performed on each text segment in the standardized text segment set, and the text segment is converted into a semantic feature vector; Based on the semantic feature vector and the pre-stored tag vectors in the tag library, the corresponding tags are automatically labeled to the agent to obtain preliminary labeling results; The preliminary annotation results are validated for dimensions to determine whether they cover multiple dimensions simultaneously. If any dimensions are missing, the description information corresponding to the missing dimensions is extracted from the capability description document, and the corresponding dimension labels are supplemented based on the description information. After dimensional verification, the labels are encoded according to preset encoding rules to generate unique identifiers. The unique identifiers are then associated with the semantic feature vectors of the corresponding labels and stored in the label library to obtain the multi-dimensional capability label vectors of the agent.

3. The method according to claim 2, characterized in that, The preliminary annotation results are validated for dimensions to determine whether they cover multiple dimensions simultaneously. If any dimensions are missing, the description information corresponding to the missing dimensions is extracted from the capability description document, and the corresponding dimension labels are supplemented based on the description information, including: Iterate through all labels in the preliminary annotation results, and based on the dimension attribute identifier of each label, count the dimension types covered by the preliminary annotation results to obtain the set of covered dimensions; The covered dimension set is compared with the preset complete dimension set to identify the missing dimensions that have not been covered, and the dimension type of the missing dimensions is recorded. For each dimension type of the missing dimension, a set of feature keywords corresponding to the dimension type is obtained from a preset dimension keyword library; Retrieve text fragments containing keywords from the feature keyword set from the capability description document, extract the text fragments as descriptive information corresponding to the missing dimension, and perform semantic encoding on the descriptive information corresponding to the missing dimension to obtain the semantic feature vector of the missing dimension; Based on the semantic feature vector of the missing dimension, the label attribute features corresponding to the missing dimension are identified through semantic parsing, and candidate labels belonging to the missing dimension type are selected from the label library according to the label attribute features. The candidate labels are added to the preliminary annotation results to obtain the dimension-complete annotation results.

4. The method according to claim 1, characterized in that, Receive user input, perform intent parsing on the user input, extract core requirement information, related scenario information, and data dependency information to generate an intent vector, including: Receive user input, perform text normalization processing on the user input, remove invalid characters and perform sentence segmentation to obtain standardized input text; The standardized input text is parsed to identify the subject-verb-object components and modification relationships, thereby obtaining a grammatical structure tree. Based on the grammar structure tree, the core phrases expressing the action target in the standardized input text are located, and the core phrases are extracted as core requirement information; Based on the grammar structure tree, the scene description phrases that describe the application background and the dependency description phrases that describe the input and output constraints in the standardized input text are identified and extracted as related scene information and data dependency information, respectively. The core requirement information and the associated scenario information are vector-concatenated according to a preset fusion strategy to generate an initial intent representation. The data dependency information is added as a constraint to the initial intent representation to obtain the intent vector.

5. The method according to claim 1, characterized in that, The intent vector is semantically matched with multi-dimensional capability tag vectors in the tag library, the similarity between the vectors is calculated, and a set of candidate agents whose similarity meets a preset threshold is selected, including: The intent vector is decomposed to obtain demand dimension features and scenario dimension features; For each agent's multi-dimensional capability label vector, extract the capability requirement dimension component corresponding to the requirement dimension feature, and the capability scene dimension component corresponding to the scene dimension feature from the multi-dimensional capability label vector; Calculate the demand similarity between the demand dimension features and the capability demand dimension components, and the scenario similarity between the scenario dimension features and the capability scenario dimension components, respectively. Based on preset dimensional weight coefficients, the demand similarity and the scene similarity are weighted and fused to obtain the comprehensive similarity between the intent vector and the multi-dimensional capability tag vector; The overall similarity is compared with a preset threshold to select agents whose overall similarity meets the preset threshold, and the selected agents are added to the candidate agent set. The agents in the candidate agent set are arranged in order according to the comprehensive similarity, resulting in a candidate agent set sorted by matching degree.

6. The method according to claim 1, characterized in that, For each agent in the candidate agent set, its real-time performance metrics and historical call records are obtained. A performance score is calculated based on the real-time performance metrics, a user preference score is calculated based on the historical call records, and a comprehensive scheduling score is calculated based on the similarity, performance score, and user preference score, including: For each agent in the candidate agent set, obtain the agent's real-time performance metrics; Based on preset performance index weights, the response latency and task success rate in the real-time performance index are weighted and summed to obtain the performance score of the agent; Retrieve the agent's historical call records from the historical call database, and count the number of calls and the number of successful calls related to the current user in the historical call records; Based on the number of calls and the number of successful calls, the call frequency and success frequency of the agent in the current user's historical usage are calculated. Based on the call frequency and the success frequency, a comprehensive evaluation is performed according to a preset preference calculation rule to obtain the user preference score of the agent for the current user. Based on preset scheduling weight coefficients, the similarity, performance score, and user preference score are weighted and fused to obtain the comprehensive scheduling score of the agent.

7. The method according to claim 6, characterized in that, Based on the number of calls and the number of successful calls, the call frequency and success frequency of the agent in the current user's historical usage are calculated. Based on the call frequency and success frequency, a comprehensive evaluation is performed according to preset preference calculation rules to obtain the user preference score of the agent for the current user, including: Retrieve the total number of calls made by the current user within a preset time window, as recorded in the historical call database, and use this as the user's call baseline. The call frequency of the intelligent agent is obtained by calculating the ratio of the number of times the agent is called to the base number of user calls. The success frequency of the agent is obtained by calculating the ratio of the number of successful calls to the total number of calls made by the agent. The call frequency is subjected to time decay processing. Based on the time interval between each call in the historical call record and the current time, the contribution value of the corresponding call record in the call frequency is decayed and weighted to obtain the time-weighted call frequency. The success frequency is processed by task type differentiation. Based on the task type processed in each successful call in the historical call record, the task type corresponding to the core requirement input by the current user is identified, and the successful call records matching the task type in the historical call record are extracted. The task-related success frequency is recalculated based on the successful call records. According to the preset preference calculation rules, the time-weighted call frequency and the task-related success frequency are weighted and fused to obtain the user preference score of the agent for the current user.

8. A cross-domain intelligent agent precise scheduling system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The first unit is used to extract multi-dimensional tags from the capability description document of the connected intelligent agent by parsing the intelligent agent's capability description document, and then encode and store the tags in a tag library to obtain a standardized multi-dimensional capability tag vector; The second unit is used to receive user input, perform intent parsing on the user input, extract core requirement information, related scenario information, and data dependency information to generate an intent vector; The third unit is used to perform semantic matching between the intent vector and the multi-dimensional capability tag vectors in the tag library, calculate the similarity between the vectors, and filter out a set of candidate agents whose similarity meets a preset threshold; The fourth unit is used to obtain the real-time performance indicators and historical call records for each agent in the candidate agent set, calculate a performance score based on the real-time performance indicators, calculate a user preference score based on the historical call records, and calculate a comprehensive scheduling score based on the similarity, performance score, and user preference score. The fifth unit is used to select a target agent from the candidate agent set based on the comprehensive scheduling score, send a scheduling instruction to the target agent, and determine the execution result.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images