ChatOps-based methods for real-time IT operations collaboration and automated work order generation

Abstract

This invention belongs to the field of automated operation and maintenance technology, and particularly relates to a method for real-time IT operation and maintenance collaboration and automatic work order generation based on ChatOps. This method monitors and parses operation and maintenance dialogues to generate structured data representing intent and entities; then intelligently allocates tasks, forming work order data to be filled and executable automated instruction sequences; under rule constraints, the work order data is verified and completed, and orchestrated control instructions are generated; finally, the work order system is synchronously driven to create structured work orders and the automation engine to execute operation and maintenance operations, realizing a closed loop from unstructured dialogues to traceable and executable operation and maintenance processes, improving collaboration efficiency and handling accuracy.

Description

Technical Field

[0001] This invention belongs to the field of automated operation and maintenance technology, and particularly relates to a method for real-time collaboration and automatic work order generation in IT operation and maintenance based on ChatOps. Background Technology

[0002] In ChatOps-based IT operations collaboration, bots are typically used as the central interaction mechanism. Based on commands entered by operations personnel in the chat platform or discussions of monitoring alarms, bots invoke backend automated scripts or API interfaces to execute corresponding operations, and the execution results are fed back to the chat channel in real time, thereby achieving transparent and traceable collaborative work. For tasks that require tracking, such as fault handling and service requests, existing technologies usually require manual creation and filling of work orders in an external IT service management system (ITSM) based on chat logs to drive subsequent process-oriented processing and knowledge accumulation. For real-time dialogues generated in complex fault scenarios, involving multiple participants, multiple rounds, and a large amount of unstructured text, the key information needed to generate standardized work orders (such as accurate fault service identification, resource location, impact description, etc.) is often scattered and inconsistently expressed. In existing technologies, the automatic or semi-automatic generation of work orders usually relies on simple keyword matching or manual triggering of fixed robot command templates in the dialogue to extract limited structured fields. This method is applicable in scenarios with sparse information and standardized formats, but its information extraction capability is significantly limited when dealing with the aforementioned complex natural language dialogues. Chatbots do not explicitly distinguish between valid factual statements, irrelevant discussions, historical context restates, and vague references in the dialogue, and also lack the ability to deeply understand the semantics and information fusion of multi-turn dialogue contexts. As a result, when faced with real collaborative dialogues with high information density and diverse expressions, simple rule-based extraction methods are prone to information omissions, identification errors, or inaccurate associations.

[0003] In summary, in real-time IT operations collaboration based on ChatOps, existing information extraction methods based on keyword matching or fixed templates are particularly ineffective in multi-turn dialogue scenarios when dealing with highly unstructured and context-dependent instant conversations. This is especially true in cases where the methods fail to effectively resolve referential inconsistencies (such as the specific meaning of "it" or "that service") and integrate fragmented information across messages. Consequently, when extracting key entities (such as fault objects and scope of impact) from the dialogue history for generating work orders, errors in identification, information fragmentation, or omissions occur, resulting in inaccurate and incomplete automated work order information. Therefore, accurately resolving referential inconsistencies and integrating key information across messages in multi-turn dialogues to generate high-fidelity structured work order data is a specific technical problem that needs to be solved. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a ChatOps-based method for real-time IT operations collaboration and automatic work order generation. This method first continuously monitors and collects real-time dialogue content from a chat platform. Then, it uses a configured semantic parsing model to deeply understand and represent unstructured dialogue, generating a first structured data set representing core intent and key entity information. Based on this set, intelligent task allocation is performed between work order information fields and automated operations, resulting in a second structured data set for filling work order templates and a set of automated operation instructions. Under preset rule constraints, the second structured data set is validated and completed. Combined with the first structured data set and the automated operation instruction set, a third control instruction is generated for orchestrating execution logic. The validated and completed work order data is output to the work order system to automatically create work orders, while the third control instruction is output to the automation engine to drive operations, thus achieving a closed loop from real-time dialogue to structured recording and automated processing.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] ChatOps-based methods for real-time IT operations collaboration and automated work order generation include:

[0007] Based on continuous monitoring of real-time conversations on chat platforms, real-time conversation content containing unstructured text is collected.

[0008] The real-time dialogue content is understood and represented by the configured semantic parsing model, and a first structured data set representing the core processing intention and key entity information of the dialogue is calculated and generated.

[0009] Based on the first structured data set, tasks are allocated between the preset work order information fields and the automated operation and maintenance operations to be triggered, resulting in a second structured data set for filling into the standardized work order template, and a set of automated operation instructions for driving the automated operation and maintenance engine to execute.

[0010] Under the constraints of preset work order integrity rules and information accuracy verification rules, the second structured data set is constrained, verified, and completed to obtain the verified and completed second structured data set. Based on the first structured data set and the automated operation instruction set, a third control instruction is generated. The third control instruction is used to characterize the arrangement of the execution logic and order of the automated operation instruction set.

[0011] The second structured dataset after verification and completion is output as a work order generation instruction to the creation interface of the work order management system, and the third control instruction is output as an automated orchestration instruction to the control loop of the automated operation and maintenance engine.

[0012] Specifically, the matching degree between the key entity information in the second structured data set after verification and completion and the corresponding key entity information in the first structured data set is not less than a first preset threshold; the matching degree between the set of automated operation instructions arranged by the third control instruction and the core dialogue processing intention represented in the first structured data set is not less than a second preset threshold.

[0013] Specifically, a first structured data set representing the core intent of the dialogue and key entity information is calculated and generated, including:

[0014] The real-time dialogue content is normalized to obtain a normalized dialogue text sequence;

[0015] Based on the normalized dialogue text sequence, each word is mapped to a dense vector through the word vector model in the embedding layer to obtain the word vector sequence;

[0016] Based on the word vector sequence, forward and backward sequence modeling is performed through a bidirectional long short-term memory network in the encoding layer to obtain a semantic vector representation of the dialogue that incorporates contextual information;

[0017] Based on the semantic vector representation of the dialogue, a nonlinear transformation is performed through the fully connected layer in the first decoding branch, and then the probability distribution is calculated through the Softmax function to obtain the intent classification label and corresponding confidence level that characterize the core processing intent of the dialogue.

[0018] Specifically, the calculation and generation of the first structured data set representing the core intent of the dialogue and key entity information also includes:

[0019] Based on the semantic vector representation of the dialogue, sequence labeling is performed through the conditional random field model in the second decoding branch. Each word in the normalized dialogue text sequence is classified to obtain a preliminary key entity information set. Each element in the preliminary key entity information set includes an entity type label and the start and end position index of the entity in the normalized dialogue text sequence.

[0020] The normalized dialogue text sequence is context-encoded to obtain the semantic vector representation of the current round of dialogue and the semantic vector representation of each historical round of dialogue. Based on a preset context window, the semantic vector representation of the current round of dialogue and the semantic vector representation of the historical rounds of dialogue are concatenated. The dependency weight of the current semantic vector on the historical semantic vector is calculated through an attention mechanism. The historical semantic vectors are weighted and summed according to the dependency weight. The weighted sum is then fused with the current semantic vector to obtain an enhanced semantic representation that incorporates the historical dialogue context.

[0021] Specifically, the calculation and generation of the first structured data set representing the core intent of the dialogue and key entity information also includes:

[0022] Based on the enhanced semantic representation, the vector cosine similarity between different entity references in the preliminary key entity information set is calculated through the coreference resolution layer. Multiple entity references with cosine similarity exceeding a preset threshold are merged into the same entity identifier to obtain a unique key entity information set that eliminates referential ambiguity and has a unified identifier.

[0023] Based on the intent classification label and the unique key entity information set, the data is assembled and serialized according to the preset key value through the information fusion layer to obtain the first structured data set representing the core processing intent and key entity information of the dialogue.

[0024] Specifically, tasks are assigned between preset work order information fields and automated operation and maintenance operations to be triggered, including:

[0025] Based on the intent classification labels and corresponding confidence levels in the first structured dataset, a decision tree classifier with confidence threshold verification is used to match predefined work order template selection rules, filter paths whose confidence levels do not meet the threshold, and obtain the target work order template identifier.

[0026] Based on the target work order template identifier, the template-field association database is queried by combining a database query function that sorts fields by priority, and the set of required field names, the set of optional field names, and the field priority sorting results are obtained.

[0027] Based on the unique key entity information set and entity identifier in the first structured data set, the mapping is completed and the entity identifier is associated by using the entity type-field mapping function that supports dynamic updates, and the mapping is completed by using the iteratively updated entity type-work order field correspondence lookup table to obtain the preliminary field filling mapping relationship.

[0028] Specifically, the task allocation between the preset work order information fields and the automated operation and maintenance operations to be triggered also includes:

[0029] Based on the preliminary field filling mapping relationship, the unique key entity information set, and the field priority sorting results, the entity text content is extracted according to priority by the field value filling function with uniqueness verification, and the field filling and deduplication are completed to generate the second structured data set.

[0030] Based on the intent classification labels, unique key entity information set and entity attribute information in the first structured data set, the rule engine loads the hierarchical condition action rule set through the automated operation triggering of multi-level condition matching, and generates an automated operation instruction set with execution priority identifier.

[0031] Based on the execution priority, logical dependencies, and execution prerequisites of the automated operation instruction set, a directed acyclic graph with dependency conflict detection is used to construct and optimize the task execution flowchart to obtain the initial automated orchestration logic.

[0032] Based on the initial automated orchestration logic, the current system load, resource utilization rate, and historical execution time data, the execution order of instructions is adjusted and system resources are allocated through a multi-objective optimized resource scheduling function, generating a third control instruction that represents the execution logic, execution order, and resource allocation scheme of the automated operation instruction set.

[0033] Specifically, under the constraints of preset work order integrity rules and information accuracy verification rules, the second structured data set is subjected to constraint verification and completion, including:

[0034] Based on the target work order template identifier, a pre-defined work order integrity rule database is queried by executing a structured query language statement to obtain a list of required field definitions and a list of field format rules associated with the target work order template identifier.

[0035] Based on the list of required fields defined and the set of field names contained in the second structured data set, the missing required field names are obtained by calculating the set difference, thus forming a list of missing fields;

[0036] Based on the field format rule list, a regular expression engine is used to perform pattern matching verification on the padding values ​​corresponding to each field in the second structured data set, identifying fields whose formats do not conform to the field format rules, and forming a list of fields with abnormal formats.

[0037] Specifically, under the constraints of preset work order integrity rules and information accuracy verification rules, the second structured data set is subjected to constraint verification and completion, which also includes:

[0038] Based on the entity identifiers contained in the unique key entity information set, authoritative configuration attribute information corresponding to each entity identifier is obtained by calling the Hypertext Transfer Protocol application programming interface provided by the configuration management database.

[0039] Based on the field fill values ​​associated with entities in the second structured data set and the corresponding authoritative configuration attribute information obtained from the configuration management database, the accuracy quantification score of each field fill value is obtained by calculating the Levenstein distance and converting it into a similarity score.

[0040] Based on a preset accuracy verification threshold, by comparing the accuracy quantification score with the accuracy verification threshold, fields with accuracy quantification scores lower than the accuracy verification threshold are identified, forming a list of fields with abnormal accuracy.

[0041] Specifically, under the constraints of preset work order integrity rules and information accuracy verification rules, the second structured data set is subjected to constraint verification and completion, which also includes:

[0042] Based on the missing field list, information retrieval is performed by executing a graph query language query statement or a structured query language query statement to access the operation and maintenance knowledge base or historical work order database. If the retrieval is successful, the corresponding fill value is generated and the second structured data set is updated.

[0043] Based on the list of format exception fields and the list of accuracy exception fields, a task to be reviewed containing exception details and context is created through the workflow engine, and the task to be reviewed is submitted to the manual review queue. At the same time, the automatic work order generation process based on the current second structured data set is paused.

[0044] After completing information completion and potential manual review and correction, a second structured data set that satisfies all preset rules is generated by performing a final integrity check and data merging.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] This invention addresses the shortcomings of existing technologies by integrating real-time chat on a chat platform with the entire IT operations and maintenance (O&M) process through a ChatOps model. The core benefits and derivations are as follows: Based on continuous monitoring and collection of unstructured dialogues, the core intent and unique key entities are accurately extracted using semantic parsing models, bidirectional long short-term memory networks, and coreference resolution techniques, solving the pain points of fragmented and ambiguous information in traditional O&M dialogues. Through confidence verification, dynamic field mapping, and priority sorting, task allocation methods achieve accurate matching of work order fields and automated operation command generation, avoiding the inefficiency and errors of manual allocation. Multi-dimensional verification rules, cross-validation of the configuration management database, and knowledge base completion ensure the completeness and accuracy of work order information, meeting O&M compliance requirements. Finally, by connecting the work order system with the automated O&M engine, a closed loop is built from dialogue understanding to work order generation and automated execution, significantly reducing manual intervention, improving O&M response speed and collaboration efficiency, and lowering O&M costs. Simultaneously, it ensures a high degree of alignment between automated operations and dialogue intent, effectively supporting the standardized and intelligent implementation of O&M scenarios such as fault handling and resource application. Attached Figure Description

[0047] Figure 1 This is a flowchart of the method for real-time IT operations collaboration and automatic work order generation based on ChatOps in Embodiment 1 of the present invention.

[0048] Figure 2 This is a logical structure diagram of the semantic parsing model in Embodiment 1 of the present invention.

[0049] Figure 3 This is a logic diagram of the rule matching engine performing multi-level condition matching in Embodiment 1 of the present invention. Detailed Implementation

[0050] Example 1

[0051] Please see Figure 1 The present invention provides an embodiment of a method for real-time IT operations collaboration and automatic work order generation based on ChatOps, the steps of which include:

[0052] S1. Based on continuous monitoring of real-time conversation content in the chat platform, collect real-time conversation content containing unstructured text.

[0053] S2. The instant dialogue content is understood and represented by the configured semantic parsing model, and a first structured data set representing the core processing intention and key entity information of the dialogue is calculated and generated.

[0054] S3. Based on the first structured data set, tasks are allocated between the preset work order information fields and the automated operation and maintenance operations to be triggered, resulting in a second structured data set for filling into the standardized work order template, and a set of automated operation instructions for driving the automated operation and maintenance engine to execute. For example, in this embodiment, the preset work order information fields are set based on the work order template structure and business requirements of the target operation and maintenance scenario. First, the core data items contained in various work orders in this scenario are sorted out, including general fields such as work order number, work order type, creation time, handler, priority, etc., as well as entity fields defined according to specific business domains, such as fault object, performance index, status value, handling action, supplier name, contract number, asset identifier, etc. These fields are standardized into fixed field names and data types to form a set of work order information fields that can be parsed by the automated system. For example, in a technology risk management scenario, the pre-defined work order information fields include a risk type field to store the dialogue intent classification results, a supplier field to store the supplier name identified by the entity, a contract field to store the associated contract number, and a risk item field to store the specific risk description. At the same time, each field is configured with corresponding automated operation parameter mapping rules. When a work order is generated, the system automatically matches and triggers the corresponding operation and maintenance operations based on the values ​​of these fields. For example, the value of the supplier field triggers the supplier information query interface, and the value of the contract field triggers the contract status verification script, ensuring that the work order information can be accurately assigned to the corresponding automated operation and maintenance operations.

[0055] S4. Under the constraints of preset work order integrity rules and information accuracy verification rules, the second structured data set is constrained, verified, and completed to obtain a verified and completed second structured data set. Based on the first structured data set and the automated operation instruction set, a third control instruction is generated. The third control instruction is used to characterize the arrangement of the execution logic and sequence of the automated operation instruction set. It should be further noted that the matching degree between the key entity information in the verified and completed second structured data set and the corresponding key entity information in the first structured data set is not less than a first preset threshold; the matching degree between the automated operation instruction set arranged by the third control instruction and the core dialogue processing intention represented in the first structured data set is not less than a second preset threshold. It should be further explained that, in this embodiment, when setting the work order integrity rules and information accuracy verification rules, rule modeling is first performed based on the work order template definition and historical work order data statistics of the target operation and maintenance scenario: the work order integrity rules are preset according to the essential field definitions, field format specifications, and dependencies between fields of each work order template. For example, in the technology risk management scenario, the preset integrity rules for the risk reporting work order template are as follows: the supplier name field cannot be empty and must conform to the predefined supplier list format; the risk level field must be an enumerated value of high / medium / low; and the occurrence time field must be a valid timestamp format and cannot be later than the current time. At the same time, association rules are preset, such as when the risk level is high, responsibility is... The "person" field cannot be empty. The information accuracy verification rules are preset based on the accuracy distribution of entity recognition in historical work orders and the sample data annotated by business experts. By statistically analyzing the matching degree distribution between entity information extracted from the dialogue and entity information confirmed in the final work order, the lower quartile of the matching degree distribution is used as the benchmark and combined with the business tolerance requirements to set a first preset threshold, such as 0.95. At the same time, by analyzing the consistency distribution between the core handling intention and the actual execution operation instructions in historical work orders, a second preset threshold is set based on the median of the consistency distribution or the business expectation value, such as 0.90. This ensures that the work order data after verification and completion not only meets the requirements of structured integrity but also truly reflects the core intention and key information of the original dialogue.For example, the first and second preset thresholds in this embodiment are determined based on the statistical analysis results of historical work order data. Specifically, the following method is adopted: collect historical work orders completed within a preset period and their corresponding original dialogue data, extract the original entity information involved in the generation process of each historical work order and the standardized entity information in the final work order, calculate the matching degree distribution between the two, and simultaneously extract the core handling intention and the set of actual executed operation instructions corresponding to each historical work order, calculate the fit degree distribution between the intention and the instructions, use the lower quartile of the matching degree distribution and the fit degree distribution as the benchmark, and make fine adjustments in combination with the review opinions of business experts on the quality of work orders, and finally set the first preset threshold to 0.95 and the second preset threshold to 0.90. For example, in a financial technology operations and maintenance scenario, historical data shows that in 95% of completed work orders, the entity information after verification and completion matches the entity information extracted from the original dialogue with a degree of over 0.95. At the same time, in 90% of the work orders, the actual operation instructions executed are highly consistent with the dialogue intent. Therefore, these two thresholds are set to 0.95 and 0.90, respectively. When the entity matching degree in the newly generated second structured data set is lower than 0.95 or the instruction arrangement consistency is lower than 0.90, a manual review or completion process is triggered to ensure the accuracy and executability of the work order generation.

[0056] S5. The second structured dataset after verification and completion is output as a work order generation instruction to the creation interface of the work order management system to drive the work order management system to automatically instantiate and persist a corresponding structured work order record based on the instruction; and the third control instruction is output as an automated orchestration instruction to the control loop of the automated operation and maintenance engine to instruct the automated operation and maintenance engine to schedule and execute the corresponding atomic operation and maintenance operation according to the execution logic, sequence and resource allocation scheme defined by the instruction.

[0057] It should be further explained that the semantic parsing model in this embodiment includes an embedding layer, an encoding layer, a first decoding branch, a second decoding branch, an attention mechanism, a coreference resolution layer, and an information fusion layer;

[0058] Further explanation is needed; please refer to [link / reference]. Figure 2 This embodiment calculates and generates a first structured data set representing the core processing intentions and key entity information of the dialogue, including:

[0059] S200. Normalize the instant conversation content by combining it with a pre-trained Chinese Bert model to obtain a normalized conversation text sequence. Specifically: Normalize the instant conversation content by combining it with a pre-trained Chinese Bert model. The specific processing logic includes first cleaning the noise in the original conversation text, removing HTML tags, extra spaces, special symbols and irrelevant characters. For example, clean the original input "Server 192.168.1.1 DOWN!! Emergency!! #Operation and Maintenance" to "Server 192.168.1.1 DOWN Emergency Operation and Maintenance". Then, use the WordPiece tokenizer自带的WordPiece分词器对清洗后的文本进行分词处理,将其切分为适合模型输入的子词或字词元序列,例如上述文本可被切分为“服”“务”“器”“192”“.”“168”“.”“1”“.”“1”“DOWN”“了”“紧”“急”“运”“维”等词元,然后对词元进行归一化操作,包括将英文单词统一转换为小写形式如将“DOWN”转为“down”,并对数字、IP地址、时间戳等特殊格式进行标准化或保留原样以便后续实体识别,同时根据实际需求过滤掉对意图识别无实际贡献的停用词如某些语气词和助词,但通常保留关键实体信息,最终将处理后的词元序列组合成规范化的对话文本序列,作为中文Bert模型的输入,从而得到包含对话核心意图和关键实体信息的向量表示,这一过程有效提升了模型对运维对话的理解准确性和鲁棒性。

[0060] S201. Based on the normalized conversation text sequence, map each token to a dense vector through the word vector model in the embedding layer to obtain a word vector sequence.

[0061] It should be further noted that in this embodiment, mapping each token to a dense vector through the word vector model in the embedding layer includes:

[0062] S2011. Based on the normalized conversation text sequence, obtain a token index sequence through the mapping operation from token to index. The mapping operation refers to matching each token string with a predefined vocabulary, which stores the correspondence between known tokens and unique integer indices, and returns the corresponding index if the match is successful. For example, for the normalized conversation text sequence "Server 192.168.1.1 CPU is too high", the token "Server" matches the vocabulary and returns the index 1024, "CPU" returns the index 356, "too high" returns the index 2001, and "192.168.1.1" does not appear in the vocabulary and returns a unified unknown token index 0. Finally, obtain the token index sequence [1024, 0, 356, 2001].

[0063] S2012. Based on the lexical index sequence, an initial dense vector sequence is obtained through an index-to-vector lookup operation. This operation utilizes a pre-trained embedding matrix, and reads the corresponding row vector from the matrix according to the lexical index as the initial dense vector of that lexical. In this embodiment, the embedding matrix is ​​trained on an operational domain corpus using the Skip-gram algorithm of the Word2Vec model. For example, for the above index sequence [1024, 0, 356, 2001], the 1024th row of the embedding matrix is ​​read to obtain the 300-dimensional initial vector v_ser corresponding to the server, the 0th row is read to obtain the initial vector v_unk corresponding to the unknown lexical, the 356th row is read to obtain the initial vector v_cpu corresponding to the CPU, and the 2001st row is read to obtain the initial vector v_high corresponding to the high. These vectors form the initial dense vector sequence [v_ser, v_unk, v_cpu, v_high] in sequence.

[0064] S2013. Based on the initial dense vector sequence, a domain-adaptive dense vector sequence is obtained through a domain-adaptive fine-tuning operation. The domain-adaptive fine-tuning operation includes the following steps: performing a nonlinear transformation on each initial dense vector in the initial dense vector sequence using a fully connected neural network; the weight parameters of the fully connected neural network are obtained through fine-tuning on preprocessed target ChatOps dialogue data; the fine-tuning learning aims to adapt to semantic understanding in the operations and maintenance domain, calculates the loss value between the model prediction result and the labeled label using the cross-entropy loss function, and minimizes the loss value using the Adam optimizer to update the weights of the fully connected neural network. The Adam optimizer has several parameters: an initial learning rate of 1e-5 and a weight decay coefficient of 1e-4; a preset iteration range of 10-50 rounds for fine-tuning learning, with an early stopping mechanism that terminates fine-tuning when the loss value on the validation set fails to decrease for three consecutive rounds; and preprocessed target ChatOps dialogue data, which is structured training data associated with O&M domain intent classification labels and entity annotation information after word segmentation, lexical mapping, and denoising. The fully connected neural network obtained through the fine-tuning operation performs a vector-wise transformation on the initial dense vector sequence, resulting in a domain-adapted dense vector sequence that fits the specific O&M context. For example, the initial dense vector sequence [v_ser, v_unk, v_cpu, v_high] is input into the fine-tuned fully connected neural network, which performs a non-linear transformation on each vector, outputting a new domain-adapted vector sequence [v'_ser, v'_unk, v'_cpu, v'_high], where vectors such as v'_ser more accurately express the semantic information within the O&M context.

[0065] S2014. For lexical units in the normalized dialogue text sequence that do not appear in the predefined vocabulary, a replacement dense vector is obtained through a sub-word vector combination operation. This operation first uses the BytePairEncoding algorithm to decompose out-of-vocabulary words into known sub-word units, and then calculates the arithmetic mean of the initial dense vectors corresponding to these sub-word units as the replacement vector. The replacement vector refers to the dense vector generated by calculating the arithmetic mean of the initial dense vectors corresponding to each sub-word unit after decomposing the out-of-vocabulary word unit into known sub-word units using the BytePairEncoding algorithm for the out-of-vocabulary word unit that does not appear in the predefined vocabulary in the normalized dialogue text sequence; this vector is used to replace the original word vector of the out-of-vocabulary word unit in subsequent processing. For example, for the unknown word "192.168.1.1", the BytePairEncoding algorithm decomposes it into the sub-word sequence "192", ".", "168", ".", "1", ".", "1". These sub-words all exist in the predefined vocabulary. The corresponding initial dense vectors v_192, v_dot1, v_168, v_dot2, v_1, v_dot3, and v_1 are taken from the embedding matrix respectively. The arithmetic mean of these vectors is calculated to obtain the replacement vector v_alt, which will be used to replace the original unknown word vector v_unk.

[0066] S2015. Based on the domain-adapted dense vector sequence and the alternative dense vector, a final word vector sequence is generated through vector selection and sorting operations. The logic of the vector selection and sorting operations in this embodiment is as follows: if a word element exists in the vocabulary, its domain-adapted dense vector is output; if it is an unregistered word element, its alternative dense vector is output. The final vectors of all words are arranged in the original text order to form a word vector sequence. For example, for the word sequence “server”, “192.168.1.1”, “CPU”, “too high”, where “server”, “CPU”, “too high” exist in the vocabulary, the domain adaptation vectors v'_ser, v'_cpu, and v'_high obtained in step S2013 are used. However, “192.168.1.1” is an out-of-vocabulary word, so the replacement vector v_alt generated in step S2014 is used. These vectors are combined in their original order to obtain the final word vector sequence [v'_ser, v_alt, v'_cpu, v'_high]. This sequence can be directly input into the subsequent encoding layer for dialogue understanding processing.

[0067] S202. Based on the word vector sequence, forward and backward sequence modeling is performed through a bidirectional long short-term memory network in the encoding layer to obtain a semantic vector representation of the dialogue that incorporates contextual information.

[0068] It should be further explained that this embodiment obtains the semantic vector representation of the dialogue that incorporates contextual information, including:

[0069] S2021. Based on the word vector sequence, a forward hidden state sequence is calculated using a forward long short-term memory network (LSTM). The LSTM performs LSM calculations on each word vector in the word vector sequence sequentially from the beginning to the end of the sequence, outputting the forward hidden state for each word. After traversing the entire sequence, the ordered forward hidden state sequence is obtained. For example, the word vector sequence generated in step S2015 is [v'_ser, v_alt, v'_cpu, v'_high], with a sequence length of 4 and each vector dimension of 300. The LSTM first receives the first word vector v'_ser, combines it with the initialized memory cell and hidden state to calculate and output the first forward hidden state h1_f, whose dimension is preset to 128. Then, h1_f and the second word vector v_alt are used as the input for the current step to calculate the second forward hidden state h2_f. This process continues until the last word vector v'_high, resulting in the fourth forward hidden state h4_f. After all calculations are completed, the forward hidden state sequence [h1_f, h2_f, h3_f, h4_f] is obtained by arranging them in order. Each element of this sequence corresponds to the positive contextual semantic information of the word at the original word order position.

[0070] S2022. Based on the word vector sequence, a backward hidden state sequence is calculated using a backward long short-term memory network. The backward long short-term memory network performs long short-term memory network calculations on each word vector in the word vector sequence in reverse order from the end to the beginning of the sequence, and outputs the backward hidden state of each word. After traversing the entire sequence, a backward hidden state sequence arranged in the original word order is obtained. For example, given the word vector sequence [v'_ser, v_alt, v'_cpu, v'_high], the backward long short-term memory network first receives the last word vector v'_high, and calculates it using the initialized memory cells and hidden states. The output is the backward hidden state h4_b corresponding to the last word in the original sequence. This state has the same 128 dimensions as the forward network, but captures the backward semantics starting from the end of the sequence. Next, h4_b and the second-to-last word vector v'_cpu are used as input to calculate the backward hidden state h3_b corresponding to the third word in the original sequence. This process is repeated in reverse until the first word vector v'_ser, yielding the backward hidden state h1_b corresponding to the first word in the original sequence. Since the network calculates in reverse order, the output states need to be rearranged according to the original word order, ultimately obtaining the backward hidden state sequence [h1_b, h2_b, h3_b, h4_b], where each element reflects the contextual semantic information from that word onwards until the end of the sequence.

[0071] S2023. Based on the forward hidden state sequence and the backward hidden state sequence, an encoder hidden state sequence is generated through vector concatenation operation, which is the semantic vector representation of the dialogue; the vector concatenation operation connects the forward hidden state vector and the backward hidden state vector at each position in the sequence to generate a bidirectional fused hidden state vector of the word at that position; the vectors generated at all positions are arranged in the original word order of the normalized dialogue text sequence to form the encoder hidden state sequence. For example, for the first word "server" in the original sequence, its forward hidden state is h1_f (128 dimensions) and its backward hidden state is h1_b (128 dimensions). Concatenating the two results in a 256-dimensional bidirectional fusion vector h1, i.e., h1 = [h1_f; h1_b]. For the second word "192.168.1.1", h2_f and h2_b are concatenated to obtain h2. The third word "CPU" corresponds to h3_f and h3_b, which are concatenated to obtain h3. The fourth word "too high" corresponds to h4_f and h4_b, which are concatenated to obtain h4. Finally, the encoder's hidden state sequence [h1, h2, h3, h4] is generated by arranging the words in the original word order. Each 256-dimensional vector in this sequence simultaneously fuses the forward and backward contextual semantics of the word's position, fully representing the semantic information of the entire dialogue text, and can be directly used for subsequent intent classification and entity recognition tasks.

[0072] S203. Based on the semantic vector representation of the dialogue, a nonlinear transformation is performed through the fully connected layer in the first decoding branch, and then the probability distribution is calculated through the Softmax function to obtain the intent classification label and corresponding confidence level that characterize the core processing intent of the dialogue.

[0073] It should be further explained that, in this embodiment, obtaining the intent classification label and corresponding confidence level that characterizes the core processing intent of the dialogue includes:

[0074] S2031. Based on the encoder hidden state sequence, the sequence features are aggregated through a global max pooling operation to obtain a fixed-dimensional dialogue-level global feature vector. For example, the generated encoder hidden state sequence is [h1, h2, h3, h4], where each vector has a dimension of 256 and the sequence length is 4. The global max pooling operation takes the maximum value of all words in each dimension of the sequence. That is, for the i-th dimension, the maximum value is selected from h1[i], h2[i], h3[i], and h4[i] as the output value of that dimension, finally obtaining a 256-dimensional dialogue-level global feature vector v_global, which aggregates the most significant semantic features in the entire dialogue sequence.

[0075] S2032. Based on the dialogue-level global feature vector, a first fully connected neural network is used to perform a nonlinear feature transformation to obtain a first intermediate feature vector; wherein the first fully connected neural network applies a rectified linear unit activation function after performing the nonlinear feature transformation. For example, the 256-dimensional dialogue-level global feature vector v_global is input into the first fully connected neural network, which contains 128 neurons, with a weight matrix size of 128×256 and a bias vector dimension of 128. After a linear transformation y=W·v_global+b, a 128-dimensional vector is obtained, and then the ReLU activation function f(x)=max(0,x) is applied element-wise to obtain the first intermediate feature vector v_mid1, which retains the nonlinear features after activation.

[0076] S2033. Based on the first intermediate feature vector, a second fully connected neural network is used to perform a nonlinear feature transformation to obtain a second intermediate feature vector; wherein the output dimension of the second fully connected neural network is the same as the preset total number of intent categories. For example, the first intermediate feature vector v_mid1 has a dimension of 128, which is input into the second fully connected neural network. The output dimension of this network is preset to 8, corresponding to 8 intent categories. Its weight matrix size is 8×128, and its bias vector dimension is 8. After linear transformation, an 8-dimensional second intermediate feature vector v_mid2 is obtained. Each element of this vector represents the original score of the corresponding category, which has not yet been normalized to a probability.

[0077] S2034. Based on the second intermediate feature vector, the probability of it belonging to each preset intent category is calculated using the Softmax function to obtain a probability distribution vector. It should be further explained that the preset intent categories in this embodiment are set based on the business logic and historical data analysis of the target operation and maintenance scenario. Specifically, this includes: First, sorting through historical work orders, operation and maintenance dialogue records, and domain knowledge documents to abstract high-frequency and critical business operation purposes, forming an initial category list; Second, combining the IT service management standard framework and internal operation and maintenance processes, merging, splitting, and standardizing the naming of these categories to ensure they are mutually exclusive and comprehensively covered; Finally, forming a clear and operable classification system such as "fault report," "resource application," "configuration change," "information query," and "compliance risk reporting," and solidifying it into the model training label set and rule matching library as the basis for intent classification. For example, the 8-dimensional second intermediate feature vector v_mid2 is input into the Softmax function to calculate the probability of each category, resulting in a probability distribution vector p=[p1, p2, p3, p4, p5, p6, p7, p8], where p1 corresponds to "fault report", p2 to "resource request", p3 to "configuration change", p4 to "information query", p5 to "compliance risk report", p6 to "permission request", p7 to "problem consultation", and p8 to "other". The sum of each probability value is 1. For example, p=[0.05, 0.80, 0.02, 0.06, 0.03, 0.02, 0.01, 0.01] indicates that the model predicts the highest probability that the current dialogue intent is "resource request".

[0078] S2035. Based on the probability distribution vector, by finding the maximum value and its index, the preset intent category corresponding to the maximum probability value is output as the intent classification label, and the maximum probability value is output as the confidence level. For example, for the above probability distribution vector p, the maximum value 0.80 appears at index 1, corresponding to the preset intent category "resource request". Therefore, the output intent classification label is "resource request", and the confidence level is 0.80, indicating that the model determines the current dialogue intent as a resource request with 80% confidence.

[0079] S204. Based on the semantic vector representation of the dialogue, sequence labeling is performed through the conditional random field model in the second decoding branch. Each word in the normalized dialogue text sequence is classified to obtain a preliminary key entity information set. Each element in the preliminary key entity information set includes an entity type label and the start and end position index of the entity in the normalized dialogue text sequence.

[0080] It should be further explained that the specific implementation steps for obtaining the preliminary set of key entity information in this embodiment include:

[0081] S2041. Based on the encoder hidden state sequence, calculate the score of each lexical belonging to each predefined entity type through a fully connected linear transformation to obtain the emission score vector sequence. For example, taking the dialogue "The CPU utilization of the database master database has reached 95%, requiring an emergency restart" in an IT operations and maintenance scenario as an example, following the encoder hidden state sequence obtained in step S2023, assume that the normalized lexical sequence of this dialogue is ["database", "master database", "CPU", "utilization", "reached", "95%", "required", "emergency", "restart"], and the hidden state vector corresponding to each lexical has a dimension of 256. Input the 256-dimensional vector of each lexical into a fully connected linear layer, which contains a weight matrix W with a dimension of [total number of entity types × 256] and a bias vector b with a dimension of [total number of entity types]. By calculating y = W·h + b, obtain the score of each lexical belonging to each predefined entity type. Assuming that the predefined entity types include 11 categories such as fault objects, performance indicators, status values, handling actions, connectors, and non-entity labels, each word will output an 11-dimensional score vector. The score vectors of all words will be arranged in the original word order to form a sequence of emission score vectors with the shape [9, 11].

[0082] S2042. Based on the emission score vector sequence and a pre-trained entity label transition matrix, the globally optimal entity label sequence is calculated using the Viterbi decoding algorithm to obtain the entity type label corresponding to each word. For example, the pre-trained entity label transition matrix T is loaded, with dimensions [11, 11], and T[u][v] represents the log probability of transitioning from entity label u to label v. The Viterbi decoding algorithm first initializes the cumulative score vector of the first word as the emission score vector of the first word plus the transition probability from the starting label to each label (the starting label is usually a preset special label); then it iterates through each subsequent word, and for each possible label v at the current position, it calculates the sum of the cumulative score corresponding to all previous position labels u, the transition probability from u to v, and the emission score of the current position label v, takes the maximum value as the cumulative score of the current position label v, and records the previous label u that makes the maximum value as the optimal path source; after the traversal is completed, starting from the last word, it backtracks according to the recorded optimal path source to determine the entity type label corresponding to each word in turn. The final result is the entity tag sequence of the dialogue, such as [fault object, fault object, performance metric, performance metric, non-entity, status value, non-entity, action, action], where non-entity means that the word does not belong to any predefined entity.

[0083] S2043. Based on the normalized dialogue text sequence and the globally optimal entity label sequence, identify word segments consecutively labeled with the same non-empty entity type label; for each identified word segment, record its corresponding entity type label, and determine the starting word segment position index and ending word segment position index based on the normalized dialogue text sequence. For example, align the word segment sequence and label sequence word by word: the word segment "database" is labeled as a fault object, "main database" is labeled as a fault object, "CPU" is labeled as a performance indicator, "utilization rate" is labeled as a performance indicator, "reached" is labeled as a non-entity, "ninety-five percent" is labeled as a status value, "needed" is labeled as a non-entity, "urgent" is labeled as a handling action, and "restart" is labeled as a handling action. This process identifies consecutive lexical segments labeled with the same non-empty entity type: Segment 1 covers lexical indices 0 to 1, entity type is faulty object, and text content is "database master"; Segment 2 covers lexical indices 2 to 3, entity type is performance indicator, and text content is "CPU utilization"; Segment 3 covers lexical indices 5 to 5, entity type is status value, and text content is "95%"; Segment 4 covers lexical indices 7 to 8, entity type is handling action, and text content is "emergency restart". The entity type, text content, start index, and end index of each segment are recorded.

[0084] S2044. Based on all identified word segments, their corresponding entity type labels, and start and end position indices, construct data elements containing the above information, and the preliminary key entity information set is composed of all data elements. For example, this embodiment takes the dialogue "The CPU utilization of the database master database has reached 95%, and an emergency restart is required" in an IT operation and maintenance scenario as an example. Specifically, based on the encoder hidden state sequence of the dialogue, the high-dimensional vector of each word is input into a fully connected linear layer. Through the linear transformation of parameter matrix multiplication and bias vector superposition, the score of each word corresponding to the predefined entity type such as fault object, performance index, status value, handling action, and connector is output. All scores constitute the emission score vector sequence. The pre-trained label transition matrix storing the transition probability between entity types is loaded. The Viterbi decoding algorithm calculates the globally optimal label sequence with a dynamic programming strategy. First, the emission score of the first word is used as the initial cumulative score. Then, subsequent words are traversed. The cumulative score of each entity type is calculated by "the cumulative score of the previous word + the transition probability + the emission score of the current word" and the optimal path is recorded. After backtracking, the entity label of each word is obtained. The dialogue text sequence and the label sequence are aligned word by word. Word segments with the same non-empty entity type are identified. The entity type, text content, and start and end position index of each segment are recorded. Data elements are constructed based on this information. After summarizing, a preliminary set of key entity information is formed. For example, based on the four identified lexical fragments, four data elements are constructed. Each data element contains an entity type label, entity text content, a starting lexical position index, and an ending lexical position index, ultimately resulting in a preliminary set of key entity information: [{Entity type: Faulty object, Text: Database master, Starting index: 0, Ending index: 1}, {Entity type: Performance indicator, Text: CPU utilization, Starting index: 2, Ending index: 3}, {Entity type: Status value, Text: 95%, Starting index: 5, Ending index: 5}, {Entity type: Action, Text: Emergency restart, Starting index: 7, Ending index: 8}]. This set accurately extracts the core entity information from the dialogue, providing structured input for subsequent work order generation or automated processing.

[0085] S205. Context-encode the standardized dialogue text sequence to obtain the semantic vector representation of the current round of dialogue and the semantic vector representation of each historical round of dialogue. Based on a preset context window, concatenate the semantic vector representation of the current round of dialogue with the semantic vector representation of the historical rounds of dialogue. Calculate the dependency weight of the current semantic vector on the historical semantic vector through an attention mechanism. Perform a weighted summation on the historical semantic vectors according to the dependency weights. Fuse the weighted summation result with the current semantic vector to obtain an enhanced semantic representation that incorporates the historical dialogue context.

[0086] The preset context window's value selection rules are as follows: the minimum threshold for window length is 3 rounds, and the maximum threshold is 10 rounds; the sliding step size is set to 1 unit based on dialogue rounds; the scope of historical dialogue rounds is traced back from the current round; the correlation between historical rounds and the current round's intent is calculated based on the semantic vector cosine similarity between the historical round's intent classification label and the current round's intent classification label, with a value range of [0, 1]; the preset correlation threshold range is [0.7, 0.9], dynamically configured according to the risk level of the operation and maintenance scenario; the window length adjustment rules are as follows:

[0087] 1) When the number of entities in the preliminary key entity information set corresponding to the historical round is ≥2 and the correlation degree is > preset correlation degree threshold, the window length is taken as the maximum threshold of 10 rounds;

[0088] 2) Conversely, the window length is calculated using the following formula: L = L_min + (L_max - L_min) × (N / N_max) × R, where L is the actual window length, L_min = 3, L_max = 10, N is the number of key entities in the historical rounds, N_max is the preset maximum key entity number threshold (value is 5), and R is the relevance value; the calculation result is rounded down and is not less than L_min;

[0089] Based on the above rules, the semantic vector representations of the historical rounds of dialogue and the semantic vector representations of the current round of dialogue are concatenated in chronological order to obtain a long semantic vector sequence.

[0090] It should be further explained that this embodiment calculates the dependency weights of the current semantic vector on the historical semantic vectors through an attention mechanism, and fuses the weighted summed vectors, including:

[0091] S2051. Based on the concatenated long semantic vector sequence, a query vector sequence, a key vector sequence, and a value vector sequence are generated through a linear projection layer and three independent fully connected layers. The first fully connected layer projects the vector sequence containing the attention positions to obtain the query vector sequence. The second fully connected layer projects the vector sequence containing the retrieved historical information to obtain the key vector sequence. The third fully connected layer projects the vector sequence containing the retrieved historical information to obtain the value vector sequence. For example, assuming the current round of dialogue is "It needs to be restarted," its normalized word sequence is ["needs," "restart," "it"], with a corresponding original semantic vector representation sequence length of 3 and each vector having 256 dimensions. The previous round of dialogue is "The database master's CPU utilization has reached 95%, requiring an emergency restart," with a semantic vector representation sequence length of 9 and each vector having 256 dimensions. Concatenating these two results in a long semantic vector sequence of length 12. The sequence is input into a linear projection layer and then passed through three independent fully connected layers with 256-dimensional input and 256-dimensional output to generate a query vector sequence Q, a key vector sequence K, and a value vector sequence V, respectively. Each sequence has a shape of 12×256, where the first 3 positions correspond to the current round's words and the last 9 positions correspond to the historical round's words.

[0092] S2052. Based on the query vector sequence and the key vector sequence, the attention weight matrix is ​​calculated through the scaling dot product attention module. First, the matrix product of the transpose of the query vector sequence and the key vector sequence is calculated to obtain the original attention score matrix. Then, each element in the original attention score matrix is ​​divided by the square root of the key vector dimension as a scaling factor. Subsequently, the Softmax function is applied to each row of the scaled attention score matrix to normalize it so that the sum of the elements in each row is one, thus obtaining the attention weight matrix that quantifies the semantic dependence of the current round dialogue word position on the word positions of all historical rounds. For example, for the query vector sequence Q (dimension 12×256) and key vector sequence K (dimension 12×256) generated in step S2051, first calculate the matrix product of the transposes of Q and K to obtain the original attention score matrix S (dimension 12×12), where S[i][j] represents the original attention score of the i-th word position to the j-th word position; then divide each element in S by the square root of the key vector dimension, √256=16, to obtain the scaled score matrix S_scaled; finally, apply the Softmax function to each row of S_scaled so that the sum of all elements in each row is 1, to obtain the attention weight matrix A (dimension 12×12), where A[i][j] represents the normalized attention weight of the i-th word position to the j-th word position. The first three rows of this weight matrix correspond to the attention distribution of the current round word to all words (including historical and current), which can clearly reflect the degree of attention of the current round word to key entities such as "database master" and "CPU utilization" in the historical round.

[0093] S2053. Based on the attention weight matrix and the value vector sequence, a weighted summation module is used to calculate the context vector. The matrix product of the attention weight matrix and the value vector sequence is calculated, and all position vectors in the value vector sequence are weighted and summed according to their attention weights. This generates a context vector aggregating relevant historical semantic information for each word position in the current round of dialogue. For example, multiplying the 12×12 dimension attention weight matrix A obtained in step S2052 with the 12×256 dimension value vector sequence V yields a 12×256 dimension context vector sequence C, where the i-th row of C equals the weighted sum of the i-th row of A and all rows of V, i.e., C[i] = Σ_jA[i][j]. V[j]. Since the first 3 rows of A focus on the historical information of the current round's word units, the first 3 vectors of C aggregate the historical information most relevant to the current word unit "need", "restart", and "it", respectively. For example, the context vector corresponding to "restart" will focus on the semantic information related to actions such as "emergency restart" in the history.

[0094] S2054. Based on the context vector and the original semantic vector representation of the current round, the semantic vector representation of the current round is fused with the residual connection and layer normalization module. First, the original semantic vector representation of the current round is added element by element to the corresponding context vector to perform residual connection. Then, the mean and variance of each vector in the addition result are calculated and standardized and rescaled to perform layer normalization processing, which stabilizes the training process and accelerates convergence. Finally, the enhanced semantic representation of the fused historical dialogue context is obtained. The enhanced semantic representation can associate entity information such as the performance index entity identifier of the fault object in the cross-round operation and maintenance dialogue, supporting subsequent referential resolution. For example, the first three vectors from the context vector sequence C are taken as the context vectors c1, c2, and c3 corresponding to the current round. These are then element-wise added to the original semantic vector representations v1, v2, and v3 of the current round (i.e., the current round vectors before concatenation with history), resulting in residual concatenated vectors r1, r2, and r3, each still 256-dimensional. Then, layer normalization is performed on each vector, that is, the mean and variance of all dimensions of each 256-dimensional vector are calculated, standardized, multiplied by a learnable scaling parameter, and a translation parameter is added to obtain the final enhanced semantic vectors e1, e2, and e3. The enhanced semantic representation sequence composed of these three vectors retains the original semantics of the current round of dialogue while incorporating key entity information from historical dialogues. For example, the vector e2 corresponding to the word "restart" already contains information such as the faulty object "database master" and the performance indicator "CPU utilization" mentioned in the history, so that when subsequent indices are delivered, they can be accurately associated with the preceding entities.

[0095] S206. Based on the enhanced semantic representation, the cosine similarity of vectors between different entity references in the preliminary key entity information set is calculated through the coreference resolution layer. Multiple entity references with cosine similarity exceeding a preset threshold are merged into the same entity identifier to obtain a unique key entity information set that eliminates referential ambiguity and has a unified identifier. In this embodiment, the preset threshold is determined by performing a grid search on the validation set with the goal of maximizing the harmonic mean of the accuracy and recall of entity reference merging.

[0096] It should be further explained that this embodiment obtains a unique set of key entity information that eliminates referential ambiguity and has a unified identifier, specifically including:

[0097] S2061. Based on the enhanced semantic representation and the preliminary key entity information set, obtain the denotation semantic vector of each entity reference. According to the start and end position index of each entity reference in the preliminary key entity information set, extract the bidirectional fusion hidden state vector of the corresponding word position from the enhanced semantic representation. Aggregate the extracted bidirectional fusion hidden state vector through the average pooling function to obtain the denotation semantic vector representing the overall semantics of each entity reference. For example, in a technology risk management scenario, following the enhanced semantic representation sequence obtained in step S2054 and the preliminary key entity information set generated in step S2044, assuming that the preliminary key entity information set contains three entity references, namely, "Supplier A" of the supplier type covering word index positions 5 to 6, "its database server" of the IT asset type covering word index positions 8 to 10, and "the server" of the IT asset type covering word index positions 12 to 13, two 256-dimensional vectors corresponding to positions 5 to 6, three 256-dimensional vectors corresponding to positions 8 to 10, and two 256-dimensional vectors corresponding to positions 12 to 13 are extracted from the enhanced semantic representation sequence. Average pooling is then performed on the vector sequence extracted for each entity reference, that is, the average value of all vectors is calculated for each dimension, resulting in three 256-dimensional reference semantic vectors v_supplier, v_asset1, and v_asset2, which represent the overall semantics of "Supplier A", "its database server", and "the server", respectively.

[0098] S2062. Based on the semantic vectors of all entity references, calculate the semantic correlation between each two different entity references using the cosine similarity function. For any two semantic vectors of references, first calculate their dot product, then calculate the Euclidean norm of each of the two semantic vectors of references, and divide the dot product result by the product of the two Euclidean norms to obtain the cosine similarity score between negative one and positive one. For example, for the three semantic vectors obtained in step S2061, the dot product of v_asset1 and v_asset2 is 0.85, the norm of v_asset1 is 1.0, the norm of v_asset2 is 1.0, and their cosine similarity is 0.85; the dot product of v_supplier and v_asset1 is 0.12, the norm of v_supplier is 1.0, the norm of v_asset1 is 1.0, and their cosine similarity is 0.12; the dot product of v_supplier and v_asset2 is 0.08, the norm of v_supplier is 1.0, the norm of v_asset2 is 1.0, and their cosine similarity is 0.08. This indicates that "its database server" and "this server" have high semantic similarity, while their semantic similarity with "supplier A" is low.

[0099] S2063. Based on the cosine similarity scores of all pairwise entity references, entity references are merged using agglomerative hierarchical clustering algorithm. During algorithm initialization, each entity reference is treated as an independent cluster. The algorithm continuously searches for the two clusters with the highest cosine similarity scores that exceed a preset threshold and merges them. During the merging process, the average linking strategy is used to calculate the distance between clusters, which is the average cosine similarity between all reference semantic vectors in the two clusters, until there are no clusters that meet the merging conditions. Each cluster formed in the end corresponds to the same entity. For example, assuming the preset merging threshold is set to 0.70, based on the cosine similarity score obtained in step S2062, during initialization, the three entity references each independently form cluster 1 corresponding to "supplier A", cluster 2 corresponding to "its database server", and cluster 3 corresponding to "the server". The highest similarity score is found to be 0.85 between cluster 2 and cluster 3. This value exceeds the threshold of 0.70, so cluster 2 and cluster 3 are merged into a new cluster 4. At this time, cluster 1 and cluster 4 remain. The inter-cluster distance between cluster 1 and cluster 4 is calculated. Since cluster 4 contains two reference vectors, the average cosine similarity between the vectors in cluster 1 and all vectors in cluster 4 is calculated using the average linking strategy. The result is 0.10, which does not exceed the threshold of 0.70, so no further merging is performed. Finally, two clusters are obtained: cluster 1 corresponds to the entity "supplier A", and cluster 4 corresponds to the two references of the same entity, "its database server" and "the server".

[0100] S2064. Based on the final cluster results obtained from the agglomerative hierarchical clustering algorithm, a globally unique entity identifier is generated for each final cluster using an identifier allocation function. The entity identifier generated by the identifier allocation function is either a string identifier based on a universally unique identification code algorithm or an integer identifier based on a sequence number. For example, for the two clusters finally obtained in step S2063, the identifier allocation function generates a globally unique entity identifier "550e8400-e29b-41d4-a716-446655440000" as a string identifier for cluster 1, and generates a globally unique entity identifier "550e8400-e29b-41d4-a716-446655440001" for cluster 4, ensuring that each entity cluster has a unique and distinguishable identity throughout the entire dialogue processing flow.

[0101] S2065. Based on globally unique entity identifiers and a preliminary set of key entity information, a unique set of key entity information is generated through information update operations: Each entity reference data element in the preliminary set of key entity information is traversed, and the globally unique entity identifier of the cluster to which the entity reference belongs is written into the corresponding data element, replacing the original temporary identifier. The set of all updated data elements is the unique set of key entity information that eliminates referential ambiguity and has a unified identifier. For example, in a technology risk management scenario, the preliminary set of key entity information contains three entity references: "Supplier A" (supplier type), "its database server" (IT asset type), and "the server" (IT asset type). After the above steps, the cosine similarity score of "its database server" and "the server" exceeds the preset merging threshold, so they are merged into the same cluster and assigned the same entity identifier "550e8400-e29b-41d4-a716-446655440001", while "Supplier A" forms an independent cluster and is assigned a different entity identifier "550e8400-e29b-41d". The code iterates through these three entity reference data elements, replacing the temporary identifier of "Supplier A" with "550e8400-e29b-41d4-a716-446655440000", and replacing the temporary identifiers of "its database server" and "the server" with "550e8400-e29b-41d4-a716-446655440001" respectively. The set of the updated three data elements is the unique key entity information set. The entity references in this set eliminate referential ambiguity, and different references to the same entity have a unified global identifier.

[0102] S207. Based on the intent classification label and the unique key entity information set, the data format is assembled and serialized according to a preset key value through an information fusion layer to obtain a first structured data set representing the core processing intent and key entity information of the dialogue, specifically:

[0103] S2071. Based on the intent classification label and its corresponding confidence level, create a top-level key-value pair. Use the string name of the intent classification label as the key and the corresponding confidence level value as the value to form a core data pair containing the core intent of the dialogue and its confidence level. For example, in a technology risk management scenario, the intent classification label is "risk reporting," and its corresponding confidence level is 0.96. The created core data pair would be {"risk reporting": 0.96}.

[0104] S2072. Based on the unique key entity information set, perform a structured transformation of entity information. Traverse each entity data element in the unique key entity information set, using the entity type label string in the entity data element as the subkey and the dictionary containing the entity text content and the unified entity identifier as the subvalue, and organize all the transformed entity subkey and subvalue pairs into a nested dictionary structure. For example, the unique key entity information set contains two entity data elements. The first entity type label is "Supplier", the entity text content is "Supplier A", and the unified entity identifier is "ENT-001". This is converted to the subkey "Supplier" corresponding to the subvalues ​​{"text": "Supplier A", "entity_id": "ENT-001"}. The second entity type label is "IT Asset", the entity text content is "Database Server", and the unified entity identifier is "ENT-002". This is converted to the subkey "IT Asset" corresponding to the subvalues ​​{"text": "Database Server", "entity_id": "ENT-002"}. Finally, we get the nested entity dictionary {"Supplier": {"text": "Supplier A", "entity_id": "ENT-001"}, "IT Asset": {"text": "Database Server", "entity_id": "ENT-002"}}.

[0105] S2073. Based on the top-level key-value pairs and the nested entity dictionary structure, perform data assembly, assign the nested entity dictionary structure as a value to the key named "entities", form a core data pair containing the key "entities", and merge this core data pair with the core data pair formed in S2071 to obtain a complete two-level structure dictionary object. For example, assigning the value of the nested entity dictionary {"Supplier": {"text": "Supplier A", "entity_id": "ENT-001"}, "IT Asset": {"text": "Database Server", "entity_id": "ENT-002"}} to the key "entities" results in {"entities": {"Supplier": {"text": "Supplier A", "entity_id": "ENT-001"}, "IT Asset": {"text": "Database Server", "entity_id": "ENT-002"}}}. This is then merged with the top-level key-value pair {"Risk Reporting": 0.96} to obtain the complete dictionary {"Risk Reporting": 0.96, "entities": {"Supplier": {"text": "Supplier A", "entity_id": "ENT-001"}, "IT Asset": {"text": "Database Server", "entity_id": "ENT-002"}}}.

[0106] S2074. Based on the complete two-layer structured dictionary object, call the standard JSON serialization function to convert the dictionary object into a machine-readable string according to the JSON specification, obtaining the first structured data set. For example, performing JSON serialization on the above complete dictionary yields the string {"Risk Reporting": 0.96, "entities": {"Supplier": {"text": "Supplier A", "entity_id": "ENT-001"}, "IT Assets": {"text": "Database Server", "entity_id": "ENT-002"}}}. This structured data fully encodes the core intent of the dialogue and key entity information, and can be directly used as input for the work order creation interface.

[0107] It should be further explained that this embodiment allocates tasks between preset work order information fields and automated operation and maintenance operations to be triggered, including:

[0108] S301. Based on the intent classification labels and corresponding confidence levels in the first structured data set, a decision tree classifier with confidence threshold verification is used to match predefined work order template selection rules to filter paths whose confidence levels do not meet the threshold, thereby obtaining the target work order template identifier. It should be further noted that the pre-set confidence threshold in this embodiment is dynamically set based on a dual dimension of historical work order accuracy statistical analysis and business scenario risk level classification. The specific setting process is as follows: First, historical intent classification label samples, corresponding confidence levels, and work order matching result data within a preset time range are collected. The work order matching accuracy, recall, and F1 score under different confidence threshold intervals are statistically analyzed, with the goal of maximizing the F1 score to determine the confidence benchmark threshold interval. Second, according to the risk level classification rules for operation and maintenance business scenarios, the operation and maintenance scenarios corresponding to the intent classification labels are divided into high-risk, medium-risk, and low-risk scenarios. High-risk scenarios include database failures, server crashes, and other events that may cause business interruptions. The system is divided into several risk levels. Medium-risk scenarios include those affecting business performance, such as abnormal resource utilization or deviations in configuration parameters. Low-risk scenarios include those without business impact, such as status queries or information inquiries. The system then maps the confidence threshold range to different risk levels. For high-risk scenarios, the lower limit of the threshold range is used to reduce the probability of missed detections. For medium-risk scenarios, the middle value of the threshold range is used to balance accuracy and recall. For low-risk scenarios, the upper limit of the threshold range is used to improve matching accuracy. Finally, a periodic threshold update mechanism is configured. Historical work order data is re-collected at a preset cycle, and the above statistical analysis and risk mapping steps are executed to achieve dynamic iterative optimization of the confidence threshold. Based on the confidence threshold determined in the above settings, a filtering operation is performed on the confidence values ​​corresponding to the intent classification tags. Branch paths with confidence values ​​below the threshold are removed to ensure that the matching results between the selected intent classification tags and the work order template meet the preset accuracy and risk control requirements.

[0109] It should be further explained that, in this embodiment, obtaining the target work order template identifier includes:

[0110] S3011. Based on predefined work order template selection rules, a rule tree is constructed using the C4.5 decision tree algorithm. The predefined work order template selection rules define the mapping relationship between intent classification labels and standardized work order template identifiers. The C4.5 decision tree algorithm uses intent classification labels as features and standardized work order template identifiers as classification targets to train on historical rule samples, generating the rule tree. The internal nodes of this rule tree contain judgment conditions based on confidence thresholds, and the leaf nodes store the corresponding standardized work order template identifiers. For example, in a technology risk management scenario, based on historical work order data, intent classification labels such as fault reporting, resource application, and risk reporting are used as input features, and corresponding work order template identifiers such as FAULT_TEMPLATE, RESOURCE_TEMPLATE, and RISK_TEMPLATE are used as classification targets. The root node of the trained rule tree is first divided according to the intent classification label. For the subtree of risk reporting intent, a verification node with a confidence threshold of 0.9 is set. Samples below this threshold need to be further divided into low-confidence branches.

[0111] S3012. Based on the intent classification label extracted from the first structured data set, a decision tree classifier is used to traverse the rule tree. The decision tree classifier takes the extracted intent classification label as input, starts from the root node of the rule tree, selects a branch path according to the node judgment conditions, and continues until a leaf node is reached. For example, if the intent classification label extracted from the first structured data set generated in the previous step is "risk reporting", this label is input into the decision tree classifier. The classifier starts from the root node and enters the subtree branch corresponding to risk reporting according to the label matching rules.

[0112] S3013. When traversing to an internal node containing a confidence threshold judgment condition, a threshold check is performed based on the confidence level extracted from the first structured data set through a numerical comparison operation. The numerical comparison operation compares the extracted confidence level with a preset confidence threshold stored in the internal node. For example, in the risk reporting branch, if the confidence threshold stored in the internal node is 0.9, and the confidence level corresponding to this intent is extracted from the first structured data set as 0.85, the numerical comparison operation determines whether 0.85 is less than 0.9.

[0113] S3014. If the result of the numerical comparison operation is that the extracted confidence level is less than the preset confidence threshold, then the traversal of the current branch path is terminated through the decision tree pruning operation. For example, since 0.85 is less than 0.9, the numerical comparison result is true, and the decision tree pruning operation immediately stops the continued traversal of the current branch and no longer visits the child nodes below this node.

[0114] S3015. When the traversal reaches a leaf node, the standardized work order template identifier is obtained by reading the data stored in that leaf node, and used as the target work order template identifier. For example, if the confidence level is greater than or equal to 0.9, the traversal continues downwards until a leaf node is reached. The standardized work order template identifier stored in this leaf node is RISK_TEMPLATE_HIGH_CONF, which is the target work order template identifier selected this time. In this example, the branch was pruned due to insufficient confidence, and it may actually enter another branch or trigger the default template selection process.

[0115] S302. Based on the target work order template identifier, query the template-field association database using a database query function that combines field priority sorting to obtain the set of required field names, the set of optional field names, and the field priority sorting result; it should be further explained that this embodiment queries the template-field association database using a database query function that combines field priority sorting, including:

[0116] S3021. Based on the target work order template identifier, access the preset template-field association database through a database query function. The template-field association database stores the mapping relationship between different work order template identifiers and their corresponding required field definitions, optional field definitions, and priority weights of each field. For example, assuming the target work order template identifier is "RISK_TEMPLATE", accessing the template-field association database through a database query function returns the mapping relationship corresponding to the template. The required field definitions include "risk description", "risk level", and "occurrence time", and the optional field definitions include "associated system" and "responsible person". The priority weights of each field are 10 for risk description, 9 for risk level, 8 for occurrence time, 5 for associated system, and 3 for responsible person.

[0117] S3022. Based on the mapping relationship returned by the database query function, the fields are separated by attributes using a field classification function. The field classification function, according to preset attribute identifiers, divides the returned field set into a set of required field names and a set of optional field names. The set of required field names contains the field names that must be filled in to create this type of work order, while the set of optional field names contains field names that can be selectively filled based on information completeness. For example, based on the above-returned mapping relationship, the field classification function divides "Risk Description," "Risk Level," and "Occurrence Time" into the set of required field names, and "Associated System" and "Responsible Person" into the set of optional field names.

[0118] S3023. Based on the field priority weights contained in the mapping relationship returned by the database query function, a sorting function is used to sort the field names in the required field name set and the optional field name set. The sorting function arranges the fields in descending order according to their priority weights, obtaining the sorting results for the required field names and the optional field names. For example, based on the returned field priority weights, the sorting function sorts the required field name set in descending order of weights to obtain the sorting results for risk description, risk level, and occurrence time, and sorts the optional field name set in descending order of weights to obtain the sorting results for associated systems and responsible persons.

[0119] S303. Based on the unique key entity information set and entity identifiers in the first structured data set, the mapping is completed and the entity identifiers are associated according to the iteratively updated entity type-work order field correspondence lookup table through a dynamically updated entity type-work order field mapping function, thereby obtaining the preliminary field filling mapping relationship. For example, in this embodiment, the correspondence between entity types and work order fields is constructed based on a combination of business rules and historical data statistics. First, domain experts define an acceptable set of entity types for each field according to the business meaning of each field in the work order template. For example, the risk description field corresponds to the risk item entity type, the supplier field corresponds to the supplier entity type, the contract field corresponds to the contract number entity type, and the asset field corresponds to the IT asset entity type. Then, the completed data is collected. The historical work order data is compiled, and the distribution of entity types actually filled in each work order field is statistically analyzed. The co-occurrence frequency and mutual information score of each entity type and work order field are calculated. Expert rules are used as prior constraints, and statistical results are used as weighting criteria. The final mapping relationship is determined by weighted voting or rule priority ranking. For example, in a technology risk management scenario, the risk description field is preferentially mapped to the risk item entity type according to expert rules. At the same time, historical statistics show that more than 90% of this field is filled by risk item entities. Therefore, the constructed correspondence is as follows: risk item entity type is mapped to risk description field, supplier entity type is mapped to supplier field, contract number entity type is mapped to contract field, and IT asset entity type is mapped to asset field. This correspondence is fixed in the mapping table for lookup during iterative updates.

[0120] It should be further explained that this embodiment uses a dynamically updated entity type-field mapping function to complete the mapping and associate entity identifiers based on an iteratively updated entity type-work order field correspondence lookup table, including:

[0121] S3031. Based on the entity type label contained in each entity data element of the unique key entity information set, access a preset entity type-work order field correspondence lookup table through a hash table lookup operation to obtain one or more candidate work order field names associated with each entity type label. The entity type-work order field correspondence lookup table is a dynamic data structure that supports entry creation, updating, and deletion operations through an external management interface, which stores the mapping relationship between entity type label strings and work order field name strings. For example, suppose the unique key entity information set contains two entity data elements: the first entity type label is supplier, the entity text content is supplier A, and the unique entity identifier is ENT-001; the second entity type label is IT asset, the entity text content is database server, and the unique entity identifier is ENT-002. Access the preset entity type-work order field correspondence lookup table through a hash table lookup operation. The mapping relationship stored in this table is that the candidate work order field name set associated with the supplier type label includes supplier name, supplier contact person, and supplier address; the candidate work order field name set associated with the IT asset type label includes asset name, asset number, and asset type.

[0122] S3032. Based on the acquired candidate work order field names, a target work order field name is determined from the candidate work order field names using a field selection strategy. The field selection strategy makes decisions based on the target work order template identifier and preset field priority rules to obtain the final target work order field name corresponding to each entity type label. For example, based on the candidate field names obtained in S3031, combined with the target work order template identifier RISK_TEMPLATE and the preset field priority rules, which stipulate that for risk reporting work orders, the supplier name field is preferred for supplier types and the asset name field is preferred for IT asset types, the target work order field name corresponding to the supplier type label is determined to be the supplier name, and the target work order field name corresponding to the IT asset type label is determined to be the asset name.

[0123] S3033. Based on the determined target work order field name and the entity text content and unique entity identifier contained in the corresponding entity data element in the unique key entity information set, a mapping relationship data unit is generated through a data assembly operation; the mapping relationship data unit contains at least three data items: target work order field name, entity text content, and unique entity identifier. For example, for a supplier entity data element, a mapping relationship data unit is generated based on its target work order field name "supplier name," entity text content "supplier A," and unique entity identifier "ENT-001." This unit contains the target work order field name "supplier name," the entity text content "supplier A," and the unique entity identifier "ENT-001." For an IT asset entity data element, a mapping relationship data unit is generated based on its target work order field name "asset name," entity text content "database server," and unique entity identifier "ENT-002." This unit contains the target work order field name "asset name," the entity text content "database server," and the unique entity identifier "ENT-002."

[0124] S3034. Based on the mapping relationship data units generated for all entities, the preliminary field filling mapping relationship is generated through a set construction operation; the preliminary field filling mapping relationship is a structured data set composed of multiple mapping relationship data units. For example, the above two mapping relationship data units are combined into a set to obtain the preliminary field filling mapping relationship. This set contains the first unit whose target work order field name is supplier name, entity text content is supplier A, and unique entity identifier is ENT-001; and the second unit whose target work order field name is asset name, entity text content is database server, and unique entity identifier is ENT-002, providing a structured data source for subsequent work order field filling.

[0125] S304. Based on the preliminary field filling mapping relationship, the unique key entity information set and the field priority sorting result, the entity text content is extracted according to priority through the field value filling function with uniqueness verification, and the field filling and deduplication are completed to generate the second structured data set.

[0126] It should be further explained that this embodiment uses a field value filling function with uniqueness verification to extract entity text content according to priority and complete field filling and deduplication, including:

[0127] S3041. Based on the sorting results of the required field names and the sorting results of the optional field names, a sequence of field names to be filled, arranged in priority order, is constructed through a list merging operation. The specific construction and purpose of the list merging operation are as follows: This operation receives the sorting results of the required field names and the sorting results of the optional field names as input. By placing all elements of the sorting results of the required field names at the beginning of the sequence and placing all elements of the sorting results of the optional field names at the end of the sequence in their original order, a new, integrated list of field names is generated. This list is the sequence of field names to be filled, and its purpose is to establish a strict processing order for subsequent field filling operations. For example, for a risk reporting work order template, the sorting results of the required field names are risk description, risk level, occurrence time, and supplier name; the sorting results of the optional field names are associated system, responsible person, and asset name; and the sequence of field names to be filled generated through the list merging operation is risk description, risk level, occurrence time, supplier name, associated system, responsible person, and asset name.

[0128] S3042. Based on the sequence of field names to be filled and the preliminary field filling mapping relationship, each field name element in the sequence of field names to be filled is processed sequentially through iterative control logic. The specific construction and purpose of this iterative control logic are as follows: This logic uses the sequence of field names to be filled as the iteration object, and uses an index counter to sequentially access the field names from the first element to the last element. For each iteration, the purpose of accessing the current field name is to initiate and sequentially execute the following four sub-steps S3042a to S3042d to calculate the filling value for a single field. For example, if the iterative control logic first accesses the current field name as "risk description," since there is no matching entry for "risk description" in the preliminary field filling mapping relationship, subsequent sub-steps will return null values. When the current field name is "supplier name," a sub-step will be initiated to calculate the filling value based on the supplier name.

[0129] S3042a. Based on the current field name, a conditional query operation is used to retrieve associated entries in the preliminary field filling mapping relationship to obtain a set of candidate field filling entries corresponding to the current field. The specific construction and purpose of the conditional query operation are as follows: This operation uses the current field name as the query key, traverses each field filling entry in the preliminary field filling mapping relationship set, and compares whether the field name stored in the entry matches the query key; all entries with matching field names are collected and returned, forming the candidate field filling entry set, whose purpose is to filter out all possible candidate data sources for filling the current field. For example, if the current field name is the supplier name, and an entry is found in the preliminary field filling mapping relationship with the target work order field name as supplier name, entity text content as supplier A, and unique entity identifier as ENT-001, this entry is collected to form the candidate field filling entry set, which contains only this one entry.

[0130] S3042b: Based on the candidate field filling item set, obtain the unique entity set associated with the current field through data extraction and deduplication operations. The specific construction and purpose of this operation are as follows: First, traverse the candidate field filling item set and extract the unique entity identifier associated with each item; then, add all extracted identifiers sequentially to an initially empty hash set data structure, utilizing the hash set's automatic exclusion of duplicate elements to achieve identifier deduplication. The hash set ultimately output by this operation is the unique entity set associated with the current field, its purpose being to ensure that subsequent content extraction is based only on non-repeating entities, avoiding information redundancy. For example, extracting the unique entity identifier ENT-001 from the candidate field filling item set and adding it to the hash set yields the unique entity set associated with the current field as ENT-001.

[0131] S3042c. Based on the unique entity set, extract the entity text content corresponding to each unique entity identifier from the unique key entity information set through a key-value retrieval operation. The specific construction and purpose of the key-value retrieval operation is as follows: This operation iterates through each unique entity identifier in the unique entity set, uses it as a retrieval key, searches for entity data elements with the same identifier in the unique key entity information set, and returns the entity text content string stored in that element. Its purpose is to obtain the original description information of each related entity. For example, iterates through the unique entity set ENT-001, retrieves the entity text content corresponding to the identifier as supplier A in the unique key entity information set, and returns that string.

[0132] S3042d. Based on all extracted entity text content, the final fill value for the current field is generated through multi-text fusion decision and execution operations. The specific construction and purpose of the multi-text fusion decision and execution operations are as follows: This operation first determines the number of received entity text contents; if the number is one, the text content is directly output as the final fill value; if the number is greater than one, multiple text contents are merged into a single string according to preset fusion rules, such as sorting all text contents lexicographically and then using a specific delimiter, and this merged string is output as the final fill value. Its purpose is to integrate information that may come from multiple entities into a standardized value suitable for a single work order field. For example, if the number of extracted entity text contents is one (supplier A), then supplier A is directly output as the final fill value for the current field "supplier name".

[0133] S3043. Based on the final fill value of each field in the sequence of field names to be filled, a mapping relationship creation operation is used to generate a set of correspondences between field names and final fill values. The specific construction and purpose of the mapping relationship creation operation is as follows: This operation synchronously traverses the sequence of field names to be filled and the final fill value corresponding to each field, creating a key-value pair data structure for each pair of field name-fill value, where the key is the field name and the value is the final fill value. All created key-value pairs are collected to form the key-value pair set, which is used to provide basic data units for constructing structured order data. For example, traversing the entire sequence of field names to be filled, the final fill value of the risk description field is empty, the risk level field is empty, the occurrence time is empty, the supplier name is supplier A, the associated system is empty, the responsible person is empty, and the asset name is database server. The corresponding key-value pair set is created, containing: Risk description: empty, Risk level: empty, Occurrence time: empty, Supplier name: supplier A, Associated system: empty, Responsible person: empty, Asset name: database server.

[0134] S3044. Based on the key-value pair set, a second structured data set is generated through a structured encapsulation operation. The specific construction and purpose of the structured encapsulation operation are as follows: This operation creates a new empty data structure conforming to a machine-readable format such as a JSON object, XML document, or a specific serialized object; then, iterates through the key-value pair set, inserting each key-value pair into the newly created data structure, where the key serves as the attribute name or element tag of the data structure, and the value serves as the corresponding attribute value or element content. The final generated data structure is the second structured data set, whose purpose is to organize discrete field-filling information into a complete structured data payload that can be directly used to drive the work order creation interface. For example, based on the above key-value pair set, a JSON object is created, and after encapsulation, the second structured data set is obtained as follows: Risk Description: null, Risk Level: null, Occurrence Time: null, Supplier Name: Supplier A, Related System: null, Responsible Person: null, Asset Name: Database Server. This JSON object can be directly used as the input payload for creating a risk reporting work order.

[0135] S305. Based on the intent classification labels, unique key entity information set and entity attribute information in the first structured data set, the rule engine loads the hierarchical condition action rule set through the automated operation triggering of multi-level condition matching, and generates an automated operation instruction set with execution priority identifier.

[0136] It should be further explained that this embodiment triggers the rule engine to load a hierarchical conditional action rule set through automated operations of multi-level condition matching, including:

[0137] S3051. Based on the pre-stored rule configuration file, the hierarchical conditional action rule set is loaded into memory and converted into an executable data structure through rule loading and parsing operations. The hierarchical conditional action rule set consists of multiple rules, each rule containing a condition part and an action part; the condition part is a multi-level logical expression used to match intent classification labels, entity types, and entity attribute values; the action part defines the atomic automated operation instructions to be triggered when the conditions are met and their execution priority identifiers. For example, in a technology risk management scenario, the rule configuration file pre-stores three rules. Rule 1's condition part is that the intent classification tag is risk reporting, the entity type is supplier, and the entity attribute value includes expired qualification. The action part is to trigger a supplier suspension of cooperation instruction and set the priority to 1. Rule 2's condition part is that the intent classification tag is risk reporting, the entity type is IT asset, and the entity attribute value includes CPU utilization exceeding 95%. The action part is to trigger a database performance diagnosis instruction and set the priority to 2. Rule 3's condition part is that the intent classification tag is risk reporting, the entity type is contract, and the entity attribute value includes near expiration. The action part is to trigger a contract renewal reminder instruction and set the priority to 3.

[0138] S3052. Based on the first structured data set, through data extraction operations, obtain the intent classification label, all entity type labels and corresponding entity text content contained in the unique key entity information set, and the entity attribute information. For example, extract the intent classification label "risk reporting" from the first structured data set, extract the entity type label "supplier" and its corresponding entity text content "supplier A" from the unique key entity information set, extract the entity type label "IT asset" and its corresponding entity text content "database server" from the entity type information set, and simultaneously extract entity attribute information from the dialogue context. The attribute information of supplier A includes expired qualifications, and the attribute information of the database server includes CPU utilization exceeding 95%.

[0139] S3053. Based on the hierarchical conditional action rule set, the intent classification label, the entity type label, and the entity attribute value, multi-level conditional matching is performed through the rule engine to obtain a list of activated rules. Please refer to [link to relevant documentation]. Figure 3 Specifically, it includes:

[0140] S3053a. Based on the intent classification label, the entity type label, and the entity attribute value, construct corresponding intent fact objects, entity type fact objects, and entity attribute fact objects, respectively. Through the fact insertion operation of the rule matching engine, insert the intent fact objects, entity type fact objects, and entity attribute fact objects into the pre-compiled Rete algorithm network. For example, risk reporting is constructed as an intent fact object, suppliers and IT assets are constructed as two entity type fact objects, and expired qualifications and CPU utilization exceeding 95% are constructed as two entity attribute fact objects, all of which are then inserted into the Rete algorithm network.

[0141] S3053b: Based on the fact objects inserted into the Rete algorithm network, perform a single-type atomic pattern matching operation through the Alpha nodes in the Rete algorithm network; wherein, the intent fact object is matched only with the pre-compiled intent matching Alpha node without any associated constraints, the entity type fact object is matched only with the pre-compiled entity type matching Alpha node without any associated constraints, and the entity attribute fact object is matched only with the pre-compiled entity attribute matching Alpha node without any associated constraints, and obtain all successfully matched Alpha nodes and their output matched fact objects. For example, if the risk reporting pattern of the intent fact object is successfully matched with the risk reporting pattern stored in the intent matching Alpha node, the risk reporting matching fact object is output; if the vendor of the entity type fact object is successfully matched with the vendor pattern stored in the entity type matching Alpha node, the vendor matching fact object is output; if the IT asset of the entity type fact object is successfully matched with the IT asset pattern stored in the entity type matching Alpha node, the IT asset matching fact object is output; if the qualification expired of the entity attribute fact object is successfully matched with the qualification expired pattern stored in the entity attribute matching Alpha node, the qualification expired matching fact object is output; if the CPU utilization rate of the entity attribute fact object exceeds 95% is successfully matched with the CPU utilization rate exceeding 95% pattern stored in the entity attribute matching Alpha node, the CPU utilization rate exceeding 95% matching fact object is output.

[0142] S3053c: Based on the matched fact object output by the successfully matched Alpha node, a multi-type fact association matching activation operation is performed through the Beta node in the Rete algorithm network. Specifically: the first-level Beta node receives the output from the intent matching Alpha node and the entity type matching Alpha node. When the matched fact objects from the two channels satisfy the predefined intent-entity type association constraint, the first-level Beta node is activated and outputs a first-level partial matching result containing intent and entity type. The second-level Beta node receives the output from the first-level Beta node and the entity attribute matching Alpha node. When the first-level partial matching result and the matched fact object from the entity attribute matching Alpha node satisfy the predefined intent-entity type-entity attribute association constraint, the second-level Beta node is activated and outputs a complete partial matching result containing intent, entity type, and entity attribute. For example, a Level 1 Beta node receives a risk reporting intent matching fact object and a supplier entity type matching fact object. If both satisfy the association constraint in Rule 1 that the intent is risk reporting and the entity type is supplier, the Beta node is activated and outputs a Level 1 partial matching result containing the risk report and the supplier. A Level 1 Beta node also receives a risk reporting intent matching fact object and an IT asset entity type matching fact object. If both satisfy the association constraint in Rule 2 that the intent is risk reporting and the entity type is IT asset, the Beta node is activated and outputs a Level 1 partial matching result containing the risk report and the IT asset. The second-level Beta node receives partial first-level matching results including risk reports and supplier information, and also receives entity attribute matching facts for entities with expired qualifications. When all three satisfy the association constraints in Rule 1 (intended for risk reporting, entity type as supplier, and entity attribute as expired qualification), the Beta node is activated and outputs a complete partial matching result including risk reports, supplier information, and expired qualification. The second-level Beta node also receives partial first-level matching results including risk reports and IT assets, and also receives entity attribute matching facts for entities with CPU utilization exceeding 95%. When all three satisfy the association constraints in Rule 2 (intended for risk reporting, entity type as IT asset, and entity attribute as CPU utilization exceeding 95%), the Beta node is activated and outputs a complete partial matching result including risk reports, IT assets, and CPU utilization exceeding 95%.

[0143] S3053d: Based on the complete partial matching results output by all activated Level 2 Beta nodes, a full match judgment operation is performed to determine whether the hierarchical condition action rule corresponding to each complete partial matching result is fully satisfied, and the rule for full matching is obtained. For example, for a complete partial matching result that includes risk reporting, supplier and qualification expiration, it is determined that its corresponding rule one is fully satisfied; for a complete partial matching result that includes risk reporting, IT assets and CPU utilization exceeding 95%, it is determined that its corresponding rule two is fully satisfied.

[0144] S3053e: Based on all rules determined to be complete matches, collect their rule identifiers through a list generation operation to obtain the list of activated rules. For example, collect the identifier R001 of rule one and the identifier R002 of rule two, and generate the list of activated rules as R001 and R002.

[0145] S3054. Based on the list of activated rules, an automated operation instruction is generated for each activated rule through an instruction generation operation. This operation reads the action part of the activated rule, instantiates the atomic operation template defined therein, and fills in the relevant entity text content extracted from the first structured data set as parameters. At the same time, it assigns the generated instruction an execution priority identifier read from the action part. All generated instructions constitute an initial set of automated operation instructions. For example, for rule one, its action part is to trigger a supplier to suspend cooperation instruction. The instruction template is instantiated, the entity text content supplier A is filled in as a parameter, and a priority identifier of 1 is assigned, generating the instruction suspend cooperation - supplier A; for rule two, its action part is to trigger a database performance diagnosis instruction. The instruction template is instantiated, the entity text content database server is filled in as a parameter, and a priority identifier of 2 is assigned, generating the instruction performance diagnosis - database server. The initial instruction set contains these two instructions.

[0146] S3055. Based on the initial set of automated operation instructions, the set of automated operation instructions with execution priority identifiers is obtained through priority sorting and conflict resolution operations. The specific logic includes:

[0147] S3055a. Based on the execution priority identifier contained in each instruction in the initial set of automated operation instructions, all instructions are sorted in descending order using a sorting algorithm to obtain the sorted instruction sequence. For example, in the initial instruction set, the priority identifier of "Pause Cooperation - Supplier A" is 1, and the priority identifier of "Performance Diagnosis - Database Server" is 2. After sorting in descending order, the sequence obtained is "Performance Diagnosis - Database Server" with priority 2 first, and "Pause Cooperation - Supplier A" with priority 1 second.

[0148] S3055b: Based on the sorted instruction sequence, analyze the resource dependencies and logical mutual exclusion relationships between instructions using a conflict detection algorithm, identify instruction pairs with resource conflicts or logical mutual exclusion, and obtain a list of conflicting instruction pairs. For example, conflict detection may find that the "Performance Diagnosis - Database Server" instruction requires the use of the database server's diagnostic interface resources, while the "Suspend Cooperation - Supplier A" instruction, although not directly using the same resource, may cause changes to the database server's access permissions if the supplier's "Suspend Cooperation" instruction is executed first, affecting the execution of the diagnostic instruction. Therefore, these two instructions are logically mutually exclusive, and are identified as conflicting instruction pairs and added to the conflicting instruction pair list.

[0149] S3055c. Based on the list of conflicting instruction pairs, execute a conflict resolution decision. This decision first selects and retains the instruction with the higher execution priority from the conflicting instruction pair based on the execution priority identifier, and removes the conflicting instruction with the lower execution priority. If the instructions in the conflicting instruction pair have the same execution priority identifier, or the high-priority instruction cannot be executed independently due to system resource constraints, or the conflicting instruction pair involves predefined critical operation and maintenance operations, then a manual arbitration sub-process is triggered. For example, in this case, the priority 2 of Performance Diagnosis - Database Server is higher than the priority 1 of Suspend Cooperation - Supplier A. Based on the priority decision, the higher-priority Performance Diagnosis - Database Server instruction is retained, and the lower-priority Suspend Cooperation - Supplier A instruction is removed, without triggering manual arbitration.

[0150] S3055d. When the manual arbitration sub-process is triggered, the detailed information, conflict type, and associated operation and maintenance objects of the conflicting instruction pair are pushed to the preset responsible person queue through the application programming interface of the operation and maintenance collaboration platform. Based on the arbitration decision fed back by the responsible person through the operation and maintenance collaboration platform within a preset time limit, the instruction selection, order adjustment, or resource coordination operation is performed to obtain the instruction update plan confirmed by arbitration. For example, assuming that the two instructions have the same priority in this case, or the high-priority instruction cannot be immediately executed for diagnosis because the database server is currently in a peak business period, manual arbitration is triggered. The detailed information of the conflicting instruction pair, Performance Diagnosis - Database Server and Suspension of Cooperation - Supplier A, is pushed to the operation and maintenance responsible person queue. The responsible person feeds back the arbitration decision on the platform, which is to execute Suspension of Cooperation - Supplier A to change permissions first, and then execute Performance Diagnosis - Database Server, to obtain the instruction update plan confirmed by arbitration.

[0151] S3055e: Based on the conflict resolution decision or the arbitration-confirmed instruction update scheme, through an instruction set update operation, abandoned conflicting instructions are removed from the sorted instruction sequence, or the execution order and resource allocation of related instructions are adjusted to generate the final automated operation instruction set with execution priority identifiers. For example, based on the priority decision result, "Suspend Cooperation - Supplier A" is removed from the sorted instruction sequence, and only "Performance Diagnosis - Database Server" is retained, generating the final automated operation instruction set, which contains one instruction, namely "Performance Diagnosis - Database Server," with a priority identifier of 2. If based on a manual arbitration scheme, the instruction order is adjusted to execute "Suspend Cooperation - Supplier A" first, followed by "Performance Diagnosis - Database Server," and the resource allocation status is updated to generate the final automated operation instruction set.

[0152] S306. Based on the execution priority, logical dependencies, and execution prerequisites of the automated operation instruction set, a directed acyclic graph with dependency conflict detection is used to construct and optimize the task execution flowchart to obtain the initial automated orchestration logic.

[0153] It should be further explained that the initial automated orchestration logic obtained in this embodiment includes:

[0154] S3061. Based on the automated operation instruction set, a directed acyclic graph node set is created through a node construction operation. The node construction operation generates a corresponding node data object for each instruction in the automated operation instruction set; the node data object includes an instruction identifier attribute, an execution priority numerical attribute, a logical dependency description attribute, and an execution precondition attribute extracted from the corresponding instruction, and includes an initially empty subsequent node reference list attribute. For example, in a technology risk management scenario, the system generates a set of automated operation instructions containing four instructions. Each instruction is assigned a corresponding node data object through node construction operations: The first instruction generates a node data object corresponding to the supplier SLA default warning instruction, with the instruction identifier attribute being CMD-001, the execution priority value set to high (9), and the logical dependency description attribute indicating no pre-dependencies; the second instruction generates a node data object corresponding to the instruction to retrieve database master-slave synchronization logs, with the instruction identifier attribute being CMD-002, the execution priority value set to medium (5), and the logical dependency description attribute indicating no pre-dependencies; the third instruction generates a node data object corresponding to the instruction to start a database cluster health check, with the instruction identifier attribute being CMD-003, the execution priority value set to medium (5), and the logical dependency description attribute indicating a pre-dependency of CMD-002; the fourth instruction generates a node data object corresponding to the instruction to push risk handling notifications to the operations and maintenance manager, with the instruction identifier attribute being CMD-004, the execution priority value set to high (9), and the logical dependency description attribute indicating a pre-dependency of CMD-001.

[0155] S3062. Based on the set of nodes in the directed acyclic graph, a set of directed edges between nodes is established through a dependency relationship construction operation. The dependency relationship construction operation traverses the logical dependency description attribute of each node data object, parses out one or more prerequisite instruction identifiers that the node depends on; searches for the corresponding prerequisite node data object in the set of nodes in the directed acyclic graph according to the prerequisite instruction identifier; establishes a directed edge from the prerequisite node to the current node data object between the found prerequisite node data object and the current node data object, and adds the reference of the current node data object to the subsequent node reference list attribute of the prerequisite node data object. For example, by traversing the data objects of each node, the logical dependency of CMD-003 is described as a predecessor dependency of CMD-002, and the logical dependency of CMD-004 is described as a predecessor dependency of CMD-001. Based on this, the corresponding predecessor node data object is found, and a directed edge is established between the CMD-002 node and the CMD-003 node, pointing from the predecessor node CMD-002 to the current node CMD-003. The reference of CMD-003 is added to the subsequent node reference list attribute of CMD-002. At the same time, a directed edge is established between the CMD-001 node and the CMD-004 node, pointing from the predecessor node CMD-001 to the current node CMD-004. The reference of CMD-004 is added to the subsequent node reference list attribute of CMD-001.

[0156] S3063. Based on the established set of directed edges in the directed acyclic graph node set, the acyclicity of the graph is verified through a cycle detection operation. The cycle detection operation uses a depth-first search algorithm to traverse the set of directed acyclic graph nodes; during the traversal, an access status marker is maintained for each node data object; if the algorithm detects a path that starts from a node data object, passes through several directed edges, and returns to the node data object or any of its ancestor nodes, a cyclic dependency is determined to exist, i.e., a dependency conflict, and a preset conflict handling process is triggered. For example, using a depth-first search algorithm to traverse the four node data objects with established directed edges, starting from CMD-001, mark it as visited, traverse its subsequent node CMD-004, mark CMD-004 as visited, CMD-004 has no subsequent nodes, backtrack; then visit CMD-002, mark it as visited, traverse its subsequent node CMD-003, mark CMD-003 as visited, CMD-003 has no subsequent nodes, backtrack; during the entire traversal, no directed edge path was found between any node and itself or a visited ancestor node, therefore the graph structure is determined to be acyclic and without dependency conflicts.

[0157] S3064. After confirming the graph is acyclic through the loop detection operation, a node execution order list is generated based on the directed edge set using a topological sorting operation. The topological sorting operation calculates a linear order of all nodes in the directed acyclic graph node set according to the sequential constraints defined by the directed edge set. This order satisfies the condition that for any directed edge, its starting node is always earlier than its ending node in the linear order. The generated linear order is the node execution order list. For example, based on the constraints defined by the directed edge set, namely that CMD-001 must be executed before CMD-004 and CMD-002 must be executed before CMD-003, the linear order of nodes calculated by the topological sorting algorithm is CMD-001, CMD-002, CMD-004, CMD-003. In this sequence, all preceding nodes appear before their dependent nodes, satisfying the dependency constraints.

[0158] S3065. Based on the node execution order list and the execution priority numerical attribute of each node data object, the node order is adjusted through a sequence optimization operation to obtain the initial automated orchestration logic. The sequence optimization operation, while strictly maintaining the relative order between nodes with direct or indirect dependencies as determined by the topology sorting operation, reorders adjacent nodes in the node execution order list that have no dependencies in descending order, placing nodes with higher execution priority values ​​at the beginning. The final order list after reordering is the initial automated orchestration logic. For example, given the node order list CMD-001, CMD-002, CMD-004, CMD-003 obtained from topological sorting, we check adjacent node pairs: there is no dependency between CMD-001 and CMD-002, and the priority value of CMD-001 is 9, which is higher than the priority value of CMD-002, so the descending order is maintained and no adjustment is needed; there is no dependency between CMD-002 and CMD-004, and the priority value of CMD-004 is 9, which is higher than the priority value of CMD-002, so we move CMD-004 before CMD-002, resulting in the new sequence CMD-001, CMD-004, CMD-002, CMD-003; we then check that there is a dependency between CMD-002 and CMD-003, so CMD-002 must be kept before CMD-003, and the current order is satisfied. The final optimized initial automated orchestration logic is CMD-001, CMD-004, CMD-002, and CMD-003. This order not only follows all dependency constraints but also executes the high-priority CMD-004 in advance, improving the timeliness of risk handling.

[0159] S307. Based on the initial automated orchestration logic, the current system load, resource utilization rate, and historical execution time data, the system adjusts the instruction execution order and allocates system resources through a multi-objective optimized resource scheduling function, generating a third control instruction that represents the execution logic, execution order, and resource allocation scheme of the automated operation instruction set.

[0160] It should be further explained that this embodiment adjusts the instruction execution order and allocates system resources through a multi-objective optimized resource scheduling function, including:

[0161] S3071. Based on the linear node execution order list included in the initial automated orchestration logic, an initial scheduling scheme chromosome is generated through a genetic algorithm encoding operation. The genetic algorithm encoding operation maps each node data object in the linear node execution order list to a gene position on the initial scheduling scheme chromosome; the instruction identifier of the corresponding node data object is stored at this gene position; simultaneously, a resource quota vector is associated with each gene position, and the resource quota vector stores the number of CPU cores, memory capacity, and network bandwidth allocated to the automated operation instruction represented by the instruction identifier. For example, in a technology risk management scenario, the initial automated orchestration logic includes two instructions, the linear node execution order list is CMD-001 and CMD-002, where CMD-001 is the instruction to generate a supplier service level agreement default warning, and CMD-002 is the instruction to initiate a deep health check of the database cluster. The genetic algorithm encoding operation generates an initial scheduling scheme chromosome, mapping CMD-001 to the first gene position and CMD-002 to the second gene position. At the same time, CMD-001 is associated with a resource quota vector of 2 CPU cores, 4GB memory, and 100Mbps bandwidth, and CMD-002 is associated with a resource quota vector of 1 CPU core, 2GB memory, and 50Mbps bandwidth, forming a chromosome structure [CMD-001:2CPU,4GB,100Mbps, CMD-002:1CPU,2GB,50Mbps].

[0162] S3072. Based on the system's current load index data, resource utilization rate data, historical execution time data, and the resource quota vector of the initial scheduling scheme chromosome, multiple optimization objective values ​​are calculated through a multi-objective fitness evaluation operation, specifically including:

[0163] S3072a. Calculate the first optimization target value, i.e., the estimated overall execution time. Based on the preset baseline execution time T0 of each instruction under no-load conditions recorded in the historical execution time data, and combined with the global load factor F reflected in the current system load index data (range 0 to 1, F=0 indicates no load, F=1 indicates full load), predict the actual execution time Ti=T0×1+α×F of each instruction using a weighted average algorithm, where α is the load influence coefficient, ranging from 0.5 to 2.0, determined by the instruction type, such as α=2.0 for compute-intensive instructions and α=0.5 for I / O-intensive instructions; sum the predicted actual execution times Ti of all instructions to obtain the estimated overall execution time as the first optimization target value. In this embodiment, the preset baseline execution time T0 is derived from statistical analysis of historical execution time data, representing the inherent execution time of an instruction under no load interference. The global load factor F is calculated by real-time monitoring of system resource utilization and is used to quantify the current environmental busyness. The load impact coefficient α is preset according to the instruction type, reflecting the difference in sensitivity of different types of instructions to system load. For example, I / O intensive instructions are less affected by load, so α takes a smaller value of 0.5, while compute-intensive instructions are sensitive to resource contention, so α takes a larger value of 2.0. The actual execution time of each instruction is calculated using the formula Ti = T0 × (1 + α × F), and finally, the total estimated execution time is obtained by summing them up, thus serving as the core target value for measuring the timeliness of the scheme in scheduling optimization.

[0164] For example, assuming the current global load factor F is 0.3, CMD-001 is an I / O intensive instruction with a baseline execution time T0 of 5 seconds and a load impact factor α of 0.5, then the actual execution time T1 = 5 × 1 + 0.5 × 0.3 = 5.75 seconds; CMD-002 is a compute intensive instruction with a baseline execution time T0 of 10 seconds and a load impact factor α of 2.0, then the actual execution time T2 = 10 × 1 + 2.0 × 0.3 = 16 seconds, and the estimated total execution time is 5.75 + 16 = 21.75 seconds.

[0165] S3072b. Calculate the second optimization target value, namely, the resource utilization balance. Based on the amount of various resources allocated to each instruction in the resource quota vector, and the initial utilization rate of various resources in the current resource utilization rate data of the system, simulate the resource utilization changes during the execution of the instruction sequence, and calculate the average planned utilization rate of the three types of resources—CPU, memory, and network bandwidth—over the entire simulation time window; then, calculate the standard deviation of the average planned utilization rate of these three types of resources; use the reciprocal of the standard deviation as a measure of the resource utilization balance to obtain the second optimization target value. For example, during the simulation execution, the average planned utilization rate of CPU is 0.6, memory is 0.5, and network is 0.4, the calculated standard deviation is 0.0816, and its reciprocal is approximately 12.25, which is used as a measure of the resource utilization balance.

[0166] S3072c. Calculate the third optimization target value, namely the resource allocation conflict penalty value. Based on the logical dependencies between instructions, determine the concurrent execution interval of the instructions. Within the concurrent interval, detect whether there are two or more instructions whose resource quota vector requests occupy the same physical resource unit and whose total request exceeds the rated capacity of that resource unit. If such resource contention is detected, perform the following calculations: 1. Calculate the resource over-quota ratio P = total request - rated capacity / rated capacity, P ≥ 0; 2. Query the predefined penalty coefficient table according to the resource type (CPU, memory, network bandwidth) to obtain the corresponding basic penalty coefficient K; where the penalty coefficient table is defined as follows: CPU resource K = 10, memory resource K = 8, network bandwidth resource K = 5; 3. Conflict penalty value = K × P, if P = 0, the penalty value is 0; if no resource contention is detected, the conflict penalty value is zero. For example, since CMD-001 and CMD-002 have a dependency relationship, CMD-002 must wait for CMD-001 to complete before it can execute. Since there is no concurrent execution interval between them, no resource contention is detected, and the conflict penalty value is 0.

[0167] S3073. Based on the initial scheduling scheme chromosome and the multi-objective fitness evaluation operation, the Pareto optimal solution set is obtained by iterative optimization using a non-dominated sorting genetic algorithm with an elite retention strategy. The non-dominated sorting genetic algorithm with an elite retention strategy executes the following sub-steps: First, an initial population consisting of multiple randomly generated scheduling scheme chromosomes is generated. Then, the multi-objective fitness evaluation operation is performed on each chromosome in the current population to obtain its corresponding multiple optimization objective values. Next, based on the multiple optimization objective values ​​of all chromosomes, a non-dominated sorting operation and a crowding distance calculation operation are performed to stratify and sort the individuals in the population. Subsequently, based on the results of the non-dominated sorting and crowding distance, a selection operation is performed to select parent individuals. The selected parent individuals are sequentially subjected to a simulated binary crossover operation and a polynomial mutation operation to generate a offspring population. The elite individuals in the parent population are merged with the offspring population to form a new generation population. The above evaluation, sorting, selection, crossover, mutation, and merging steps are repeated until the preset iteration termination condition is met. Finally, all non-dominated chromosomes are obtained from the last generation population to form the Pareto optimal solution set. For example, the initial population contains 50 randomly generated scheduling scheme chromosomes. After 100 generations of iterative optimization, three non-dominant chromosomes are selected from the final population to form a Pareto optimal solution set. Chromosome A corresponds to an estimated total execution time of 20.5 seconds, a resource utilization balance of 12.5, and a conflict penalty of 0. Chromosome B corresponds to 21.0 seconds, 13.0 seconds, and 0 seconds, and chromosome C corresponds to 22.0 seconds, 14.0 seconds, and 0 seconds.

[0168] S3074. Based on the Pareto optimal solution set, a final scheduling scheme chromosome is selected through a preset decision strategy. The preset decision strategy is used to determine a suitable chromosome from multiple candidate chromosomes in the Pareto optimal solution set, according to a preset optimization objective preference, as the final scheduling scheme chromosome. For example, the preset decision strategy prioritizes resource utilization balance, selecting the scheme with the shorter estimated overall execution time when the balance is similar. Based on this preference, chromosome B is selected from the Pareto optimal solution set as the final scheduling scheme chromosome.

[0169] S3075. Based on the final scheduling scheme chromosome, the third control instruction is generated through a genetic algorithm decoding operation. The genetic algorithm decoding operation parses the gene sequence of the final scheduling scheme chromosome, obtains the instruction identifiers stored at the gene locations and their arrangement order, thereby determining the final instruction execution order; simultaneously, it reads the resource quota vector associated with each gene location to obtain the number of CPU cores, memory capacity, and network bandwidth allocated to each instruction; finally, based on the instruction execution order and resource allocation parameters, a structured instruction sequence conforming to the automated operation and maintenance engine control interface specification is constructed, and this structured instruction sequence is the third control instruction. For example, after the chromosome decoding of the final scheduling scheme, the execution order of the instructions is CMD-001 first and CMD-002 second. CMD-001 is allocated 2 CPU cores, 4GB memory, and 100Mbps bandwidth, while CMD-002 is allocated 1 CPU core, 2GB memory, and 50Mbps bandwidth. The generated third control instruction is to first execute the instruction to generate a supplier service level agreement default warning and allocate the corresponding resources. After its completion, the instruction to start a deep health check of the database cluster is executed and the corresponding resources are allocated. This instruction sequence can be directly sent to the automated operation and maintenance engine for execution.

[0170] It should be further explained that this embodiment performs constraint verification and completion under the constraints of preset work order integrity rules and information accuracy verification rules, including:

[0171] S401. Based on the target work order template identifier, a pre-defined work order integrity rule database is queried by executing a structured query language statement to obtain the list of required field definitions and the list of field format rules associated with the target work order template identifier. For example, in a technology risk management scenario, the target work order template identifier is a supplier compliance review work order. By executing an SQL query, the list of required field definitions corresponding to this template is obtained from the work order integrity rule database. This list includes the supplier name, contract number, risk description, and non-compliance items. The list of field format rules includes the data type and validation rules for each field.

[0172] S4011. Based on the target work order template identifier, establish a network connection with the preset work order integrity rule database through a database connection driver. The database connection driver is a standard relational database connection component used to manage network sessions, authentication, and data transmission. For example, a network connection to a MySQL database is established through a JDBC driver, and authentication is completed by passing in the database URL, username, and password, successfully establishing a session channel.

[0173] S4012. Based on the established network connection, generate a parameterized query statement using the Structured Query Language (SCL) statement builder. This statement contains a query condition that specifies the target work order template identifier as the matching key, used to retrieve relevant records in the rule definition table of the work order integrity rule database. For example, the statement builder can be used to generate the parameterized query statement SELECT required_fields, format_rules FROM template_rules WHERE template_id = ?, and the supplier compliance review work order can be passed in as a parameter.

[0174] S4013. Based on the generated parameterized query statement, a query request is sent to the work order integrity rule database through the database connection executor, and the query result set returned by the database is received. For example, the above query statement is executed through an established JDBC connection. The result set returned by the database contains one record, which contains the required_fields field valued as supplier name, contract number, risk description, and non-compliance item, and the format_rules field valued as a rule definition in JSON string format.

[0175] S4014. Based on the received query result set, read the results line by line using the data row parser. Each row of result data contains at least two fields: one field to store the definition information of required fields, and the other field to store the definition information of field format rules. For example, use the result set parser to read the only row of data returned, and extract the string value of the required_fields field and the JSON string value of the format_rules field from that row.

[0176] S4015. Based on each row of parsed result data, the required field definition information items are added to a list using a data structure builder to form the required field definition list; simultaneously, the corresponding field format rule definition information items are added to another list to form the field format rule list. For example, the parsed supplier name, contract number, risk description, and non-compliance items are separated by commas to construct the required field definition list, and the parsed JSON string is deserialized into a structured field format rule list, which includes the supplier name as a string type and the contract number needing to meet the format constraints of the regular expression ^HT-[0-9]{4}-[0-9]{4}$.

[0177] S402. Based on the list of required fields and the set of field names contained in the second structured data set, the missing required field names are obtained by calculating the set difference, forming a list of missing fields.

[0178] It should be further explained that this embodiment obtains the names of missing required fields through set difference operations. The specific implementation steps and logic are described below with examples:

[0179] S4021. Based on the second structured data set, obtain a set of field names through a key set extraction operation. The key set extraction operation is used to traverse the key-value pair data structure of the second structured data set, collect all the keys, i.e., all the filled field names, to form a set of field names. For example, in a supplier compliance review scenario, the second structured data set may contain key-value pairs {"vendor": "supplier B", "risk_desc": "failed security audit"}, then the set of field names obtained through this operation is ["vendor", "risk_desc"].

[0180] S4022. Based on the list of required field definitions and the set of field names, a comparison is performed using a set difference operation. This operation matches each element in the list of required field definitions (i.e., each required field name) with an element in the set of field names. The core logic is: for each required field name in the list of required field definitions, check if the name exists in the set of field names. If it does not exist, the required field name is determined to be a missing item.

[0181] S4023. Based on the judgment result of the set difference calculation operation, a missing field list is generated through a result collection operation. This operation adds the names of all required fields that are determined to be missing items to an initially empty list data structure. This list is the missing field list. Continuing the previous example, if the list of required field definitions is ["vendor", "contract_id", "risk_desc"], and the set of field names is ["vendor", "risk_desc"], then by comparison, it is found that "contract_id" does not exist in the set of field names, so the generated missing field list is ["contract_id"].

[0182] S403. Based on the field format rule list, perform pattern matching verification on the padding values ​​corresponding to each field in the second structured data set using a regular expression engine to identify fields whose formats do not conform to the field format rules, and form a list of fields with abnormal formats.

[0183] S404. Based on the entity identifiers contained in the unique key entity information set, obtain the authoritative configuration attribute information corresponding to each entity identifier by calling the Hypertext Transfer Protocol application programming interface provided by the configuration management database.

[0184] S405. Based on the field fill value V1 associated with the entity in the second structured data set and the corresponding authoritative configuration attribute information V2 obtained from the configuration management database, the accuracy quantification score of each field fill value is obtained by calculating the Levenstein distance and converting it into a similarity score; wherein, the formula for converting the Levenstein distance into a similarity score is: S = 1 - (LD(V1,V2) / max(len(V1),len(V2))), S is the accuracy quantification score, with a value range of [0,1], LD is the Levenstein distance function, and len is the string length function; the closer S is to 1, the higher the accuracy;

[0185] S406. Based on a preset accuracy verification threshold, by comparing the accuracy quantification score with the accuracy verification threshold, identify fields whose accuracy quantification score is lower than the accuracy verification threshold, and form a list of fields with abnormal accuracy; wherein, the preset accuracy verification threshold is dynamically configured according to the field type: the threshold for key fields (such as contract number, server ID) is 0.95, and the threshold for ordinary fields (such as description information) is 0.8.

[0186] S407. Based on the missing field list, access the operation and maintenance knowledge base or historical work order database for information retrieval by executing a graph query language query statement or a structured query language query statement. If the retrieval is successful, generate the corresponding fill value and update the second structured data set; wherein, the query logic is as follows:

[0187] 1) For each missing field F, extract its associated entity identifier ID;

[0188] 2) If accessing the operations and maintenance knowledge base, execute the graph query language statement (such as Cypher): MATCH (n:Entity{id:ID}) RETURN nF AS value;

[0189] 3) If accessing the historical work order database, execute the following structured query language statement: SELECT F FROM work_order WHERE entity_id=ID ORDER BY create_time DESC LIMIT 1;

[0190] 4) Successful retrieval criteria: If the returned value is not empty and conforms to the field format rules, then the value will be filled into the missing field and the second structured data set will be updated;

[0191] 5) Degradation strategy for failed retrieval: If no valid data is found, the missing field is marked as "to be manually completed" and a manual completion reminder process is triggered.

[0192] S408. Based on the format exception field list and the accuracy exception field list, create a task to be reviewed containing exception details and context through the workflow engine, and submit the task to be reviewed to the manual review queue, while pausing the automatic work order generation process based on the current second structured data set.

[0193] S409. After completing the information completion and potential manual review and correction, the second structured data set after verification and completion is generated by performing the final integrity check and data merging, which satisfies all preset rules.

[0194] It should be further explained that this embodiment performs an output operation based on the verified and completed second structured data set and the third control instruction, including:

[0195] S5.1. Based on the second structured data set after verification and completion, create a work order by calling the operation-driven work order management system through the application programming interface; specifically including:

[0196] S5.1a. Based on the second structured data set after verification and completion, and the predefined work order management system application programming interface specification, a standard format request payload is generated through a serialization function.

[0197] S5.1b: Based on the generated request payload, a request is initiated to the creation interface of the work order management system through a Hypertext Transfer Protocol client.

[0198] S5.1c. Based on the processing result of the request by the work order management system, the work order creation status is determined by the response parsing function. If a success status code and a unique work order identifier are parsed, the work order instantiation and persistence are determined to be successful; if an error status code is parsed, the creation is determined to have failed and an exception handling process is triggered.

[0199] S5.2. Based on the third control command, the automated operation and maintenance engine is driven to perform operations through message delivery. Specifically, this includes:

[0200] S5.2a. Based on the third control instruction, encapsulate it into a message body that conforms to the control interface specification of the automated operation and maintenance engine through a formatting function.

[0201] S5.2b. Based on the encapsulated message body, the message is delivered to the control loop entry point of the automated operation and maintenance engine through a message queue or remote procedure call client.

[0202] S5.2c. Based on the automated operation and maintenance engine's reception of messages and subsequent execution feedback, the task execution status is monitored through a status listener.

[0203] For example, in a technology risk management scenario, when a conversation appears in a chat group stating, "Contract CON-2024-012 stipulates that supplier H's field engineer must pass a safe operation certification, but this inspection found that he does not have this qualification, and relevant equipment logs are missing," the system, after the aforementioned process, generates a verified and completed second structured data set (containing fields and values ​​such as supplier, contract, and risk items) and a pre-arranged third control instruction (containing two instructions, "Generate Supplier Default Alarm" and "Initiate Safe Equipment Log Review," along with their resource allocation). When executing the output operation, the system completes two tasks in parallel: First, it serializes the second structured data set into JSON format and sends it to the RESTful API interface of the work order management system via an HTTP POST request. This successfully drives the latter to create a "Supplier Compliance Risk Work Order" with a status of "New" in the database and returns the work order ID. Meanwhile, the system encapsulates the third control command into a specific message format and sends it to the automated operation and maintenance engine through the message middleware. The engine then prioritizes scheduling resources to execute high-priority alarm generation tasks according to the command logic, and then triggers log review tasks. The status of the entire process is fed back to the chat group in real time through the callback interface, thus realizing a complete closed loop from risk dialogue reporting to structured recording and automated handling response.

[0204] This embodiment addresses the technical pain points in real-time IT operations collaboration based on ChatOps, where existing methods struggle to resolve the issues of referential resolution and cross-message information fusion in multi-turn unstructured dialogues, leading to inaccurate and incomplete work order generation information. It constructs a semantic parsing model comprising an embedding layer, encoding layer, dual decoding branches, attention mechanism, and coreference resolution layer. Through core technologies such as domain-adaptive fine-tuning of word vector mapping, contextual semantic encoding using bidirectional long short-term memory networks, historical dialogue semantic splicing using dynamic adaptive context windows, dependency weight calculation using attention mechanisms, and entity referential merging using cosine similarity and agglomerative hierarchical clustering, a complete technical solution is formed, encompassing dialogue content acquisition, semantic parsing, task allocation, work order generation, and automated operation orchestration, achieving significant technical effects and application value. Specifically, in the referential resolution stage, this application extracts bidirectional fusion hidden state vectors of entity referentials based on enhanced semantic representation, generates referential semantic vectors through average pooling, quantifies the entity referential relevance through cosine similarity calculation, and then uses agglomerative hierarchical clustering... Hierarchical clustering algorithms group highly related references into the same entity and assign globally unique identifiers, effectively solving the ambiguity problem of pronouns such as "it" and "the server". For example, when processing cross-turn dialogues such as "the database instance of server A experienced a connection timeout" and "restart it and check performance metrics", "it" can be accurately associated with the database instance in the historical dialogue, avoiding incorrect entity identification. In the cross-message information fusion stage, this application uses a dynamically adaptive context window to adjust the window length based on the number of key entities in the historical rounds and the relevance of the intent in the current round. After concatenating the semantic vectors of the current and historical rounds, the dependency weight of the current semantic on the historical semantic is calculated with the help of an attention mechanism, realizing deep fusion of cross-turn information. For example, when processing the dialogues such as "Contract SLA-2024-015 requires supplier G's server to pass the monthly security audit" and "This audit report shows that neither of its two target servers met the standards", the scattered "supplier G" and "its two target servers" can be associated as a unified entity, avoiding the fragmentation and omission of key information.Building upon this foundation, this application employs a decision tree classifier to match target work order templates, field filling with uniqueness verification to generate a second structured data set, a multi-level condition matching rule engine to generate automated operation instructions, and a third control instruction involving the construction of a directed acyclic graph with dependency conflict detection and multi-objective optimization of resource scheduling orchestration. After work order integrity and information accuracy verification and completion, it ultimately outputs high-fidelity structured work orders and precise automated operation and maintenance instructions. For example, in supplier compliance risk scenarios, it successfully generates compliance review work orders containing all required fields such as supplier name, contract number, risk description, and non-compliance items. Simultaneously, it prioritizes triggering high-priority supplier SLA default warning instructions, followed by executing a database cluster health deep check instruction. This completely resolves the technical deficiencies of existing methods, significantly improves the real-time performance, accuracy, and automation level of ChatOps operation and maintenance collaboration, reduces manual intervention costs, and achieves closed-loop management of the entire process from operation and maintenance dialogue to work order generation and automated processing.

[0205] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

[0206] If the technical solution disclosed herein involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. A ChatOps-based method for real-time IT operations collaboration and automatic work order generation, characterized in that: include: Based on continuous monitoring of real-time conversations on chat platforms, real-time conversation content containing unstructured text is collected. The real-time dialogue content is understood and represented by the configured semantic parsing model, and a first structured data set representing the core processing intention and key entity information of the dialogue is calculated and generated. Based on the first structured data set, tasks are allocated between the preset work order information fields and the automated operation and maintenance operations to be triggered, resulting in a second structured data set for filling into the standardized work order template, and a set of automated operation instructions for driving the automated operation and maintenance engine to execute. Under the constraints of preset work order integrity rules and information accuracy verification rules, the second structured data set is constrained, verified, and completed to obtain the verified and completed second structured data set. Based on the first structured data set and the automated operation instruction set, a third control instruction is generated. The third control instruction is used to characterize the arrangement of the execution logic and order of the automated operation instruction set. The second structured dataset after verification and completion is output as a work order generation instruction to the creation interface of the work order management system, and the third control instruction is output as an automated orchestration instruction to the control loop of the automated operation and maintenance engine.

2. The method for real-time IT operations collaboration and automatic work order generation based on ChatOps as described in claim 1, characterized in that, The matching degree between the key entity information in the second structured data set after verification and completion and the corresponding key entity information in the first structured data set is not lower than the first preset threshold. The degree of consistency between the set of automated operation instructions arranged by the third control instruction and the core processing intent of the dialogue represented in the first structured data set is not less than the second preset threshold.

3. The method for real-time IT operations collaboration and automatic work order generation based on ChatOps as described in claim 2, characterized in that, The first structured data set representing the core intent of the dialogue and key entity information is calculated and generated, including: The real-time dialogue content is normalized to obtain a normalized dialogue text sequence; Based on the normalized dialogue text sequence, each word is mapped to a dense vector through the word vector model in the embedding layer to obtain the word vector sequence; Based on the word vector sequence, forward and backward sequence modeling is performed through a bidirectional long short-term memory network in the encoding layer to obtain a semantic vector representation of the dialogue that incorporates contextual information; Based on the semantic vector representation of the dialogue, a nonlinear transformation is performed through the fully connected layer in the first decoding branch, and then the probability distribution is calculated through the Softmax function to obtain the intent classification label and corresponding confidence level that characterize the core processing intent of the dialogue.

4. The method for real-time IT operations collaboration and automatic work order generation based on ChatOps as described in claim 3, characterized in that, The calculation generates a first structured data set representing the core intent of the dialogue and key entity information, which also includes: Based on the semantic vector representation of the dialogue, sequence labeling is performed through the conditional random field model in the second decoding branch. Each word in the normalized dialogue text sequence is classified to obtain a preliminary key entity information set. Each element in the preliminary key entity information set includes an entity type label and the start and end position index of the entity in the normalized dialogue text sequence. The normalized dialogue text sequence is context-encoded to obtain the semantic vector representation of the current round of dialogue and the semantic vector representation of each historical round of dialogue. Based on a preset context window, the semantic vector representation of the current round of dialogue and the semantic vector representation of the historical round of dialogue are concatenated. The dependency weight of the current semantic vector on the historical semantic vector is calculated through an attention mechanism. The historical semantic vectors are weighted and summed according to the dependency weight. The weighted sum is then fused with the current semantic vector to obtain an enhanced semantic representation that incorporates the historical dialogue context.

5. The ChatOps-based real-time IT operations collaboration and automatic work order generation method as described in claim 4, characterized in that, The calculation generates a first structured data set representing the core intent of the dialogue and key entity information, which also includes: Based on the enhanced semantic representation, the vector cosine similarity between different entity references in the preliminary key entity information set is calculated through the coreference resolution layer. Multiple entity references with cosine similarity exceeding a preset threshold are merged into the same entity identifier to obtain a unique key entity information set that eliminates referential ambiguity and has a unified identifier. Based on the intent classification label and the unique key entity information set, the data is assembled and serialized according to the preset key value through the information fusion layer to obtain the first structured data set representing the core processing intent and key entity information of the dialogue.

6. The method for real-time IT operations collaboration and automatic work order generation based on ChatOps as described in claim 5, characterized in that, The process of allocating tasks between preset work order information fields and automated operation and maintenance operations to be triggered includes: Based on the intent classification labels and corresponding confidence levels in the first structured dataset, a decision tree classifier with confidence threshold verification is used to match predefined work order template selection rules, filter paths whose confidence levels do not meet the threshold, and obtain the target work order template identifier. Based on the target work order template identifier, the template-field association database is queried by combining a database query function that sorts fields by priority, and the set of required field names, the set of optional field names, and the field priority sorting results are obtained. Based on the unique key entity information set and entity identifier in the first structured data set, the mapping is completed and the entity identifier is associated by using the entity type-field mapping function that supports dynamic updates, and the mapping is completed by using the iteratively updated entity type-work order field correspondence lookup table to obtain the preliminary field filling mapping relationship.

7. The method for real-time IT operations collaboration and automatic work order generation based on ChatOps as described in claim 6, characterized in that, The process of allocating tasks between preset work order information fields and automated operation and maintenance operations to be triggered also includes: Based on the preliminary field filling mapping relationship, the unique key entity information set, and the field priority sorting results, the entity text content is extracted according to priority by the field value filling function with uniqueness verification, and the field filling and deduplication are completed to generate the second structured data set. Based on the intent classification labels, unique key entity information set and entity attribute information in the first structured data set, the rule engine loads the hierarchical condition action rule set through the automated operation triggering of multi-level condition matching, and generates an automated operation instruction set with execution priority identifier. Based on the execution priority, logical dependencies, and execution prerequisites of the automated operation instruction set, a directed acyclic graph with dependency conflict detection is used to construct and optimize the task execution flowchart to obtain the initial automated orchestration logic. Based on the initial automated orchestration logic, the current system load, resource utilization rate, and historical execution time data, the execution order of instructions is adjusted and system resources are allocated through a multi-objective optimized resource scheduling function, generating a third control instruction that represents the execution logic, execution order, and resource allocation scheme of the automated operation instruction set.

8. The method for real-time IT operations collaboration and automatic work order generation based on ChatOps as described in claim 6, characterized in that, The step of performing constraint verification and completion on the second structured data set under the constraints of preset work order integrity rules and information accuracy verification rules includes: Based on the target work order template identifier, a pre-defined work order integrity rule database is queried by executing a structured query language statement to obtain a list of required field definitions and a list of field format rules associated with the target work order template identifier. Based on the list of required fields defined and the set of field names contained in the second structured data set, the missing required field names are obtained by calculating the set difference, thus forming a list of missing fields; Based on the field format rule list, a regular expression engine is used to perform pattern matching verification on the padding values ​​corresponding to each field in the second structured data set, identifying fields whose formats do not conform to the field format rules, and forming a list of fields with abnormal formats.

9. The method for real-time IT operations collaboration and automatic work order generation based on ChatOps as described in claim 8, characterized in that, The step of performing constraint verification and completion on the second structured data set under the constraints of preset work order integrity rules and information accuracy verification rules also includes: Based on the entity identifiers contained in the unique key entity information set, authoritative configuration attribute information corresponding to each entity identifier is obtained by calling the Hypertext Transfer Protocol application programming interface provided by the configuration management database. Based on the field fill values ​​associated with entities in the second structured data set and the corresponding authoritative configuration attribute information obtained from the configuration management database, the accuracy quantification score of each field fill value is obtained by calculating the Levenstein distance and converting it into a similarity score. Based on a preset accuracy verification threshold, by comparing the accuracy quantification score with the accuracy verification threshold, fields with accuracy quantification scores lower than the accuracy verification threshold are identified, forming a list of fields with abnormal accuracy.

10. The method for real-time IT operations collaboration and automatic work order generation based on ChatOps as described in claim 9, characterized in that, The step of performing constraint verification and completion on the second structured data set under the constraints of preset work order integrity rules and information accuracy verification rules also includes: Based on the missing field list, information retrieval is performed by executing a graph query language query statement or a structured query language query statement to access the operation and maintenance knowledge base or historical work order database. If the retrieval is successful, the corresponding fill value is generated and the second structured data set is updated. Based on the list of format exception fields and the list of accuracy exception fields, a task to be reviewed containing exception details and context is created through the workflow engine, and the task to be reviewed is submitted to the manual review queue. At the same time, the automatic work order generation process based on the current second structured data set is paused. After completing information completion and potential manual review and correction, a second structured data set that satisfies all preset rules is generated by performing a final integrity check and data merging.

Images