A method for identifying service order intent based on a large language model

Abstract

The application provides a method for identifying the intention of a service order based on a large language model, which comprises the following steps: obtaining the text of a service order, calling a large language model in parallel by using differentiated initial instructions, generating candidate thinking chains from the dimensions of customer expression, technician investigation and accessory demand, and forming a structured reasoning result from entity elements, fault phenomena to service items; aligning and verifying the entity elements in the candidate thinking chains with a knowledge graph in the field of service, and obtaining a knowledge verification confidence weight; calculating the semantic similarity of key nodes of each thinking chain, constructing a consistency relationship matrix and calculating a thinking chain divergence index; when the index is lower than a first threshold, output a single intention according to the confidence weight; when the index is greater than or equal to the first threshold and lower than a second threshold, positioning the conflict nodes and issuing a discrimination instruction, and outputting a composite intention; when the index is greater than or equal to the second threshold, a low random reconstruction verification mechanism is used for re-reasoning until the output condition is met or the artificial review process is entered.

Description

Technical Field

[0001] This application belongs to the field of large language models, and in particular relates to a method for recognizing the intent of automotive service work orders based on large language models. Background Technology

[0002] Large language models, with their strong capabilities in natural language understanding, semantic extraction, and logical reasoning, have been increasingly applied to the field of text intent recognition. Using large language models for automotive service work order processing can, to some extent, reduce manual review costs and improve the efficiency of work order information extraction and preliminary classification. However, automotive service work order texts typically possess strong professional, complex, and unstructured characteristics. They contain not only vertical domain terms such as vehicle parts, fault symptoms, and repair items, but may also simultaneously mix in multi-source information such as customer subjective descriptions, technician inspection records, fault codes, and parts and consumables details. Conventional intent recognition methods often struggle to fully capture the deep semantic relationships within the context when processing such texts, especially when customer descriptions are ambiguous, diagnostic information is incomplete, or multiple repair items are intertwined, leading to inaccurate recognition results and unstable service item classification.

[0003] Existing automotive service work order intent recognition methods based on large language models typically rely on single prompt words and linear reasoning paths, lacking mechanisms for comprehensive analysis and cross-validation from multiple perspectives, including customer statements, technician inspections, and parts requests. When discrepancies exist between the customer's subjective statements and the technician's objective inspections, the model is prone to misunderstandings and may even generate inference results lacking domain factual basis. Furthermore, existing thought chain reasoning processes are largely not fully integrated with automotive service domain knowledge bases, resulting in a lack of effective verification of generated entity elements, fault relationships, and service items, making it difficult to quantify the reliability and confidence of the reasoning results. Faced with potential intent divergences between different thought chains, existing technologies also lack mechanisms for assessing the degree of divergence, identifying conflict nodes, and adaptive correction, making it difficult to select weighted outputs, composite adjudication, or re-reasoning strategies based on the severity of the divergence. Therefore, there is an urgent need for an automotive service work order intent recognition method that integrates multi-dimensional perspectives, domain knowledge verification, and thought chain divergence handling mechanisms to improve the accuracy, stability, and interpretability of the recognition process. Summary of the Invention

[0004] To address the shortcomings of existing technologies in automotive service work order intent recognition, such as reliance on a single reasoning path, lack of domain knowledge verification, and difficulty in quantifying divergences in the thought chain, leading to insufficient accuracy and stability, this invention provides, in a first aspect, a method for automotive service work order intent recognition based on a large language model. The method includes:

[0005] Acquire the text data of the automotive service work order to be identified; use the initial non-homogeneous perspective instructions to drive the large language model in parallel, and generate multiple candidate thought chains from the dimensions of customer expression, technician inspection and parts demand. The candidate thought chains contain a structured reasoning process from entity elements, fault phenomena to service items. The structured entity elements of the candidate thought chains are aligned and verified with a predefined automotive service domain knowledge graph. The knowledge verification confidence is calculated and used as a weighting factor. The semantic similarity of key reasoning nodes between each candidate thought chain is calculated, and a relationship matrix is ​​constructed based on the semantic similarity. The thought chain divergence index is calculated based on the relationship matrix, and an adaptive mechanism is activated according to the thought chain divergence index. If the thought chain divergence index is lower than the first threshold, it is weighted in combination with the knowledge verification confidence level, and the single intent with the highest confidence level is output. If the thought chain divergence index is greater than or equal to the first threshold but lower than the second threshold, the conflict node is located, a conflict analysis instruction is generated to drive the large language model to make a decision, and a composite intent is output. If the thought chain divergence index is greater than or equal to the second threshold, the generation temperature of the large language model is adjusted to a preset low random range, a clarifying reconstruction instruction is generated based on the root cause of the divergence to optimize the initial instruction set, and a new round of reasoning verification loop is started until the thought chain divergence index is lower than the second threshold, or when the preset maximum number of retries is reached, a manual review mark is output.

[0006] On the other hand, the present invention also provides an automotive service work order intent recognition system based on a large language model, the system comprising the following modules: The generation module is used to acquire the automotive service work order text data to be identified; it uses the initial non-homogeneous perspective instructions to drive the large language model in parallel, and generates multiple candidate thought chains from the dimensions of customer expression, technician inspection and parts demand. The candidate thought chains contain a structured reasoning process from entity elements, fault phenomena to service items. The construction module is used to perform entity alignment verification operations between the structured entity elements of the candidate thinking chains and the predefined automotive service domain knowledge graph, calculate the knowledge verification confidence and use the knowledge verification confidence as a weight factor; calculate the semantic similarity of key reasoning nodes between each candidate thinking chain, and construct a relationship matrix based on the semantic similarity. The output module is used to calculate the thought chain divergence index based on the relation matrix, and activate an adaptive mechanism according to the thought chain divergence index. If the thought chain divergence index is lower than a first threshold, it is weighted in combination with the knowledge verification confidence level, and the single intent with the highest confidence level is output. If the thought chain divergence index is greater than or equal to the first threshold and lower than a second threshold, the conflict node is located, a conflict analysis instruction is generated to drive the large language model to make a decision, and a composite intent is output. If the thought chain divergence index is greater than or equal to the second threshold, the generation temperature of the large language model is adjusted to a preset low random range, a clarifying reconstruction instruction is generated based on the root cause of the divergence to optimize the initial instruction set, and a new round of reasoning verification loop is started until the thought chain divergence index is lower than the second threshold, or when the preset maximum number of retries is reached, a manual review mark is output.

[0007] This invention utilizes a parallel-driven large language model with non-homogeneous perspective instructions to generate candidate thought chains from dimensions such as customer statements, technician inspections, and parts requests, thereby fully extracting multi-source semantic features from automotive service work orders. By aligning and validating the entity elements in the candidate thought chains with the automotive service domain knowledge graph, a knowledge verification confidence score is obtained and used as the weighting basis for subsequent outputs, helping to reduce the risk of the model's inference results lacking domain factual support. Simultaneously, by constructing a consistency relation matrix and calculating the thought chain divergence index, the consistency between different inference paths is quantified, and weighted output, conflict resolution, or low-random reconstruction verification processing is performed according to the degree of divergence, improving the accuracy, reliability, and stability of automotive service work order intent recognition. Attached Figure Description

[0008] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram of the semantic similarity matrix of candidate thought chains. Detailed Implementation

[0009] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0010] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0011] In the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0012] References to "one embodiment" or "some embodiments" as described in this application specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in another embodiment," "in yet another embodiment," etc., appearing in different parts of this application specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0013] In the first embodiment, the present invention proposes a method for recognizing the intent of automotive service work orders based on a large language model, such as... Figure 1 As shown, it includes: S1. Obtain the text data of the automotive service work order to be identified; use the initial non-homogeneous perspective instructions to drive the large language model in parallel, and generate multiple candidate thought chains from the dimensions of customer expression, technician inspection and parts demand. The candidate thought chains contain a structured reasoning process from entity elements, fault phenomena to service items.

[0014] Acquire the text data of the automotive service work orders to be identified and store it in a relational database. Construct an initial non-homogeneous perspective instruction set, which includes prompts for customer description, technician inspection, and parts demand dimensions. Specifically, the customer description dimension corresponds to customer anomaly statements and customer experience instruction templates; the technician inspection dimension corresponds to diagnostic equipment fault codes, technician inspection records, or vehicle structure query information; and the parts demand dimension corresponds to parts requirements and consumable details. This establishes a correspondence between the three analytical dimensions and specific input sources.

[0015] Utilizing a parallel task scheduling mechanism, asynchronous inference requests are initiated for different dimensions of prompts, and a large language model deployed on a graphics processor is invoked to analyze the automotive service work order text data. The large language model generates three candidate thought chains based on three dimensions: customer description, technician inspection, and parts requirements.

[0016] The large language model generates text conforming to a preset structured format. This structured text is parsed to obtain sequentially associated entity element nodes, fault phenomenon nodes, and service item nodes. These nodes are then connected to form a directed acyclic reasoning graph structure. Each candidate thought chain also includes an intent tag field corresponding to a service item node.

[0017] In an optional embodiment, the method of using initial non-homogeneous perspective instructions to drive a large language model in parallel generates multiple candidate thought chains from the dimensions of customer expression, technician inspection, and parts requirements, including: Fill the customer's abnormal description in the auto service work order text data into the customer experience instruction template, and extract the target part and the fault phenomenon as the first reasoning node in the candidate thinking chain of customer description dimension. Extract fault codes generated by diagnostic equipment, combine them with vehicle structure query instructions, and generate in-depth component repair records as the second reasoning node in the technician inspection dimension candidate thinking chain; The parts requirements and consumable details of the automotive service work order text data are analyzed, and the related operation items are deduced by the use of parts to generate the third reasoning node in the candidate thinking chain of parts requirements. The first inference node, the second inference node, and the third inference node are respectively combined with the entity elements, fault phenomena, and service item nodes extracted under their respective dimensions to complete the candidate thought chain for the corresponding dimension.

[0018] The large language model employs a deep neural network containing multiple Transformer decoder blocks. Each decoder block integrates a masked multi-head self-attention mechanism and a position-fed forward neural network. The network input is a discrete serialized text embedding vector processed by a tokenizer and appended with positional encoding. The output is a probability distribution matrix predicting the next word. The calculation formula for multi-head self-attention feature aggregation is as follows: Where Q represents the query tensor, K represents the key tensor, and V represents the value tensor. This represents the feature scaling constant.

[0019] The character limit for the work order text input is typically between 50 and 500 characters. The extracted customer anomaly description is then filled into a pre-defined customer experience instruction template, such as: "Please extract the faulty component and its surface manifestation from the text {text}". A named entity recognition algorithm is used to extract the target component, such as brake pads, and the fault symptom, such as abnormal vibration or unusual noise. This is then encapsulated in JSON format as the first inference node. Based on this first inference node, a candidate thought chain for the customer description dimension is further generated, for example: brake pads -> abnormal vibration or unusual noise -> brake system inspection service, and this service item is mapped to the corresponding intent tag.

[0020] If the automotive service work order text data contains a standard OBD-II fault code generated by diagnostic equipment, then the vehicle structure query information or the underlying fault database is invoked to map the fault code to a deep component repair record expressed in natural language. This deep component repair record is then used as the second inference node in the technician inspection dimension. If the automotive service work order text data does not contain a diagnostic equipment fault code, then a technician inspection node is generated based on the technician's text inspection record or the vehicle structure query information. This technician inspection node is then used as the second inference node in the subsequent candidate thought chain generation.

[0021] The consumables list is parsed using regular expressions, such as "4L of 5W-30 fully synthetic engine oil" and "1 oil filter". Based on the mapping relationship between the Bill of Materials (BOM) and the standard time library, related standard operation items are deduced in reverse, such as "routine minor maintenance" or "oil and oil filter change", thus generating a third inference node. Based on this third inference node, a candidate thought chain for parts demand dimensions is further generated, for example: oil filter -> maintenance consumables demand -> oil and oil filter change service, and this service item is mapped to the corresponding intent tag.

[0022] Following the chronological order of "fault phenomenon -> diagnosis and attribution -> repair solution" and constrained by a causal logic matrix, a graph-structured connection or a directed acyclic graph (DAG) is used to connect nodes within the candidate thought chains corresponding to each dimension, generating structured candidate thought chains, such as: brake noise -> ABS sensor fault code C0020 -> replace the left front wheel ABS sensor. The maximum length of each generated thought chain is preferably set to 3 to 5 logical levels to improve the compactness and business interpretability of the large language model's reasoning process.

[0023] S2, perform entity alignment verification operation between the structured entity elements of the candidate thinking chain and the predefined automotive service domain knowledge graph, calculate the knowledge verification confidence and use the knowledge verification confidence as a weight factor; calculate the semantic similarity of key reasoning nodes between each candidate thinking chain, and construct a relationship matrix based on the semantic similarity.

[0024] A standard entity library from a predefined automotive service domain knowledge graph is retrieved. Entity elements and action feature words are extracted for each candidate thought chain, and the extraction results are matched with standard entity nodes in the graph. A first parameter is calculated based on the ratio of the number of valid matching nodes to the total number of feature words. A second parameter is calculated based on the proportion of matching node connections that conform to the graph's preset constraint direction. The first and second parameters are then weighted and summed according to preset weight coefficients to obtain the knowledge verification confidence score of the candidate thought chain. This knowledge verification confidence score is used as a weighting factor for subsequent intent output.

[0025] Key reasoning nodes are extracted from each candidate thought chain. These key reasoning nodes include textual descriptions of fault phenomenon nodes and service item nodes. The textual descriptions of these key reasoning nodes are input into a pre-trained text embedding model to obtain the node feature vectors corresponding to each key reasoning node. Subsequently, the cosine similarity between corresponding key reasoning nodes of different candidate thought chains is calculated. When a candidate thought chain contains multiple key reasoning nodes, the semantic similarity between fault phenomenon nodes and between service item nodes are calculated separately, and a weighted average is performed to obtain the comprehensive semantic similarity of the candidate thought chain combination.

[0026] The comprehensive semantic similarity scores corresponding to each candidate thought chain combination are filled into a two-dimensional array of order equal to the total number of candidate thought chains to obtain a candidate thought chain semantic similarity matrix. Subsequently, each element in the semantic similarity matrix is ​​compared with a preset similarity threshold to obtain a binary relation matrix representing the consistency between candidate thought chains. While constructing the binary relation matrix, the original semantic similarity scores and their matrix coordinates of the fault phenomenon nodes and service item nodes in each candidate thought chain combination are recorded to form a node-level similarity detail table.

[0027] In an optional embodiment, the step of performing entity alignment verification between the structured entity elements of the candidate thought chain and a predefined automotive service domain knowledge graph, and calculating the knowledge verification confidence score, includes: Entity elements and action feature words in the candidate thought chain are extracted as feature words using word segmentation tools; The feature words are matched with graph nodes in the automotive service domain knowledge graph, and the ratio of the number of validly matched nodes to the total number of feature words is used as the first parameter. Verify whether the relationship lines between matching nodes conform to the preset constraint direction of the graph, and calculate the ratio of the number of lines that conform to the constraint to the total number of relationship lines between matching nodes as the second parameter; The knowledge verification confidence level is calculated by multiplying the first parameter and the second parameter by preset weight coefficients and then summing them.

[0028] A word segmentation tool based on the BiLSTM-CRF architecture and equipped with a dedicated automotive service dictionary is used. The neural network model involved in this architecture consists of a bidirectional long short-term memory network layer and a conditional random field layer cascaded together. The network input is a dense vector sequence of character embeddings that maps characters to real-valued arrays. The output is the sequence entity boundary label corresponding to each character, generated by decoding according to the optimal path. The formula for calculating the forget gate within the model is as follows: ,in Represents the output tensor of the forget gate. Represents the activation function. and Represents the weight transformation matrix. Represents a hidden historical state. Represents the input features at the current moment. This represents the bias parameter.

[0029] Candidate thought chains undergo text preprocessing to remove stop words and extract entity element words and action feature words. Entity element words include "spark plug" and "throttle body," while action feature words include "cleaning" and "replacement." The extracted feature words are matched against standard entity nodes in the automotive service domain knowledge graph using string comparison and synonym vector mapping. The number of feature words that successfully match standard entity nodes in the graph is counted, and the ratio of this number to the total number of feature words is used as the first parameter. For example, if 10 feature words are extracted from a single candidate thought chain, and 8 of them match standard entity nodes in the graph, the first parameter is 0.8. For nodes obtained through synonym vector mapping, if they can be uniquely assigned to a standard entity node in the graph, they are counted using the same method as string matching nodes in the calculation of the first parameter.

[0030] During the connection direction verification phase, the graph database is queried to obtain the actual relationship connections formed by the above 8 matching nodes in the candidate thought chain. These connections are then compared with the standard subject-verb-object triplet edges defined within the graph. For example, the "cleaning" action must point directionally to the "throttle body" and cannot point in the opposite direction or to fluid consumables. Assuming these 8 matching nodes generate 12 relationship connections in the thought chain, and 10 of them match the preset constraint directions of the graph, the ratio of the number of connections that meet the constraints to the total number of relationship connections is calculated, resulting in a second parameter of 0.833.

[0031] Using a pre-defined graph weighting system, the entity matching weight coefficient is preferably set to 0.4, the relationship connection weight coefficient is preferably set to 0.6, and the sum of the two is 1.0. Multiplying the first and second parameters by the aforementioned weights and summing them yields a knowledge verification confidence level of 0.82. This value is limited to the range of 0 to 1.

[0032] In an optional embodiment, calculating the semantic similarity of key reasoning nodes among the candidate thought chains and constructing a relationship matrix based on the semantic similarity includes: Extract the key reasoning nodes of each candidate thought chain and convert them into high-dimensional vector representations using a text embedding model; Calculate the cosine similarity between the high-dimensional vectors of the key reasoning nodes corresponding to any two candidate thought chains to obtain the node semantic similarity score; A similarity threshold is set. If the semantic similarity score of the node is greater than or equal to the similarity threshold, the corresponding element of the relationship matrix is ​​recorded as 1; if the semantic similarity score of the node is less than the similarity threshold, the corresponding element of the relationship matrix is ​​recorded as 0. Traverse all candidate thought chain combinations and generate a binary square matrix as a relation matrix to represent the consistency between candidate thought chains. While constructing the relation matrix, the original semantic similarity scores and their matrix coordinates between key reasoning nodes in each candidate thought chain combination are recorded to form a node-level similarity detail table.

[0033] Key reasoning nodes representing maintenance decisions are extracted from each candidate thought chain, with a maximum truncation length controlled within 128 tokens. These truncations are then input into a Sentence-BERT pre-trained text embedding model fine-tuned using automotive service industry corpus. This pre-trained model features a parameter-shared Siamese dual-channel encoder with an external mean pooling layer at the top for feature fusion. The network input is a truncated sequence of natural language text composed of concatenated discrete words, and the output is a fixed-dimensional continuous dense floating-point vector representation expressing the overall sentence semantic information. The feature calculation formula for the mean pooling layer is as follows: Where u represents the sentence-level feature representation vector, and L represents the length of the input sequence. The context hidden state tensor represents the word output by the encoder layer.

[0034] Natural language strings are encoded and mapped to dense floating-point high-dimensional vectors of 768 or 1024 dimensions. For the three candidate thought chains generated in parallel, their node vectors are paired, and the dot product between any two high-dimensional vectors is calculated and divided by the product of their magnitudes. This yields the cosine similarity distributed in the floating-point interval [-1, 1]. This result is the node semantic similarity score. For example, the cosine similarity score between the nodes "replace brake pads" and "replace brake pads" can reach 0.92. The structure of the candidate thought chain semantic similarity matrix is ​​illustrated below. Figure 2 As shown.

[0035] After obtaining all pairing scores, a pre-set similarity threshold is used, preferably fixed between 0.80 and 0.90, for example, a value of 0.85. Each score is compared to this threshold using a Boolean comparison: if the node semantic similarity score is greater than or equal to 0.85, it indicates that the two thought chains are semantically convergent at key nodes, and the corresponding coordinates (i,j) and symmetric coordinates (j,i) in the N×N relation matrix are assigned a value of 1; if the score is less than 0.85, semantic divergence exists, and the corresponding coordinate elements are assigned a value of 0. Simultaneously, specific fault phenomenon nodes and service item nodes in the candidate thought chain combination that are below the threshold, along with their original similarity scores, are written into a node-level similarity detail table.

[0036] By performing N×(N-1) / 2 traversals and judgments, a binary symmetric matrix with all elements on the main diagonal being 1 is generated, thereby realizing the expression of the consistent state of the output of multiple candidate thought chains.

[0037] S3. Calculate the thought chain divergence index based on the relationship matrix, and activate the adaptive mechanism according to the thought chain divergence index. If the thought chain divergence index is lower than the first threshold, weight the single intent with the highest confidence level by combining it with the knowledge verification confidence level. If the thought chain divergence index is greater than or equal to the first threshold but lower than the second threshold, locate the conflict node, generate a conflict analysis instruction to drive the large language model to make a decision, and output the composite intent. If the thought chain divergence index is greater than or equal to the second threshold, adjust the large language model generation temperature to a preset low random range, generate a clarifying reconstruction instruction based on the root cause of the divergence to optimize the initial instruction set, and start a new round of reasoning verification loop until the thought chain divergence index is lower than the second threshold, or output a manual review mark when the preset maximum number of retries is reached.

[0038] The number of elements with a value of 0 outside the main diagonal of the relationship matrix is ​​counted, and this number is divided by the total number of off-diagonal elements to obtain the inconsistency ratio. This inconsistency ratio is used as the thought chain divergence index. When the thought chain divergence index is less than a preset first threshold, the intent labels generated by each candidate thought chain are converted into one-hot encoded vectors. The intent labels are mapped from the service item nodes in the corresponding candidate thought chain. The one-hot encoded vectors are multiplied by the knowledge verification confidence and then weighted and summed. The index position of the highest weighted score is found, and the single intent corresponding to that index is output.

[0039] When the divergence index of the thought chain is greater than or equal to the first threshold and less than the second threshold, the coordinates of elements with a value of 0 in the relationship matrix are traversed, and the node-level similarity details table of the corresponding candidate thought chain combinations is read. Key reasoning nodes with original semantic similarity lower than the similarity threshold or the largest textual differences are identified as conflict nodes. The text content of the conflict nodes is extracted, and a conflict resolution instruction containing a requirement for the large language model to analyze the cause of the contradiction and provide a comprehensive conclusion is generated through string concatenation. The large language model is called again for reasoning to generate a composite intent that integrates multiple intent labels and outputs it.

[0040] When the thought chain divergence index is greater than or equal to the second threshold, the decoding configuration of the large language model is adjusted to a preset low random sampling configuration to reduce randomness and divergence in the re-inference process. Specifically, the generated temperature parameter is adjusted to a preset low random range, enabling the model to perform convergent inference around the located conflict nodes. Subsequently, semantic difference words between conflict nodes are extracted as the root cause of divergence, and a clarifying reconstruction instruction is generated based on a preset reconstruction template. The clarifying reconstruction instruction is used to guide the model to re-infer around conflict entities, fault phenomena, and service items, and the initial instruction set is updated with this clarifying reconstruction instruction. The reconstructed instruction set is used to drive the large language model to generate new candidate thought chains, and the above entity verification, similarity calculation, and divergence index evaluation process is repeated. The loop terminates when the latest calculated thought chain divergence index is less than the second threshold; if the number of loops reaches the preset maximum number of retries and the thought chain divergence index is still not lower than the second threshold, the loop terminates and a manual review mark is output.

[0041] In an optional embodiment, calculating the thought chain divergence index based on the relationship matrix includes: Count the total number of elements with a value of 0 in the relation matrix, excluding self-comparison elements on the diagonal; The inconsistency ratio is calculated by dividing the total number of elements with a value of 0 by the total number of off-diagonal elements in the relation matrix. The inconsistency ratio is used as the thought chain divergence index, and the thought chain divergence index is compared with the first threshold and the second threshold, wherein the first threshold is less than the second threshold.

[0042] After receiving the N×N binary relation matrix corresponding to N candidate thought chains, the discrete distribution elements of the matrix are scanned. During this process, the N self-comparison elements at position i=j on the matrix diagonal, which are necessarily 1 due to their identical vectors, are excluded.

[0043] At this point, the total number of elements with a value of 0 among the remaining elements is counted. Assuming that the number of elements with a value of 0 is 2, these 2 mismatched elements are divided by the total number of off-diagonal elements, which is 6. The inconsistency ratio is calculated to be approximately 0.33. This floating-point value quantifies the dispersion of the current inference batch.

[0044] This inconsistency ratio of 0.33 is defined as the Disagreement Index (DCI) of the thought chain in this round of inference. This index is compared with a predefined first threshold (e.g., 0.20) and a second threshold (e.g., 0.50). Based on the judgment result of 0.20 < 0.33 < 0.50, the output generated by the current model is determined to fall into the medium disagreement range, thus driving the downstream composite decision-making or low-random reconstruction verification process with quantitative parameters.

[0045] In an optional embodiment, the step of locating conflict nodes, generating conflict resolution instructions to drive a large language model to make decisions, and outputting a composite intent includes: Traverse the key reasoning nodes of the candidate thought chain corresponding to the elements with a value of 0 in the relation matrix, and read the corresponding node-level similarity details table to extract the original reasoning fragments with different text content as the conflict nodes. The text fragments of the conflicting nodes are concatenated into a preset comparison and adjudication instruction template, driving the large language model to output a composite maintenance operation object that integrates conflicting features from the given entity options; The adjudication results returned by the large language model are structurally merged with the original consistency nodes to generate the composite intent containing conflict resolution paths and service items, and then output.

[0046] When a moderate divergence logic is triggered, the system traces back along the (i,j) coordinates of the relation matrix to extract two or more candidate thought chains corresponding to elements with a value of 0. It then reads the node-level similarity details table corresponding to that coordinate, identifying fault phenomenon nodes or service item nodes with similarity below a threshold as conflict nodes; nodes with similarity meeting the threshold are marked as consistent nodes and retained. For example, if node A provides a solution of "replacing the catalytic converter" and node B provides a solution of "cleaning the throttle body only," the text truncation length is controlled within 50 to 100 characters.

[0047] The extracted contradictory fragments are filled into a structured comparison and adjudication instruction template. For example, a prompt is constructed: "A conflict in maintenance plans has been detected: Plan 1 [{Node A}] and Plan 2 [{Node B}]. Please provide a comprehensive solution strategy based on the standard working hour dictionary". This drives the large language model to perform second-order logical deduction within the limited given entity options and outputs a composite maintenance operation object that covers the effective features of both parties, such as: "Joint cleaning of three-way catalytic converter and throttle body without disassembly".

[0048] After obtaining the JSON or XML adjudication results returned by the large language model, a graph subgraph merging operation is performed. Consistent nodes with a similarity calculation threshold of 1 and no disputes in the original candidate thought chain, such as the basic maintenance node, are retained and structurally merged and stitched with the newly generated joint cleaning adjudication node according to the execution order and tree-like hierarchical relationship.

[0049] After the merger is completed, the data dictionary containing conflict resolution reasoning paths, such as "due to simultaneous fault reports from the exhaust and intake systems, joint cleaning is performed based on the combined opinions of A and B", and the service item array will be packaged to generate a standard composite intent object and output to the automotive service terminal system. The time taken from locating the entire conflict node to outputting the composite intent is preferably controlled within 500 to 1200 ms.

[0050] In an optional embodiment, adjusting the large language model generation temperature to a preset low random range, generating clarifying reconstruction instructions based on the root causes of divergence to optimize the initial instruction set, and initiating a new round of inference verification loop until the thought chain divergence index is lower than the second threshold, or outputting a manual review marker when the preset maximum number of retries is reached, includes: Modify the configuration parameters of the large language model to adjust the generated temperature parameter to a preset low random range; Extract the original text that caused the divergence and the reasoning path that led to the divergence, integrate the re-reasoning requirements to generate clarifying reconstruction instructions, and update the initial instruction set; The parallel reasoning process of the large language model is retried to generate new candidate thought chains, and the entity alignment verification operation and the construction steps of the relation matrix used to represent the consistency between candidate thought chains are re-executed. When the recalculated thought chain divergence index is lower than the second threshold, the new round of reasoning verification loop is terminated and the recognition result is output according to the interval of the thought chain divergence index. When the number of iterations reaches the preset maximum number of retries and the recalculated thought chain divergence index is still not lower than the second threshold, the new round of reasoning verification loop is terminated and a manual review mark is output.

[0051] When the divergence index of the thought chain is greater than or equal to a preset second threshold, for example, a measured index of 0.65 that is higher than the second threshold of 0.50, the current candidate thought chain is determined to be in a state of logical divergence. At this point, the generation and decoding configuration of the large language model is adjusted to a preset low random sampling configuration. Specifically, the generation temperature parameter is adjusted to 0.1 to 0.3 to make the probability distribution of candidate lexical units more concentrated, thereby reducing the randomness and divergence of the regenerated text. Through this method, the regenerated candidate thought chain performs convergent reasoning around the identified conflicting entities, fault phenomena, and service items, improving the consistency of multi-round reasoning results.

[0052] Extract the original text corpus and corresponding reasoning paths that caused the discrepancies in the previous round of reasoning, and generate clarifying reconstruction instructions based on a preset reasoning template. The clarifying reconstruction instructions are used to prompt the large language model to re-derive the identified conflicting entities, conflicting fault phenomena, and conflicting service items. For example: "There are inconsistencies in the fault attribution paths in the previous reasoning. Please re-derive based on the identified conflicting entities, fault phenomena, and service items."

[0053] The preset reasoning template is a pre-configured prompt word framework, including conflict entity filling positions, conflict fault phenomenon filling positions, conflict service item filling positions, divergence reason filling positions, and output format constraint positions. When generating a clarifying refactoring instruction, the conflict entities, fault phenomena, service items, and corresponding divergence reasons located in the previous round of reasoning are filled into the preset reasoning template to obtain a clarifying refactoring instruction for generating the next round of candidate thought chains, and the initial instruction set is updated using this clarifying refactoring instruction.

[0054] This reconstruction instruction set triggers the parallel regeneration process of the large language model, generating three entirely new candidate thought chains. This newly generated data is then fed back into the pipeline, where it sequentially performs entity alignment confidence checks based on the knowledge graph and reconstruction operations on a 3×3 relation matrix based on cosine similarity. An internal maximum retry counter variable is included. In a new round or multiple iterations, if the recalculated thought chain divergence index falls below the second threshold, the new inference verification loop terminates, and the business identification result is output based on the latest interval the index falls into. If convergence is not achieved even after reaching the maximum retry limit, the loop terminates, and an abnormal work order log requiring manual review is output, improving the feasibility and runtime efficiency control capabilities of the entire business chain.

[0055] In the second embodiment, the present invention also proposes an automotive service work order intent recognition system based on a large language model, comprising: The generation module is used to acquire the automotive service work order text data to be identified; it uses the initial non-homogeneous perspective instructions to drive the large language model in parallel, and generates multiple candidate thought chains from the dimensions of customer expression, technician inspection and parts demand. The candidate thought chains contain a structured reasoning process from entity elements, fault phenomena to service items. The construction module is used to perform entity alignment verification operations between the structured entity elements of the candidate thinking chains and the predefined automotive service domain knowledge graph, calculate the knowledge verification confidence and use the knowledge verification confidence as a weight factor; calculate the semantic similarity of key reasoning nodes between each candidate thinking chain, and construct a relationship matrix based on the semantic similarity. The output module is used to calculate the thought chain divergence index based on the relation matrix, and activate an adaptive mechanism according to the thought chain divergence index. If the thought chain divergence index is lower than a first threshold, it is weighted in combination with the knowledge verification confidence level, and the single intent with the highest confidence level is output. If the thought chain divergence index is greater than or equal to the first threshold and lower than a second threshold, the conflict node is located, a conflict analysis instruction is generated to drive the large language model to make a decision, and a composite intent is output. If the thought chain divergence index is greater than or equal to the second threshold, the generation temperature of the large language model is adjusted to a preset low random range, a clarifying reconstruction instruction is generated based on the root cause of the divergence to optimize the initial instruction set, and a new round of reasoning verification loop is started until the thought chain divergence index is lower than the second threshold, or when the preset maximum number of retries is reached, a manual review mark is output.

[0056] In an optional embodiment, the method of using initial non-homogeneous perspective instructions to drive a large language model in parallel generates multiple candidate thought chains from the dimensions of customer expression, technician inspection, and parts requirements, including: Fill the customer's abnormal description in the auto service work order text data into the customer experience instruction template, and extract the target part and the fault phenomenon as the first reasoning node in the candidate thinking chain of customer description dimension. Extract fault codes generated by diagnostic equipment, combine them with vehicle structure query instructions, and generate in-depth component repair records as the second reasoning node in the technician inspection dimension candidate thinking chain; The parts requirements and consumable details of the automotive service work order text data are analyzed, and the related operation items are deduced by the use of parts to generate the third reasoning node in the candidate thinking chain of parts requirements. The first inference node, the second inference node, and the third inference node are respectively combined with the entity elements, fault phenomena, and service item nodes extracted under their respective dimensions to complete the candidate thought chain for the corresponding dimension.

[0057] In an optional embodiment, the step of performing entity alignment verification between the structured entity elements of the candidate thought chain and a predefined automotive service domain knowledge graph, and calculating the knowledge verification confidence score, includes: Entity elements and action feature words in the candidate thought chain are extracted as feature words using word segmentation tools; The feature words are matched with graph nodes in the automotive service domain knowledge graph, and the ratio of the number of validly matched nodes to the total number of feature words is used as the first parameter. Verify whether the relationship lines between matching nodes conform to the preset constraint direction of the graph, and calculate the ratio of the number of lines that conform to the constraint to the total number of relationship lines between matching nodes as the second parameter; The knowledge verification confidence level is calculated by multiplying the first parameter and the second parameter by preset weight coefficients and then summing them.

[0058] In an optional embodiment, calculating the semantic similarity of key reasoning nodes among the candidate thought chains and constructing a relationship matrix based on the semantic similarity includes: Extract the key reasoning nodes of each candidate thought chain and convert them into high-dimensional vector representations using a text embedding model; Calculate the cosine similarity between the high-dimensional vectors of the key reasoning nodes corresponding to any two candidate thought chains to obtain the node semantic similarity score; A similarity threshold is set. If the semantic similarity score of the node is greater than or equal to the similarity threshold, the corresponding element of the relationship matrix is ​​recorded as 1; if the semantic similarity score of the node is less than the similarity threshold, the corresponding element of the relationship matrix is ​​recorded as 0. Traverse all candidate thought chain combinations and generate a binary square matrix as a relation matrix to represent the consistency between candidate thought chains. While constructing the relation matrix, the original semantic similarity scores and their matrix coordinates between key reasoning nodes in each candidate thought chain combination are recorded to form a node-level similarity detail table.

[0059] In an optional embodiment, calculating the thought chain divergence index based on the relationship matrix includes: Count the total number of elements with a value of 0 in the relation matrix, excluding self-comparison elements on the diagonal; The inconsistency ratio is calculated by dividing the total number of elements with a value of 0 by the total number of off-diagonal elements in the relation matrix. The inconsistency ratio is used as the thought chain divergence index, and the thought chain divergence index is compared with the first threshold and the second threshold, wherein the first threshold is less than the second threshold.

[0060] In an optional embodiment, the step of locating conflict nodes, generating conflict resolution instructions to drive a large language model to make decisions, and outputting a composite intent includes: Traverse the key reasoning nodes of the candidate thought chain corresponding to the elements with a value of 0 in the relation matrix, and read the corresponding node-level similarity details table to extract the original reasoning fragments with different text content as the conflict nodes. The text fragments of the conflicting nodes are concatenated into a preset comparison and adjudication instruction template, driving the large language model to output a composite maintenance operation object that integrates conflicting features from the given entity options; The adjudication results returned by the large language model are structurally merged with the original consistency nodes to generate the composite intent containing conflict resolution paths and service items, and then output.

[0061] In an optional embodiment, adjusting the large language model generation temperature to a preset low random range, generating clarifying reconstruction instructions based on the root causes of divergence to optimize the initial instruction set, and initiating a new round of inference verification loop until the thought chain divergence index is lower than the second threshold, or outputting a manual review marker when the preset maximum number of retries is reached, includes: Modify the configuration parameters of the large language model to adjust the generated temperature parameter to a preset low random range; Extract the original text that caused the divergence and the reasoning path that led to the divergence, integrate the re-reasoning requirements to generate clarifying reconstruction instructions, and update the initial instruction set; The parallel reasoning process of the large language model is retried to generate new candidate thought chains, and the entity alignment verification operation and the construction steps of the relation matrix used to represent the consistency between candidate thought chains are re-executed. When the recalculated thought chain divergence index is lower than the second threshold, the new round of reasoning verification loop is terminated and the recognition result is output according to the interval of the thought chain divergence index. When the number of iterations reaches the preset maximum number of retries and the recalculated thought chain divergence index is still not lower than the second threshold, the new round of reasoning verification loop is terminated and a manual review mark is output.

[0062] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for recognizing the intent of automotive service work orders based on a large language model, characterized in that, include: Obtain the text data of the auto service work order to be identified; By using initial non-homogeneous perspective instructions to drive the large language model in parallel, multiple candidate thought chains are generated from the dimensions of customer expression, technician inspection, and parts demand. The candidate thought chains contain a structured reasoning process from entity elements, fault phenomena to service items. The structured entity elements of the candidate thinking chain are aligned and verified with the predefined automotive service domain knowledge graph. The knowledge verification confidence is calculated and used as a weighting factor. Calculate the semantic similarity of key reasoning nodes among each candidate thought chain, and construct a relationship matrix based on the semantic similarity; The thought chain divergence index is calculated based on the relationship matrix, and an adaptive mechanism is activated according to the thought chain divergence index. If the thought chain divergence index is lower than the first threshold, it is weighted in combination with the knowledge verification confidence and the single intent with the highest confidence is output. If the thought chain divergence index is greater than or equal to the first threshold and lower than the second threshold, the conflict node is located, a conflict analysis instruction is generated to drive the large language model to make a decision, and a composite intent is output. If the thought chain divergence index is greater than or equal to the second threshold, the large language model generation temperature is adjusted to a preset low random range, a clarifying reconstruction instruction is generated based on the root cause of the divergence to optimize the initial instruction set, and a new round of reasoning verification loop is started until the thought chain divergence index is lower than the second threshold, or when the preset maximum number of retries is reached, a manual review mark is output.

2. The method according to claim 1, characterized in that, The method utilizes initial non-homogeneous perspective instructions to drive a large language model in parallel, generating multiple candidate thought chains from the dimensions of customer expression, technician inspection, and parts requirements, including: Fill the customer's abnormal description in the auto service work order text data into the customer experience instruction template, and extract the target part and the fault phenomenon as the first reasoning node in the candidate thinking chain of customer description dimension. Extract fault codes generated by diagnostic equipment, combine them with vehicle structure query instructions, and generate in-depth component repair records as the second reasoning node in the technician inspection dimension candidate thinking chain; The parts requirements and consumable details of the automotive service work order text data are analyzed, and the related operation items are deduced by the use of parts to generate the third reasoning node in the candidate thinking chain of parts requirements. The first inference node, the second inference node, and the third inference node are respectively combined with the entity elements, fault phenomena, and service item nodes extracted under their respective dimensions to complete the candidate thought chain for the corresponding dimension.

3. The method according to claim 2, characterized in that, The step of performing entity alignment verification between the structured entity elements of the candidate thought chain and the predefined automotive service domain knowledge graph, and calculating the knowledge verification confidence score, includes: Entity elements and action feature words in the candidate thought chain are extracted as feature words using word segmentation tools; The feature words are matched with graph nodes in the automotive service domain knowledge graph, and the ratio of the number of validly matched nodes to the total number of feature words is used as the first parameter. Verify whether the relationship lines between matching nodes conform to the preset constraint direction of the graph, and calculate the ratio of the number of lines that conform to the constraint to the total number of relationship lines between matching nodes as the second parameter; The knowledge verification confidence level is calculated by multiplying the first parameter and the second parameter by preset weight coefficients and then summing them.

4. The method according to claim 1, characterized in that, The step of calculating the semantic similarity of key reasoning nodes among various candidate thought chains and constructing a relationship matrix based on the semantic similarity includes: Extract the key reasoning nodes of each candidate thought chain and convert them into high-dimensional vector representations using a text embedding model; Calculate the cosine similarity between the high-dimensional vectors of the key reasoning nodes corresponding to any two candidate thought chains to obtain the node semantic similarity score; A similarity threshold is set. If the semantic similarity score of the node is greater than or equal to the similarity threshold, the corresponding element of the relationship matrix is ​​recorded as 1; if the semantic similarity score of the node is less than the similarity threshold, the corresponding element of the relationship matrix is ​​recorded as 0. Traverse all candidate thought chain combinations and generate a binary square matrix as a relation matrix to represent the consistency between candidate thought chains. While constructing the relation matrix, the original semantic similarity scores and their matrix coordinates between key reasoning nodes in each candidate thought chain combination are recorded to form a node-level similarity detail table.

5. The method according to claim 1, characterized in that, The calculation of the thought chain divergence index based on the relationship matrix includes: Count the total number of elements with a value of 0 in the relation matrix, excluding self-comparison elements on the diagonal; The inconsistency ratio is calculated by dividing the total number of elements with a value of 0 by the total number of off-diagonal elements in the relation matrix. The inconsistency ratio is used as the thought chain divergence index, and the thought chain divergence index is compared with the first threshold and the second threshold, wherein the first threshold is less than the second threshold.

6. The method according to claim 1, characterized in that, The process involves locating conflict nodes, generating conflict resolution instructions to drive the large language model to make a decision, and outputting a composite intent, including: Traverse the key reasoning nodes of the candidate thought chain corresponding to the elements with a value of 0 in the relation matrix, and read the corresponding node-level similarity details table to extract the original reasoning fragments with different text content as the conflict nodes. The text fragments of the conflicting nodes are concatenated into a preset comparison and adjudication instruction template, driving the large language model to output a composite maintenance operation object that integrates conflicting features from the given entity options; The adjudication results returned by the large language model are structurally merged with the original consistency nodes to generate the composite intent containing conflict resolution paths and service items, and then output.

7. The method according to claim 1, characterized in that, The process of adjusting the large language model generation temperature to a preset low random range, generating clarifying reconstruction instructions based on the root causes of divergence to optimize the initial instruction set, and starting a new round of inference verification loop until the thought chain divergence index is lower than the second threshold, or outputting a manual review marker when the preset maximum number of retries is reached, includes: Modify the configuration parameters of the large language model to adjust the generated temperature parameter to a preset low random range; Extract the original text that caused the divergence and the reasoning path that led to the divergence, integrate the re-reasoning requirements to generate clarifying reconstruction instructions, and update the initial instruction set; The parallel reasoning process of the large language model is retried to generate new candidate thought chains, and the entity alignment verification operation and the construction steps of the relation matrix used to represent the consistency between candidate thought chains are re-executed. When the recalculated thought chain divergence index is lower than the second threshold, the new round of reasoning verification loop is terminated and the recognition result is output according to the interval of the thought chain divergence index. When the number of iterations reaches the preset maximum number of retries and the recalculated thought chain divergence index is still not lower than the second threshold, the new round of reasoning verification loop is terminated and a manual review mark is output.

8. A vehicle service work order intent recognition system based on a large language model, characterized in that, include: The generation module is used to obtain the text data of the auto service work order to be identified; By using initial non-homogeneous perspective instructions to drive the large language model in parallel, multiple candidate thought chains are generated from the dimensions of customer expression, technician inspection, and parts demand. The candidate thought chains contain a structured reasoning process from entity elements, fault phenomena to service items. The construction module is used to perform entity alignment verification operations between the structured entity elements of the candidate thinking chain and the predefined automotive service domain knowledge graph, calculate the knowledge verification confidence score, and use the knowledge verification confidence score as a weighting factor. Calculate the semantic similarity of key reasoning nodes among each candidate thought chain, and construct a relationship matrix based on the semantic similarity; The output module is used to calculate the thought chain divergence index based on the relationship matrix, and activate the adaptive mechanism according to the thought chain divergence index. If the thought chain divergence index is lower than the first threshold, it is weighted in combination with the knowledge verification confidence and outputs the single intent with the highest confidence. If the thought chain divergence index is greater than or equal to the first threshold and lower than the second threshold, the conflict node is located, a conflict analysis instruction is generated to drive the large language model to make a decision, and a composite intent is output. If the thought chain divergence index is greater than or equal to the second threshold, the large language model generation temperature is adjusted to a preset low random range, a clarifying reconstruction instruction is generated based on the root cause of the divergence to optimize the initial instruction set, and a new round of reasoning verification loop is started until the thought chain divergence index is lower than the second threshold, or when the preset maximum number of retries is reached, a manual review mark is output.

9. The system according to claim 8, characterized in that, The method utilizes initial non-homogeneous perspective instructions to drive a large language model in parallel, generating multiple candidate thought chains from the dimensions of customer expression, technician inspection, and parts requirements, including: Fill the customer's abnormal description in the auto service work order text data into the customer experience instruction template, and extract the target part and the fault phenomenon as the first reasoning node in the candidate thinking chain of customer description dimension. Extract fault codes generated by diagnostic equipment, combine them with vehicle structure query instructions, and generate in-depth component repair records as the second reasoning node in the technician inspection dimension candidate thinking chain; The parts requirements and consumable details of the automotive service work order text data are analyzed, and the related operation items are deduced by the use of parts to generate the third reasoning node in the candidate thinking chain of parts requirements. The first inference node, the second inference node, and the third inference node are respectively combined with the entity elements, fault phenomena, and service item nodes extracted under their respective dimensions to complete the candidate thought chain for the corresponding dimension.

10. The system according to claim 8, characterized in that, The step of performing entity alignment verification between the structured entity elements of the candidate thought chain and the predefined automotive service domain knowledge graph, and calculating the knowledge verification confidence score, includes: Entity elements and action feature words in the candidate thought chain are extracted as feature words using word segmentation tools; The feature words are matched with graph nodes in the automotive service domain knowledge graph, and the ratio of the number of validly matched nodes to the total number of feature words is used as the first parameter. Verify whether the relationship lines between matching nodes conform to the preset constraint direction of the graph, and calculate the ratio of the number of lines that conform to the constraint to the total number of relationship lines between matching nodes as the second parameter; The knowledge verification confidence level is calculated by multiplying the first parameter and the second parameter by preset weight coefficients and then summing them.

Images