Electronic component classification method based on LLM and classification tree
By constructing a standardized classification knowledge base and hierarchical classification process using large language models and clustering technology, the problems of inconsistent classification criteria, unbalanced data, and non-standard indicators in the electronic component classification system are solved, and efficient and stable automated component classification is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NAT INNOVATION INST OF DEFENSE TECH PLA ACAD OF MILITARY SCI
- Filing Date
- 2026-03-12
- Publication Date
- 2026-06-09
AI Technical Summary
The existing classification system for electronic components suffers from problems such as inconsistent classification criteria, uneven data distribution, sparse samples, and non-standardized indicator descriptions, which makes model learning difficult and hinders efficient and automated classification.
We employ a method based on Large Language Model (LLM) and clustering techniques to construct a standardized classification knowledge base and hierarchical classification process. Through initial classification, cluster screening, indicator system construction, and classification tree constraints, we achieve coarse-to-fine classification decisions.
It significantly reduces the cost of building high-quality component datasets, improves the model's generalization ability and recall rate for sparse categories, and solves the classification stability problem in scenarios with mixed classification criteria and missing data.
Smart Images
Figure CN122176390A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of artificial intelligence and electronic information, specifically to a method for constructing an index system for electronic components based on a large language model and clustering, and a device classification method constrained by a classification tree. Background Technology
[0002] The classification of electronic components is a crucial foundation for supporting component management, selection, and application. Existing electronic component classification systems typically include multiple primary types (such as digital integrated circuits, resistors, capacitors, and inductors), further subdivided into a total of 244 tertiary types. However, this system faces a series of severe challenges in practical applications and data construction: (1) Inconsistent classification criteria: There are significant differences in the classification logic of the three-level types under different first-level types. For example, capacitors in resistive-capacitive-inductive devices are mainly subdivided according to the process principle, while microprocessors in digital integrated circuits are divided according to function, which leads to difficulties in feature extraction and model generalization.
[0003] (2) The data distribution is severely unbalanced: The existing actual data (e.g., data randomly selected from the data manual page of the official website of Electronic Engineering World) exhibits an extreme long-tail distribution among different third-level types. For example, electrical connectors account for as high as 41.3%, while key semiconductor devices such as infrared detectors, amplifier chips, FPGAs, and CPUs account for less than 2%, which severely restricts the model's ability to identify a few categories.
[0004] (3) The problem of sample sparsity is prominent: there are a large number of third-level types (244 types), but the labeled data (784 records) only covers 174 types, of which 42 types have only 1 sample data, which is far from enough to support the training needs of traditional supervised learning models.
[0005] (4) Inconsistent and incomplete description of indicators: Component data from different sources vary greatly in terms of the definition, description format and completeness of technical indicators, making it difficult to construct a unified and comparable feature representation. These problems are intertwined, resulting in high costs for constructing high-quality, category-balanced component datasets and making it difficult to establish a consistent, complete and unified system of component performance indicators, which in turn seriously hinders the development and application of efficient and automated component classification models.
[0006] Existing research has explored cue learning methods using large language models for zero-shot and few-shot classification, such as thought chain cueing, automatic cue generation, and rule-based cue and instruction fine-tuning. However, these methods struggle to systematically encode a complete classification system, including definitions, core indicators, negation constraints, and hierarchical decision logic, into a coherent and operable cue structure. They still face significant challenges when dealing with the unique characteristics of electronic component classification, such as inconsistent classification criteria, over 200+ types, extremely sparse long-tail categories, and complex reasoning described by multi-source heterogeneous indicators. This makes it difficult to stably and accurately complete hierarchical, rule-constrained classification tasks. Summary of the Invention
[0007] (a) Technical problems to be solved This invention aims to solve the problems faced by existing technologies in the automated classification of electronic components, such as mixed classification criteria leading to difficulties in model learning, extremely unbalanced data distribution affecting minority class identification, sparse samples insufficient for training effective models, and non-standard descriptions of multi-source heterogeneous indicators hindering unified feature representation.
[0008] (II) Technical Solution To address the aforementioned problems, this invention proposes a hierarchical classification method based on a unified indicator system constructed through large language models and clustering collaboration, constrained by a classification tree. The core idea is as follows: First, utilizing limited expert knowledge and publicly available data, a standardized classification knowledge base (including high-quality datasets, a structured indicator system, and hierarchical conceptual categories) is automatically constructed using LLM and clustering techniques. Then, based on this knowledge base and the inherent classification tree structure, a step-by-step, hierarchically constrained reasoning process is designed to guide LLM in accurately completing coarse-to-fine classification decisions. The process is as follows: Figure 1 As shown, it specifically includes the following two stages and six steps: 1. Classification Knowledge Construction Stage: Step 1, Initial three-level coarse classification: Using a small number of expert-annotated samples to guide the large language model, the collected publicly available component description data is initially classified into three levels.
[0009] Step 2, High-quality data screening: Based on the preliminary results of Step 1, the data is cleaned using an attribute similarity clustering algorithm to output a balanced, high-quality three-level type dataset, alleviating the problems of data imbalance and sample sparsity.
[0010] Step 3, Construction of a unified three-level type indicator system: Based on the high-quality dataset, supplementary expert samples, and basic definitions of the three-level types obtained in Step 2, a structured performance indicator system covering all three-level types is constructed using a large language model to provide a standardized reference for subsequent classification.
[0011] Step 4, Abstraction of First-Level Type Concepts: Based on the product system classification table, the complete indicator system of third-level types obtained in Step 3, and typical product examples, a large language model is used to summarize and analyze from three dimensions: function, structure, and application. The conceptual connotation of each first-level type is abstracted and its semantic boundaries are clarified. 2. Automatic device sorting stage: Step 5, Initial screening of three-level types: Input the target device description, integrate the classification tree structure of the type system conversion with the knowledge obtained in steps 3 and 4, and initially screen and recommend up to 2 candidate three-level types under each type level to make full use of hierarchical constraints to narrow the search range.
[0012] Step 6, Precise determination of unique tertiary types: For the candidate tertiary types obtained in Step 1, input their detailed definition, core indicators, and lower-level classification tree structure. Perform deep indicator matching through a large language model, and combine self-consistency and voting verification mechanisms to guide the model to strictly follow the predefined domain rules and indicator system to determine the final unique tertiary type.
[0013] (III) Beneficial Effects Compared with the prior art, the present invention has the following beneficial effects: 1. This invention automatically generates high-quality, balanced classification datasets and standardized performance indicator systems through clustering screening and LLM construction, significantly reducing the cost and threshold for building high-quality component datasets.
[0014] 2. This invention creatively integrates the inherent hierarchical classification tree structure as a strong constraint into the inference process of LLM. Through a step-by-step mechanism of first coarse screening and then fine judgment, it effectively solves the problems of model attention dispersion and decision instability caused by the large number of types and excessively long prompts.
[0015] 3. The method of this invention exhibits excellent generalization ability and robustness when facing incomplete scenarios with mixed classification criteria and highly missing data. Experiments show that its classification completeness F1 score is significantly better than methods without classification tree constraints or using only coarse-grained trees under various data completeness conditions.
[0016] 4. This invention unifies heterogeneous feature representations through the standardization of indicator systems and the abstraction of conceptual categories, enabling the model to more accurately understand and match the key features of long-tail categories and marginal samples, thereby improving the recognition recall rate of sparse categories and easily confused types. Attached Figure Description
[0017] Figure 1 This is a technical flowchart of an electronic component classification method based on LLM and classification trees; Figure 2 To compare the metrics of each method on the exp_3 dataset; Figure 3 This provides a comparison of metrics between Parameters and the leading Trees method on the exp_3 dataset. Detailed Implementation
[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can fully implement the present invention.
[0019] 1. Model Configuration and Deployment Step 1: Coarse Classification. Due to the massive amount of data and low accuracy requirements, the Gemma2-27B model is used and deployed locally via Ollam. Step 4 requires high-level conceptual abstraction using the DeepSeekR1 model, deployed via MindIE; Steps 2, 3, 5, and 6 all use the Qwen3-32B model and are deployed via MindIE.
[0020] like Figure 1 As shown, the entire system comprises the following core modules: ① Initial three-level type classification module, ② High-quality data screening module based on clustering, ③ Three-level type indicator system construction module based on multi-source data, ④ First-level type concept summary module, ⑤ Initial screening module for three-level types, and ⑥ Three-level type classification module. Modules ① to ④ typically only need to be executed once to build the basic data resources (high-quality label dataset, indicator system, and first-level type concepts) when the classification system remains stable. Modules ⑤ and ⑥ are the service modules that actually perform the automatic classification of components.
[0021] Step 1: Three-level classification based on small sample size This step aims to perform three-level type identification of product information in publicly available data by learning from a small number of expert-annotated samples and leveraging the semantic understanding capabilities of a large model. The specific process is as follows: 1) Input data preparation. Reference data: Sample data annotated by experts. Data to be determined Product names and technical specifications derived from publicly available data.
[0022] 2) Type matching and scoring. Traverse all 174 third-level types. For each type Based on the product name and main technical parameters, a semantic similarity score is calculated between the product and the sample. (0~100 points); each type receives a confidence score. For any test sample The scoring process is as follows:
[0023] 3) Comprehensive judgment and selection. Based on the scores of all types, select the top three types with the highest scores. The three types and their corresponding prompts are reorganized and then input back into the model to obtain a more accurate final classification result.
[0024]
[0025]
[0026] 4) Output: The output is the preliminary three-level type label for each record. .
[0027]
[0028] Step 2: Filtering high-quality data based on iterative clustering This step aims to refine and optimize the preliminary third-level type labels generated in Step 1 through cluster analysis and attribute similarity filtering mechanisms, thereby obtaining high-quality public data labels. The specific process is as follows: 1) Input data preparation. Combine the publicly available data with the preliminary three-level type labels generated in step 1. The main performance indicators are directly structured into attribute-value pairs using a large model, and the structured information is encoded into vectors using BGE encoding. Grouped by three-level type tags .
[0029] 2) Clustering and filtering. For the same tertiary type... The manual data was grouped using the K-Means clustering method; product performance parameters for each cluster center were calculated, and the correctness of the types was verified using a large model; records with a similarity greater than 0.8 to the products at the center were retained based on attribute similarity; when the number of remaining products under a certain third-level type is less than 40% of the original number, the type selection is complete. The attribute similarity calculation formula is as follows:
[0030] 3) Iterative control strategy. The number of clusters is dynamically adjusted between 500 and 200; the deeper the iteration layer, the smaller the number of clusters; ensuring that the screening process is fully efficient.
[0031] 4) Output Results. Output high-quality, publicly available level 3 category tags after filtering and optimization. .
[0032] Step 3: Construction of a three-level type indicator system based on multi-source data This step aims to construct a standardized performance parameter index system for each level-three type based on high-quality level-three type labels and related reference knowledge, utilizing a large model. The specific process is as follows: 1) Input data preparation. Reference data: expert-annotated sample data, basic definitions and core parameters of the three-level types; publicly available data. (Including performance parameter descriptions), high-quality three-level type tags output from step 2.
[0033] 2) Indicator System Construction. Example samples and high-quality public data for each of the three levels are integrated to form prompt words. These prompt words are then used to ask questions of the large model to construct the indicator system for each level. A step-by-step reasoning method is adopted: learning concept definitions and core parameters → identifying performance parameter items → eliminating redundant parameters → verifying supplementary parameters → validating physical units; standardized JSON format performance parameter items and their corresponding physical units are output. (Level 3 types) The formula for integrating prompts is as follows:
[0034] 3) Output Results. Output the complete performance parameter index system corresponding to each of the three levels. Example output results for the DC / DC converter are as follows: {'Input Voltage': 'V', 'Output Voltage': 'V', 'Output Power': 'W', 'Efficiency': '-', 'Output Current': 'A', 'Temperature': '°C', 'Input Voltage Range': 'V', 'Output Voltage Range': 'V', 'Conversion Efficiency': '%', 'Operating Temperature Range': '°C', 'Ripple Voltage': 'mV', 'Overcurrent Protection': 'A', 'Short Circuit Protection': '-', 'Insulation Resistance': 'Ω', 'Isolation Method': '-'}. Three Levels The extracted indicator system is represented as follows:
[0035]
[0036] Step 4: Summarize the concepts of primary types based on thought chains and self-consistency. This step aims to construct a complete primary-level type classification system based on a multi-dimensional summary of the three-level type definition, and to achieve accurate primary-level type classification of publicly available data. The specific process is as follows: 1) Input Data Preparation. Input data includes: product system classification table, basic definitions and core parameters of third-level types, and third-level type indicator system; data preprocessing: according to the correspondence between first-level and third-level types in the product system classification table, group the basic definitions and core parameters of third-level types and the third-level type indicator system; randomly shuffle the third-level type data under each first-level type to form 10 groups of data with different orders; integrate the above reference knowledge into structured prompts according to the prompt template. Corresponding to the first-level type. The collection of three types The process of integrating prompts is as follows, and the integration process uses... Randomly shuffle the order:
[0037] 2) Summary of Primary Type Concepts and Connotations. Input preprocessed prompts and ask questions of the large model, summarizing the definition, connotation, and scope of primary types from three dimensions: function, structure, and application. Ten sets of summary results are obtained for each primary type, used for subsequent random concatenation and self-consistency. Primary Types The summary process is as follows:
[0038] Step 5: Initial screening of three-level types based on classification tree constraints 1) Initial screening by level three types. Input the target device description. Combining the conceptual scope of the level one type, the classification tree structure of the type system conversion, and the complete indicator system of the level three types, prompts are organized according to the product system classification structure under each level one type. LLM questions are asked concurrently. A maximum of two level three types are recommended under each level one type. These types are then aggregated as candidates. This step fully utilizes the hierarchical constraints of the classification tree and the semantic information of the indicator system to narrow the search scope.
[0039] 2) Output Results. Output the initial set of three-level types. for:
[0040] Step 6: Precise Recommendation of Three-Level Types Based on Classification Tree Constraints This step aims to accurately determine the three-level type of component products through a multi-round thought chain reasoning mechanism combined with the deep understanding capabilities of a large language model. The specific process is as follows: 1) Input Data Preparation. Input data includes: publicly available data, product system classification table, basic definitions and core parameters of third-level types, and third-level type indicator system; Data preprocessing: Based on the correspondence between first-level and third-level types in the product system classification table, determine the range of third-level types for each product; obtain the basic definitions and core parameters of third-level types and the third-level type indicator system according to the range of third-level types; randomly shuffle the third-level type data under each group of first-level types to form 20 groups of data with different orders; integrate the above reference knowledge into structured prompts according to the prompt template. For products in the publicly available data... Public data description The results of the three-stage initial screening The three-level indicator system is as follows: The prompt word templates are integrated as follows. The integration process used... Randomly shuffle the order:
[0041] 2) Summary of Level 3 Type Scoring: Input the prompt words of the preprocessed data, ask questions to the large model, and obtain the level 3 type score for each product; each product obtains 20 sets of self-consistent scoring results; based on the scoring results, conduct a comprehensive analysis, select the three level 3 types with the highest scores, and construct the final judgment criteria.
[0042]
[0043] 3) Comprehensive determination of three-level types: Combine prompt words to ask the big model again; determine the final unique three-level type; output the confidence score.
[0044]
[0045] 4) Output results: Output the final three-level classification results and confidence level for each public data product.
[0046] 2. Result Verification: To evaluate the effectiveness and generalization ability of the proposed electronic component index system construction and classification tree constraint method based on LLM and clustering (referred to as the T_Parameters method in subsequent comparative analysis), this paper selects several classification methods based solely on product index classification systems as benchmarks and conducts systematic experimental verification on six datasets with different data completeness from five data sources. An example classification tree structure is shown below. If it is a multi-level structure, it is simplified to two levels by tiling for easier understanding: Ability classification tree structure (level 1 → level 3): ├─ Digital Integrated Circuits │├─ Central Processing Unit │├─ Digital Signal Processor │├─ Graphics Processor │... Analog integrated circuits │├─ A / D Converter │├─ D / A Converter │├─ Other Converters │... Additional explanation (easily confused three-level classifications), (Level 1 → Level 2 → Level 3 → all four levels included in Level 3): ├─ Hybrid Circuit → Power Converter → DC / DC Converter │└─ Detailed classification: │├─ Special DC / DC Converters │... ├─ Hybrid Circuits → ... ├─ RF devices → RF discrete devices → RF transistors │└─ Detailed classification: │├─ Radio Frequency Power Transistor │... All methods face the challenges of numerous three-level categories and long prompt words (approximately 25,000 characters). Stability is improved through multiple rounds of self-consistent voting with random shuffling. To verify the effectiveness of the classification tree constraint, a Parameters experiment was set up; to verify the granularity effect of the classification tree, an experiment with Trees_1 was set up; and to verify the effectiveness of the indicator system, an experiment with Trees_2 was configured. The configuration comparisons for each experiment are explained below: Table 1 Experimental Configuration List for Each Scheme
[0047] To verify the generality of the method, relevant fields were collected from five different data sources and categorized into device name, core indicator description, indirect indicator description, and historical classification information fields, forming six datasets with different levels of data completeness. The configurations are as follows: The experimental configurations for exp_1-6 in Table 2 are as follows:
[0048] The accuracy verification results are as follows: Since there are cases where the same third-level type labels exist, but belong to different first-level types, the complete accuracy rate is defined as the first-level type being correct and the third-level type being correct.
[0049] Table 3 shows the comparison results of each method on the exp_1-6 dataset. Dataset method Complete accuracy Complete recall rate Full F1 First-level type F1 Complete F1 difference exp_1 T_Parameters 0.9243 0.9126 0.9112 0.9617 Parameters 0.8514 0.7553 0.7718 0.8797 0.1394 Trees_1 0.8258 0.7592 0.7600 0.9208 0.1512 Trees_2 0.8638 0.8251 0.8216 0.9361 0.0896 exp_2 T_Parameters 0.9014 0.8908 0.8907 0.9477 Parameters 0.8422 0.7674 0.7809 0.8831 0.1098 Trees_1 0.8279 0.7645 0.7682 0.9114 0.1215 Trees_2 0.8623 0.8031 0.8136 0.9191 0.0771 exp_3 T_Parameters 0.9170 0.9051 0.9038 0.9567 Parameters 0.8562 0.7678 0.7795 0.8852 0.1243 Trees_1 0.8415 0.7698 0.7707 0.9317 0.1331 Trees_2 0.8668 0.8070 0.8161 0.9236 0.0877 exp_4 T_Parameters 0.9112 0.9033 0.9000 0.9632 Parameters 0.8535 0.7634 0.7722 0.8822 0.1278 Trees_1 0.8273 0.764 0.7651 0.9444 0.1349 Trees_2 0.8639 0.8171 0.8123 0.9513 0.0877 exp_5 T_Parameters 0.8929 0.8876 0.8820 0.9562 Parameters 0.8684 0.7876 0.794 0.8929 0.088 Trees_1 0.8391 0.7655 0.7667 0.9319 0.1153 Trees_2 0.8669 0.8292 0.8233 0.9529 0.0587 exp_6 T_Parameters 0.9131 0.9046 0.9014 0.9652 Parameters 0.8479 0.7701 0.7745 0.8838 0.1301 Trees_1 0.8173 0.7630 0.7590 0.9399 0.1456 Trees_2 0.8826 0.8459 0.8413 0.9532 0.0633 The T_Parameters method significantly outperforms in scenarios with mixed classification criteria and missing data. Its multi-level classification tree constraint and step-by-step inference mechanism effectively solve the problems of multiple three-level types and long prompt words, especially demonstrating outstanding generalization advantage when data is incomplete. The Trees_2 method only improves classification performance when data is complete, but it is still slightly lower than T_Parameters. Figure 2 As shown.
[0050] ①T_Parameters vs. Parameters (without classification tree constraints) In scenarios with missing data (exp_1 / exp_2), T_Parameters with complete F1 score leads Parameters by 13.94% (exp_1) and 11.98% (exp_2). This is because the classification tree constraint provides strong prior knowledge, which mitigates the impact of missing core indicators (e.g., exp_2 still maintains 90.07% accuracy even without core indicators).
[0051] ②T_Parameters vs. Trees_1 (Level 1 to Level 3 Trees) Refined constraints of the four-level classification tree: T_Parameters leads Trees_1 by an average of 113.2% in the full F1 score from exp_1 to exp_6 because the four-level tree provides a finer-grained index and reduces ambiguity in the three-level types (such as the same three-level label belonging to different first-level types).
[0052] ③T_Parameters vs. Trees_2 (Additional explanation of common mistakes at level 4) For types that are easily confused (such as optoelectronic devices, sensors and components), Trees_2 added "error-prone level 4 type descriptions" but did not significantly improve performance (exp_3 complete F1 81.61% vs T_Parameters 90.38%), because the long prompt words interfered with the model's focus on core indicators.
[0053] Because the quantity distribution of each primary category is uneven, the relevant indicator analysis indicators are pushed down to the primary category. The results on the exp_3 dataset, which has relatively good data completeness, are as follows: Figure 3 The T_Parameters method performs exceptionally well across the 11 primary types, excluding electromechanical components.
[0054] It dominates in categories with mixed classification criteria. In solid-state microwave devices and circuits (F1=0.88, leading by 34%↑), facing inconsistent classification criteria (mixed materials / functions) and easy confusion, T_Parameters unifies heterogeneous features through a four-level classification tree and performance parameter standardization (such as frequency range and power threshold), and reduces ambiguity through a step-by-step screening mechanism (first-level constraint → third-level candidate ≤2), with a recall rate (0.86) 56% higher than Trees_2 (0.55); in sensors and components (F1=0.92, leading by 16%↑), due to classification based on process principles, there are many marginal samples (such as special coating components), and T_Parameters has the advantage of using an index system to cover long-tail features (such as corrosion resistance and process complexity), with a recall rate (0.90) that crushes Trees_2 (0.73).
[0055] This specific embodiment is only used to illustrate the present invention and is not intended to limit the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the protection scope of the present invention.
Claims
1. A method for classifying electronic components based on LLM and classification trees, characterized in that, Includes the following steps: Step 1: Use a small number of expert-annotated samples to guide the large language model to perform an initial three-level coarse classification of the collected publicly available component description data; Step 2: Based on the preliminary classification results obtained in Step 1, clean the data using an attribute similarity clustering algorithm to output a balanced, high-quality three-level type dataset. Step 3: Based on the high-quality dataset, supplementary expert samples, and basic definitions of the three-level types obtained in Step 2, construct a structured performance index system covering all three-level types using a large language model; Step 4: Based on the product system classification table, the complete indicator system of the three-level types obtained in Step 3, and typical product examples, use the large language model to conduct inductive analysis from three dimensions: function, structure, and application, and abstract and summarize the conceptual scope of each primary type. Step 5: Input the target device description, integrate the classification tree structure of the type system conversion with the knowledge obtained in Step 3 and Step 4, perform preliminary screening of the three-level types under each level type and recommend up to two candidate three-level types; Step 6: For the candidate tertiary types obtained in Step 5, input their detailed definitions, core indicators, and lower-level classification tree structures. Perform deep indicator matching through a large language model, and combine self-consistency and voting verification mechanisms to determine the final unique tertiary type.
2. The method according to claim 1, characterized in that, The specific method of step 1 includes: traversing all three types, calculating the semantic similarity score between the data to be judged and the reference samples of each type, and making a comprehensive judgment based on the top three types with the highest scores to obtain preliminary classification labels.
3. The method according to claim 1, characterized in that, The specific method of step 2 includes: structuring the main performance indicators of data under the same third-level type into attribute-value pairs and encoding them into vectors, using the K-Means clustering method to group and verify the correctness of the cluster center type, and retaining records whose similarity to the center product attributes is greater than a preset threshold.
4. The method according to claim 3, characterized in that, The formula for calculating attribute similarity is: .
5. The method according to claim 1, characterized in that, In step 3, a step-by-step reasoning method is used to construct a structured performance index system, which involves sequentially learning concept definitions, identifying performance parameter items, eliminating redundant parameters, supplementing and verifying parameters, and verifying physical units.
6. The method according to claim 1, characterized in that, In step 4, the data and indicator system of the three-level types under each primary type are randomly sorted and multiple sets of prompt words are generated. The large language model is used to summarize from the dimensions of function, structure and application to obtain multiple sets of self-consistent results.
7. The method according to claim 1, characterized in that, The classification tree structure in step 5 includes a hierarchical relationship of first-level types, third-level types, and fourth-level types.
8. The method according to claim 1, characterized in that, In step 6, multiple sets of randomly ordered prompts are generated for each candidate type to obtain multiple rounds of scoring, and the final type is determined based on voting or comprehensive analysis mechanisms.
9. The method according to any one of claims 1 to 8, characterized in that, The large language model includes Gemma2-27B, Qwen3-32B, and DeepSeekR1 models, and is selectively deployed according to the computational requirements of different steps.
10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.