An automobile accessory name matching method, device and equipment based on a large model

Abstract

The application discloses a kind of based on big model's automobile accessory name matching method, device and equipment, comprising: receiving the original accessory description information input by user, and the original accessory description information is segmented, generates to be matched name;Based on the standard accessory knowledge base of pre-set, through the fine-tuning of big language model to the semantic understanding and standardization processing of to-be-matched name, obtain preliminary standardized accessory name and corresponding accessory category;The semantic similarity of preliminary standardized accessory name and accessory name in standard accessory knowledge base is calculated, according to semantic similarity and similarity threshold value, accessory name matching result is output, and standard accessory knowledge base is updated according to accessory name matching result, wherein, similarity threshold value is dynamically adjusted according to preliminary standardized accessory name and accessory category.Effectively improve the accuracy and efficiency of accessory name matching, and the self-adapting ability of system is enhanced by closed-loop learning.

Description

Technical Field

[0001] This invention relates to the field of automotive parts technology, and in particular to a method, apparatus, and equipment for matching automotive parts names based on a large model. Background Technology

[0002] In the digitalization of the automotive parts supply chain, achieving automatic and accurate matching of parts names from different sources is a crucial foundation for breaking down data silos and promoting information interconnection and business collaboration. However, this field has long faced the severe challenge of highly non-standardized parts naming. The same physical component can have numerous variations due to differences in regional customs, brand terminology, industry abbreviations, or colloquial descriptions, such as "brake pads / brake discs / brake shoes," "headlights / headlamps," and "ABS sensors / ABS." Furthermore, the names often suffer from missing directional terms (such as "left / right") or redundant modifiers (such as "light alloy rims"), collectively forming a massive, complex, and dynamically changing non-standard naming system that severely restricts the automated connection and integration of data across all links of the supply chain.

[0003] Currently, the industry primarily relies on keyword-based rule-based methods or general natural language processing models for name matching. However, rule-based methods perform precise or fuzzy matching through predefined rule bases, while general models depend on their semantic understanding capabilities to process text. Nevertheless, the unique complexity of name variations, the need for locative term processing, and the dynamic nature of terminology specific to the auto parts industry limit the accuracy and adaptability of existing methods in practical applications, thus impacting the efficiency of supply chain data integration. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a method, device and equipment for matching automotive parts names based on a large model, which aims to solve the problems of low matching accuracy, poor adaptability and lagging knowledge base updates in existing automotive parts name matching methods.

[0005] To achieve the above objectives, the present invention provides a method for matching automotive parts names based on a large model, the method comprising:

[0006] Receive the original accessory description information input by the user, segment the original accessory description information, and generate a name to be matched;

[0007] Based on a pre-set standard parts knowledge base, the name to be matched is semantically understood and standardized by a finely tuned large language model to obtain a preliminary standardized parts name and the corresponding parts category.

[0008] Calculate the semantic similarity between the preliminary standardized part name and the part name in the standard part knowledge base. Based on the semantic similarity and the similarity threshold, make a judgment, output the part name matching result, and update the standard part knowledge base according to the part name matching result. The similarity threshold is dynamically adjusted according to the preliminary standardized part name and the part category.

[0009] Preferably, the step of using a pre-set standard parts knowledge base to perform semantic understanding and standardization processing on the name to be matched through a fine-tuned large language model includes:

[0010] The name to be matched is matched with the standard parts names and aliases in the standard parts knowledge base;

[0011] If the name to be matched has only one more locative word than the standard part name or alias name, then the corresponding locative word is ignored, and the standard part name associated with the matched standard part name or alias name is taken as the preliminary standardized part name.

[0012] Preferably, the step of judging based on semantic similarity and a similarity threshold, and outputting the accessory name matching result, includes:

[0013] Calculate the first semantic similarity between the preliminary standardized part name and all alias names in the standard part knowledge base;

[0014] If the highest value in the first semantic similarity is greater than or equal to the similarity threshold, then the alias name corresponding to the first semantic similarity and the standard accessory name associated with the alias name are determined as the accessory name matching result;

[0015] If the highest value in the first semantic similarity is less than the similarity threshold, then the second semantic similarity between the preliminary standardized part name and all standard part names in the standard part knowledge base is calculated.

[0016] If the highest value of the second semantic similarity is greater than or equal to the similarity threshold, then the corresponding standard accessory name is determined as the accessory name matching result.

[0017] Preferably, the similarity threshold is dynamically adjusted based on the preliminary standardized accessory name and the accessory category, including:

[0018] A first adjustment factor is determined based on the text length of the preliminarily standardized accessory name, a second adjustment factor is determined based on the accessory category, and a third adjustment factor is determined based on historical matching feedback data.

[0019] The similarity threshold is obtained by calculating based on the first adjustment factor, the second adjustment factor, and the third adjustment factor.

[0020] Preferably, the step of calculating the similarity threshold based on the first adjustment factor, the second adjustment factor, and the third adjustment factor includes:

[0021] The similarity threshold is calculated using the formula θ = α * AdjustByLength(L) + β * AdjustByCategory(C) + γ * LearnFromFeedback(F). In this formula, θ represents the similarity threshold, L represents the text length of the initially standardized accessory name, C represents the accessory category, F represents the historical matching feedback data, α, β, and γ are the corresponding weight coefficients, AdjustByLength represents the first function for negative correlation adjustment based on the text length L, AdjustByCategory represents the second function for differential adjustment based on the accessory category C, and LearnFromFeedback represents the third function for online learning adjustment based on the historical matching feedback data F.

[0022] Preferably, updating the standard parts knowledge base based on the parts name matching results includes:

[0023] If the matching result of the accessory name is a standard accessory name, then check whether the name to be matched already exists in the alias name set of the standard accessory knowledge base;

[0024] If it does not exist, the name to be matched is added as a new alias name to the standard parts knowledge base, and the name to be matched is associated with the corresponding standard parts name;

[0025] If the matching result of the accessory name indicates a failure, the name to be matched is added to the standard accessory knowledge base and marked as pending manual review.

[0026] Preferably, the large language model is obtained by using low-rank adaptive technology for efficient parameter fine-tuning and progressive training using auto parts domain name data. The progressive training using auto parts domain name data includes initial training using standard parts names with high confidence, and incremental training using variant names containing multiple non-standard parts names.

[0027] To achieve the above objectives, the present invention also provides a vehicle parts name matching device based on a large model, the device comprising:

[0028] The input unit is used to receive the original accessory description information input by the user, and to segment the original accessory description information to generate a name to be matched;

[0029] The preprocessing unit is used to perform semantic understanding and standardization on the name to be matched based on a preset standard parts knowledge base and a finely tuned large language model to obtain a preliminary standardized parts name and the corresponding parts category.

[0030] The matching unit is used to calculate the semantic similarity between the preliminary standardized part name and the part name in the standard part knowledge base, make a judgment based on the semantic similarity and the similarity threshold, output the part name matching result, and update the standard part knowledge base based on the part name matching result. The similarity threshold is dynamically adjusted according to the preliminary standardized part name and the part category.

[0031] To achieve the above objectives, the present invention also proposes a large-model-based automotive parts name matching device, including a processor, a memory, and a computer program stored in the memory. The computer program is executed by the processor to implement the steps of the large-model-based automotive parts name matching method as described in the above embodiments.

[0032] To achieve the above objectives, the present invention also proposes a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of a large-model-based automotive parts name matching method as described in the above embodiments.

[0033] Beneficial effects:

[0034] The above solution improves the quality of initial queries by using a domain-fine-tuned large language model for deep semantic understanding, introduces dynamically adjusted similarity thresholds to enhance the adaptability and accuracy of matching decisions, and employs an automated knowledge base update mechanism based on matching results. This effectively improves the accuracy and efficiency of accessory name matching. At the same time, the closed-loop learning mechanism enhances the system's adaptability, overcomes the limitations of traditional methods that rely on manual rules and have poor fault tolerance, and effectively reduces manual maintenance costs.

[0035] By matching the name to be matched with the standard part name and aliases in the knowledge base, and ignoring the directional words and directly returning the standard part name when only directional words are extra, intelligent fault tolerance processing of directional word redundancy is achieved. This significantly improves the system's ability to parse common variant names, reduces matching failures caused by inconsistent directional words, and enhances the practicality and robustness of the method.

[0036] By first calculating the semantic similarity with aliases and then with standard part names, and then making tiered decisions based on thresholds, the comprehensiveness and accuracy of the matching process are ensured. This two-layer matching strategy expands the coverage of non-standard names and improves the recognition rate and accuracy of complex variants.

[0037] By comprehensively considering multiple factors such as text length, accessory category, and historical feedback data, and using a formulaic weighted function for fusion calculation, the similarity threshold can be intelligently adjusted according to specific scenarios and real-time feedback, achieving adaptive optimization of matching conditions. This dynamic adaptive threshold mechanism overcomes the shortcomings of fixed thresholds, making matching decisions more flexible and accurate, improving adaptability to texts of different lengths and accessories of different complexities. At the same time, through online learning, the decision performance is continuously improved, significantly enhancing the robustness and overall accuracy of the system in scenarios of different complexities.

[0038] The knowledge base is automatically updated based on the matching results. When a match is successful, a new alias is added to associate the standard part name. When a match fails, it is marked as pending review, forming a closed-loop learning cycle. This enables the knowledge base to be automatically and optimized in real time. At the same time, it efficiently guides difficult cases that cannot be handled to the manual review process, forming a human-machine collaborative optimization loop. This improves the system's ability to quickly adapt to new terms, thereby maintaining a high matching accuracy rate in the long term.

[0039] By employing efficient parameter fine-tuning techniques (such as LoRA) and combining them with a progressive training strategy from easy to difficult, a general-purpose large language model can be efficiently transformed into a specialized model for deep understanding of terminology variations in the auto parts field under limited computing power. This ensures that the model has strong domain semantic understanding and generalization capabilities, providing a high-quality semantic parsing foundation for the overall matching process. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart illustrating a method for matching the names of automotive parts based on a large model, as provided in an embodiment of the present invention.

[0042] Figure 2 This is a schematic diagram of the overall process for matching automotive parts names according to an embodiment of the present invention.

[0043] Figure 3This is a schematic diagram of a large-model-based automotive parts name matching device provided in an embodiment of the present invention.

[0044] The realization of the invention's objective, its functional characteristics, and advantages will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] The present invention will be described in detail below with reference to the embodiments.

[0047] Reference Figure 1 The diagram shown is a flowchart of a method for matching the names of automotive parts based on a large model, provided by an embodiment of the present invention.

[0048] In this embodiment, the method is implemented based on an automotive parts name matching system; wherein, the system includes:

[0049] The parts translation agent receives and segments the original parts description information to obtain the name to be matched. Based on the fine-tuned large language model and combined with the standard parts knowledge base, the name to be matched is semantically understood and standardized, and the preliminary standardized parts name and parts category are output.

[0050] The accessory matching agent is connected to the accessory translation agent. It is used to calculate a similarity threshold based on the dynamic features of the preliminary standardized accessory name, and use the similarity threshold as a decision condition to perform semantic matching between the preliminary standardized accessory name and the standard accessory knowledge base, and output the accessory name matching result.

[0051] The updated and optimized intelligent agent is connected to the accessory matching intelligent agent and used to perform an update operation on the standard accessory knowledge base based on the accessory name matching result.

[0052] In this embodiment, the method includes:

[0053] S11, Receive the original accessory description information input by the user, and segment the original accessory description information to generate a name to be matched.

[0054] The step of segmenting the original accessory description information to generate a name to be matched includes:

[0055] Identify the first space character in the original accessory description information, define the string portion before the first space character as the name to be matched, and define the string portion after the first space character as supplementary description information.

[0056] Reference Figure 2 As shown. In this embodiment, the system receives raw part description information input by the user through the operation interface or API interface. This raw part description information is usually non-standardized text directly obtained or filled in by the user from the business scenario, and it follows the common format of "part core name supplemented with description information". For example, the user may input "left front headlight assembly with LED daytime running lights" or "ZF-802-3T4G1 transmission oil filter". In order to perform structured parsing of the input content, it is segmented. That is, by locating the position index of the first space character in the entire input string, the entire character sequence (substring) before the space is extracted and defined as the "name to be matched". This part is the core processing object of all subsequent semantic understanding and standardized matching processes; at the same time, the entire character sequence after the space is defined as "supplementary description information". In this way, the system can accurately extract the key text required for subsequent deep processing from unstructured input in the early stages without the need for complex natural language word segmentation. This lays a clear input foundation for the semantic parsing of the "parts translation agent" while retaining supplementary descriptive information for the model to refer to the context when needed, thus balancing the dual requirements of processing efficiency and semantic integrity.

[0057] S12, based on the preset standard parts knowledge base, the name to be matched is semantically understood and standardized by a finely tuned large language model to obtain the preliminary standardized parts name and the corresponding parts category.

[0058] Furthermore, in step S12, the semantic understanding and standardization processing of the name to be matched is performed on the name based on the preset standard parts knowledge base through a fine-tuned large language model, including:

[0059] S12-1, Match the name to be matched with the standard parts name and alias name in the standard parts knowledge base;

[0060] S12-2, if the name to be matched has only one more locative word than the standard part name or alias name, then the corresponding locative word is ignored, and the standard part name associated with the matched standard part name or alias name is taken as the preliminary standardized part name.

[0061] In this embodiment, semantic calibration is performed using a pre-defined standard parts knowledge base (KB_std). The segmented names to be matched are then input into a large language model (LLM) fine-tuned with automotive parts corpus for deep semantic understanding and preliminary standardization. During training, this model has absorbed a large amount of semantic knowledge from professional texts such as repair manuals and parts catalogs, enabling it to recognize colloquial and regional expressions that may exist in user input and accurately capture core parts concepts.

[0062] Specifically, the system first searches for the semantically closest existing standard part name (oeName) or its alias (alias_name) in KB_std for the name to be matched. During this process, the system incorporates a crucial intelligent fault-tolerance rule: if the model analysis reveals that the input name to be matched definitively adds a valid directional word (such as left, right, front, back, etc.) to a known name (oeName or alias_name) in KB_std, while other core parts remain the same, then the redundant directional word is ignored, and the known name (or its associated oeName if it's an alias) is directly output as the preliminary standardized part name (std_name). For example, for the input "left front door," if "front door assembly" exists in the knowledge base, the model will ignore the directional word "left" and directly output "front door assembly" as std_name. If no such direct match exists in the knowledge base, the model relies on its own domain knowledge reserves to generate a std_name that conforms to industry standards, removes redundant modifications, and retains core concepts and necessary directional words. Meanwhile, during this deep semantic parsing process, the model will naturally identify the accessory category to which the accessory belongs (such as "braking system" or "lighting system"), and output this accessory category as an associated attribute of std_name, providing key context for subsequent dynamic threshold matching.

[0063] Furthermore, the construction process of the standard parts knowledge base includes:

[0064] Obtain raw parts data from multiple authoritative data sources, including original parts catalogs from vehicle manufacturers and industry standard databases;

[0065] The original component data is integrated and standardized, and structured standard component names are generated according to preset naming rules;

[0066] Obtain multiple alias names that match the standard part name, and associate the multiple alias names with the standard part name to form initial knowledge base entries;

[0067] The initial knowledge base entries are verified to obtain the standard parts knowledge base.

[0068] In this embodiment, the construction of the standard parts knowledge base (KB_std) is a system "cold start" initialization process, designed to provide a high-quality, structured initial data foundation for subsequent intelligent matching. The first step involves the collection and integration of authoritative data from multiple sources. Data sources primarily include original equipment manufacturer (OE) catalogs, industry-standardized databases (such as ACES / PIES), and product data from mainstream auto parts platforms, ensuring the authority and broad coverage of the knowledge base. Subsequently, by integrating this data, the integrated raw data undergoes deep standardization and structuring. Based on clear naming conventions (e.g., "core part name + locative term + vehicle model constraint"), standardized standard parts names (oeName) are generated from the OE catalog. Simultaneously, natural language processing and data mining techniques are used to automatically extract various synonyms, abbreviations, and regional names corresponding to the standard parts name from market data, defining them as aliases (alias_name). Each alias name is associated with a specific oeName, forming an "oeName–alias_name" mapping pair, constituting the initial knowledge base entry.

[0069] To ensure the quality and reliability of the initial knowledge base, a verification process involving both automated testing and human expert review is crucial. The system uses historically accumulated, validated "non-standard name - standard name" pairing data as a test set to conduct offline batch testing of the knowledge base's matching logic, evaluating its initial accuracy. For results with low confidence in automated matching, and for high-frequency key terms in the business, the system marks them and submits them to domain experts for manual review. Experts correct erroneous mappings and supplement missing associations, thereby optimizing the knowledge base. This pre-populated and rigorously validated KB_std is used as the core reference library upon system deployment, and the system's update and optimization agent will be responsible for its subsequent continuous dynamic optimization and updates based on this.

[0070] S13, calculate the semantic similarity between the preliminary standardized part name and the part name in the standard part knowledge base, judge based on the semantic similarity and similarity threshold, output the part name matching result, and update the standard part knowledge base based on the part name matching result, wherein the similarity threshold is dynamically adjusted according to the preliminary standardized part name and the part category.

[0071] Furthermore, in step S13, the step of judging based on semantic similarity and a similarity threshold, and outputting the accessory name matching result, includes:

[0072] S13-1, Calculate the first semantic similarity between the preliminary standardized part name and all alias names in the standard part knowledge base;

[0073] S13-2, if the highest value in the first semantic similarity is greater than or equal to the similarity threshold, then the alias name corresponding to the first semantic similarity and the standard accessory name associated with the alias name are determined as the accessory name matching result;

[0074] S13-3, If the highest value in the first semantic similarity is less than the similarity threshold, then calculate the second semantic similarity between the preliminary standardized part name and all standard part names in the standard part knowledge base;

[0075] S13-4, if the highest value of the second semantic similarity is greater than or equal to the similarity threshold, then the corresponding standard accessory name is determined as the accessory name matching result.

[0076] Furthermore, when the accessory name matching result indicates a candidate standard accessory name, if the preliminary standardized accessory name lacks a locative word, and the standard accessory knowledge base contains multiple standard accessory names that have the same core part as the candidate standard accessory name but differ only in locative words, then based on the historical frequency of the standard accessory name pre-statistically recorded in the standard accessory knowledge base, the standard accessory name with the highest historical frequency is taken as the accessory name matching result.

[0077] Furthermore, the similarity threshold is dynamically adjusted based on the initially standardized accessory name and the accessory category, including:

[0078] A first adjustment factor is determined based on the text length of the preliminarily standardized accessory name, a second adjustment factor is determined based on the accessory category, and a third adjustment factor is determined based on historical matching feedback data.

[0079] The similarity threshold is calculated based on the first adjustment factor, the second adjustment factor, and the third adjustment factor; including:

[0080] The similarity threshold is calculated using the formula θ = α * AdjustByLength(L) + β * AdjustByCategory(C) + γ * LearnFromFeedback(F). In this formula, θ represents the similarity threshold, L represents the text length of the initially standardized accessory name, C represents the accessory category, F represents the historical matching feedback data, α, β, and γ are the corresponding weight coefficients, AdjustByLength represents the first function for negative correlation adjustment based on the text length L, AdjustByCategory represents the second function for differential adjustment based on the accessory category C, and LearnFromFeedback represents the third function for online learning adjustment based on the historical matching feedback data F.

[0081] In this embodiment, the preliminary standardized part name (std_name) and its corresponding part category output are input into the part matching agent for precise judgment to determine whether it is the unique standard part name corresponding to it in the standard part knowledge base (KB_std) or whether the match fails. Specifically, a similarity threshold (θ) is dynamically calculated. This threshold is not a fixed value, but an intelligent judgment threshold generated in real time. Its calculation comprehensively considers information from multiple dimensions, including: adaptive adjustment based on the text length (L) of std_name, using a stricter threshold for short texts (such as part numbers) to ensure accuracy, and moderately relaxing it for long texts to improve recall; calling a differentiated baseline threshold based on the part category (C) identified in the previous step to address the differences in the complexity of part name variants in different categories; and introducing an online learning factor based on historical user feedback data (F). The system continuously collects confirmation and correction information of matching results. If the recent accuracy is high, the threshold is finely adjusted to try to cover more variants, and vice versa to ensure accuracy. The similarity threshold is calculated using the formula θ=α*AdjustByLength(L) +β*AdjustByCategory(C) +γ*LearnFromFeedback(F).

[0082] After obtaining the similarity threshold θ, the system performs two levels of semantic matching and threshold judgment, including: calculating the cosine similarity of the semantic vectors (e.g., generated based on the BERT model) of std_name and all alias names (alias_name) in KB_std, and obtaining the highest score max_score; if max_score ≥ θ, then the standard part name (oeName) associated with the alias name is determined as a candidate result; if the first level of matching fails (max_score < θ), then the second level of matching is started: directly calculating the semantic similarity of std_name and all standard part names (oeName) in KB_std, and if its highest score max_score_oe ≥ θ, then the oeName is determined as a candidate result; otherwise, the matching is judged to have failed, and a null value is output as the output result.

[0083] After obtaining the candidate oeName, the system executes the intelligent directional word default decision. That is, if the input std_name does not contain an explicit directional word, but there are multiple standard part names in KB_std that are the same as the core part of the candidate oeName but different only in directional words (such as "left headlight" and "right headlight"), then based on the historical frequency data of each directional variant in all relevant aliases and standard part names preprocessed in KB_std, the standard part name with the highest historical frequency is taken as the output result. That is, the most frequent directional word is added to form the final output standard part name matching result with a clear direction.

[0084] Furthermore, updating the standard parts knowledge base based on the parts name matching results includes:

[0085] If the matching result of the accessory name is a standard accessory name, then check whether the name to be matched already exists in the alias name set of the standard accessory knowledge base;

[0086] If it does not exist, the name to be matched is added as a new alias name to the standard parts knowledge base, and the name to be matched is associated with the corresponding standard parts name;

[0087] If the matching result of the accessory name indicates a failure, the name to be matched is added to the standard accessory knowledge base and marked as pending manual review.

[0088] In this embodiment, after the matching is completed, the intelligent agent performs operations based on the final result by updating and optimizing. If the matching is successful, it checks whether the name to be matched is already an alias name under the oeName (or other oeName). If not, it automatically adds it as a new alias_name to KB_std and associates it with the oeName that was successfully matched this time, thus realizing the accumulation of knowledge. If the matching fails (the result is null), the system stores the name to be matched as a new alias_name in KB_std, but sets the associated oeName field to null and marks it as "awaiting manual review" status, thereby recording new variants that the system cannot recognize, driving the continuous expansion and optimization of the knowledge base, and forming a closed-loop learning.

[0089] Furthermore, the large language model is obtained by using low-rank adaptive technology for efficient parameter fine-tuning and progressive training using auto parts domain name data. The progressive training using auto parts domain name data includes initial training using standard parts names with high confidence, and incremental training using variant names containing multiple non-standard parts names.

[0090] In this embodiment, a general, large-scale pre-trained language model (such as a Transformer-based model) is selected as the base. To achieve efficient and targeted domain adaptation, low-rank adaptive technique (LoRA) is used for efficient parameter fine-tuning. Specifically, most (e.g., 90%) of the original parameters of the base model are frozen, and only the attention layers crucial for semantic understanding (e.g., the projection matrices of query q_proj and value v_proj) are injected with trainable low-rank matrices (B and A). During forward propagation, the original weight matrix W is added to the low-rank increment ΔW = α·B·A (where α is the scaling factor) (i.e., W' = W + ΔW). This allows the model to quickly learn and internalize the unique semantic patterns of the auto parts domain with a very small number of trainable parameters (only about 0.1%-1% of the original model), while preserving the original general language understanding capabilities of the base model to the greatest extent possible, significantly reducing the computational resources and memory consumption required for training.

[0091] To further improve the model's robustness and accuracy when handling complex variations in auto parts names, this embodiment employs a progressive training strategy. The core training data includes pairings of non-standard part names with their corresponding standard part names. Training is divided into two closely linked phases:

[0092] Phase 1: Easy Sample Training; High-confidence, easily matched samples are selected from the training data to form the first training set. These samples typically meet one or more of the following characteristics: the non-standard part names are highly similar to standard part names (e.g., differing only in punctuation or stop words), can be directly mapped using predefined simple keyword rules, or are industry-recognized and unambiguous basic terms. This first training set is used to initially train the model configured with the LoRA adapter. The goal of this phase is to enable the model to quickly and stably establish a basic understanding of core terms and fundamental mapping relationships in the automotive parts field, forming a well-performing initial version.

[0093] Phase Two: Incremental Learning on Difficult Samples; After the model completes training on easy samples, a more challenging second training set is introduced for incremental training. This second training set contains complex variations of various non-standard accessory names, such as local colloquialisms (e.g., "fire nozzle"), industry abbreviations (e.g., "ABS"), names lacking directional terms (e.g., only "car door"), and long texts containing redundant modifiers. These samples are the key challenges that cause traditional matching methods to fail. By allowing the model to continuously learn and optimize its handling of these difficult samples in this phase, its semantic understanding and normalization capabilities for complex and irregular inputs in real-world business scenarios are significantly enhanced. The stopping condition for the entire training process is set to ensure that the fluctuation of key performance indicators (e.g., F1-score) on the independent validation set is less than a preset threshold (e.g., 0.5%), to ensure that the model fully converges and its performance is stable.

[0094] By combining the above-mentioned "LoRA efficient fine-tuning" with "progressive training from easy to difficult", the final domain-specific large language model not only possesses profound automotive parts expertise, but also has excellent generalization and fault tolerance capabilities, thus providing a powerful and reliable intelligent core for the "translation" stage of the entire name matching process.

[0095] Reference Figure 3 The diagram shown is a structural schematic of an automotive parts name matching device based on a large model, according to an embodiment of the present invention.

[0096] In this embodiment, the device 20 includes:

[0097] Input unit 21 is used to receive the original accessory description information input by the user, and to segment the original accessory description information to generate a name to be matched;

[0098] Preprocessing unit 22 is used to perform semantic understanding and standardization processing on the name to be matched based on a preset standard parts knowledge base and a fine-tuned large language model to obtain a preliminary standardized parts name and the corresponding parts category.

[0099] The matching unit 23 is used to calculate the semantic similarity between the preliminary standardized part name and the part name in the standard part knowledge base, make a judgment based on the semantic similarity and the similarity threshold, output the part name matching result, and update the standard part knowledge base based on the part name matching result. The similarity threshold is dynamically adjusted according to the preliminary standardized part name and the part category.

[0100] Each unit module of the device 20 can execute the corresponding steps in the above method embodiment, so the details of each unit module will not be elaborated here. Please refer to the description of the corresponding steps above for details.

[0101] This invention also provides a large-model-based automotive parts name matching device, which includes the large-model-based automotive parts name matching apparatus described above. The large-model-based automotive parts name matching apparatus can employ... Figure 3 The structure of the embodiment, correspondingly, can be executed Figure 1 The technical solutions of the method embodiments shown are similar in implementation principle and technical effect. For details, please refer to the relevant records in the above embodiments, which will not be repeated here.

[0102] The device includes: a mobile phone, digital camera, or tablet computer, or other device with a camera function; or a device with an image processing function; or a device with an image display function. The device may include components such as a memory, processor, input unit, display unit, and power supply.

[0103] The memory can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, application programs required for at least one function (such as image playback function), etc.; the data storage area can store data created according to the use of the device. In addition, the memory can include high-speed random access memory, and can also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory can also include a memory controller to provide access to the memory for the processor and input units.

[0104] The input unit can be used to receive input numerical, character, or image information, and to generate keyboard, mouse, joystick, optical, or trackball signal inputs related to user settings and function control. Specifically, in addition to a camera, the input unit of this embodiment may also include a touch-sensitive surface (e.g., a touch screen) and other input devices.

[0105] The display unit can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces of the device. These graphical user interfaces can be composed of graphics, text, icons, video, and any combination thereof. The display unit may include a display panel, optionally configured as an LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), or other similar display panel. Furthermore, a touch-sensitive surface may cover the display panel. When the touch-sensitive surface detects a touch operation on or near it, it transmits the information to the processor to determine the type of touch event. Subsequently, the processor provides corresponding visual output on the display panel based on the type of touch event.

[0106] This invention also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the memory described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement... Figure 1 The method shown is a large-model-based automotive parts name matching method. The computer-readable storage medium can be a read-only memory, a disk, or an optical disk, etc.

[0107] This invention also provides a computer program product, including a computer program / instructions, which are loaded and executed by a processor to achieve [the desired result]. Figure 1 This illustrates a method for matching the names of automotive parts based on a large model.

[0108] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, device embodiments, and storage medium embodiments, since they are basically similar to the method embodiments, the descriptions are relatively simple, and relevant parts can be referred to the descriptions in the method embodiments.

[0109] Furthermore, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0110] The foregoing description illustrates and describes preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for matching the names of automotive parts based on a large model, characterized in that, The method includes: Receive the original accessory description information input by the user, segment the original accessory description information, and generate a name to be matched; Based on a pre-set standard parts knowledge base, the name to be matched is semantically understood and standardized by a finely tuned large language model to obtain a preliminary standardized parts name and the corresponding parts category. Calculate the semantic similarity between the preliminary standardized part name and the part names in the standard part knowledge base. Based on the semantic similarity and a similarity threshold, determine the part name matching result and output it. Update the standard part knowledge base based on the part name matching result. The similarity threshold is dynamically adjusted based on the preliminary standardized part name and the part category. This includes: A first adjustment factor is determined based on the text length of the preliminarily standardized accessory name, a second adjustment factor is determined based on the accessory category, and a third adjustment factor is determined based on historical matching feedback data. The similarity threshold is calculated based on the first adjustment factor, the second adjustment factor, and the third adjustment factor, including: Through the formula θ=α AdjustByLength(L) +β AdjustByCategory(C) +γ The similarity threshold is obtained by calculating LearnFromFeedback(F); where θ represents the similarity threshold, L represents the text length of the initially standardized accessory name, C represents the accessory category, F represents the historical matching feedback data, α, β, and γ are the corresponding weight coefficients, AdjustByLength represents the first function for negative correlation adjustment based on text length L, AdjustByCategory represents the second function for differential adjustment based on accessory category C, and LearnFromFeedback represents the third function for online learning adjustment based on historical matching feedback data F. The step of judging based on semantic similarity and similarity threshold, and outputting the accessory name matching result, includes: Calculate the first semantic similarity between the preliminary standardized part name and all alias names in the standard part knowledge base; If the highest value in the first semantic similarity is greater than or equal to the similarity threshold, then the alias name corresponding to the first semantic similarity and the standard accessory name associated with the alias name are determined as the accessory name matching result; If the highest value in the first semantic similarity is less than the similarity threshold, then the second semantic similarity between the preliminary standardized part name and all standard part names in the standard part knowledge base is calculated. If the highest value of the second semantic similarity is greater than or equal to the similarity threshold, then the corresponding standard accessory name is determined as the accessory name matching result.

2. The method for matching automotive parts names based on a large model according to claim 1, characterized in that, The pre-defined standard parts knowledge base performs semantic understanding and standardization processing on the names to be matched using a fine-tuned large language model, including: The name to be matched is matched with the standard parts names and aliases in the standard parts knowledge base; If the name to be matched has only one more locative word than the standard part name or alias name, then the corresponding locative word is ignored, and the standard part name associated with the matched standard part name or alias name is taken as the preliminary standardized part name.

3. The method for matching automotive parts names based on a large model according to claim 1, characterized in that, The step of updating the standard parts knowledge base based on the parts name matching results includes: If the matching result of the accessory name is a standard accessory name, then check whether the name to be matched already exists in the alias name set of the standard accessory knowledge base; If it does not exist, the name to be matched is added as a new alias name to the standard parts knowledge base, and the name to be matched is associated with the corresponding standard parts name; If the matching result of the accessory name indicates a failure, the name to be matched is added to the standard accessory knowledge base and marked as pending manual review.

4. The method for matching automotive parts names based on a large model according to claim 1, characterized in that, The large language model is obtained by using low-rank adaptive technology for efficient parameter fine-tuning and progressive training using auto parts domain name data. The progressive training using auto parts domain name data includes initial training using standard parts names with high confidence, and incremental training using variant names that include multiple non-standard parts names.

5. A device for matching the names of automotive parts based on a large model, characterized in that, The apparatus is used to perform a large-model-based automotive parts name matching method as described in any one of claims 1 to 4, the apparatus comprising: The input unit is used to receive the original accessory description information input by the user, and to segment the original accessory description information to generate a name to be matched; The preprocessing unit is used to perform semantic understanding and standardization on the name to be matched based on a preset standard parts knowledge base and a finely tuned large language model to obtain a preliminary standardized parts name and the corresponding parts category. A matching unit is used to calculate the semantic similarity between the preliminary standardized part name and the part names in the standard part knowledge base, judge based on the semantic similarity and a similarity threshold, output the part name matching result, and update the standard part knowledge base based on the part name matching result. The similarity threshold is dynamically adjusted based on the preliminary standardized part name and the part category. This includes: A first adjustment factor is determined based on the text length of the preliminarily standardized accessory name, a second adjustment factor is determined based on the accessory category, and a third adjustment factor is determined based on historical matching feedback data. The similarity threshold is obtained by calculating based on the first adjustment factor, the second adjustment factor, and the third adjustment factor.

6. A vehicle parts name matching device based on a large model, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory, the computer program being executed by the processor to implement the steps of a large-model-based automotive parts name matching method as described in any one of claims 1 to 4.

7. A computer program product, characterized in that, Includes a computer program / instructions, which, when executed by a processor, implement the steps of a large-model-based automotive parts name matching method as described in any one of claims 1 to 4.

Images