A method and system for identifying entities in the automotive field based on a large language model

Abstract

The application provides a kind of automobile field entity recognition method and system based on large language model, the method extracts candidate entity in the text to be processed by first identification model, and the entity that cannot be mapped to standard knowledge base is marked as pending entity;When there is pending entity, retrieve relevant documents from the automobile field dynamic text library as supplementary knowledge, and conduct secondary entity recognition combined with large language model to output the final recognition result;The application adopts a hybrid knowledge base architecture, combines structured database and entity vector library to realize dual-channel retrieval, and supports dynamic incremental update based on the recognition result.The application effectively solves the problem of missed recognition caused by the strong timeliness and high ambiguity of professional terms in the automobile field, significantly improves the accuracy and recall rate of entity recognition, and at the same time, through the dynamic knowledge updating mechanism, the system's ability to quickly adapt to new car series and new brand release is guaranteed, which is suitable for application scenarios such as automobile vertical search and intelligent question answering.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and system for entity recognition in the automotive field based on a large language model. Background Technology

[0002] With the rapid development of the automotive industry, user search queries in automotive vertical applications are characterized by high specialization, diverse entity types, and high timeliness requirements. Accurately identifying key entities such as car brands, models, and series in the search text is crucial for understanding user intent and improving the relevance of search results.

[0003] Currently, entity recognition technologies mainly include the following categories:

[0004] Traditional machine learning methods employ statistical models such as support vector machines and conditional random fields, relying on manually designed bag-of-words and part-of-speech features for entity classification. These methods require extensive involvement from domain experts in feature engineering, struggle to adapt to the rapidly emerging new car models and terminology in the automotive industry, and have limited ability to understand complex contexts.

[0005] Deep learning-based methods utilize convolutional neural networks, recurrent neural networks, or pre-trained language models for end-to-end entity recognition. While this reduces the burden of feature engineering, deep learning models typically require large-scale, high-quality labeled data, which is prohibitively expensive to acquire in the automotive field. Furthermore, pre-trained models, trained on general corpora, lack sufficient coverage of automotive terminology, leading to issues such as missed recognition of new car models and failure to identify aliases.

[0006] Static knowledge base-based methods: Some solutions introduce knowledge graphs or ontology bases to assist in recognition, but existing knowledge bases have long construction cycles and are updated late, failing to cover newly released models and emerging brands in a timely manner. Furthermore, static knowledge bases struggle to handle colloquial expressions, abbreviations, and context-dependent ambiguities in user search queries, resulting in limited recall rates.

[0007] Limitations of single-model inference: Existing technologies mostly employ a single recognition process, lacking verification and compensation mechanisms for low-confidence recognition results. When the model encounters unfamiliar car model names or emerging technical terms, it lacks the ability to correct errors, directly impacting the user experience of downstream search, recommendation, and other services.

[0008] In summary, existing technologies face multiple challenges in entity recognition tasks in the automotive field, such as poor timeliness of technical terms, weak ambiguity resolution capabilities, lack of compensation mechanisms for missed identifications, and high knowledge base maintenance costs. There is an urgent need for an efficient entity recognition solution that can integrate dynamic knowledge sources and has self-correction capabilities. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for entity recognition in the automotive field based on a large language model, thereby solving the aforementioned problems existing in the prior art.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A method for entity recognition in the automotive domain based on a large language model includes the following steps:

[0012] Receive the text to be processed, which contains named entities in the automotive domain;

[0013] Candidate entities are extracted from the text to be processed using the first recognition model;

[0014] Candidate entities are mapped and matched with standard knowledge bases in the automotive field, and candidate entities that cannot be successfully mapped are marked as pending entities;

[0015] When there are entities to be identified, relevant documents are retrieved from a dynamic text library in the automotive field as supplementary knowledge based on the text to be processed.

[0016] The entity to be identified and supplementary knowledge are input into the large language model. The large language model performs secondary entity recognition to determine the standard name of the entity to be identified and outputs the final entity recognition result.

[0017] Preferably, the standard knowledge base in the automotive field is a hybrid knowledge base system, including:

[0018] A structured database used to store standard names, synonyms, and hierarchical relationships for car brands, car series, and car models;

[0019] Entity vector library, used to store semantic representation vectors of standard names and synonyms.

[0020] Preferably, the relevant documents retrieved include:

[0021] Based on the entity name of the entity to be determined, perform an exact match query in the structured database to obtain the first query result;

[0022] Based on the contextual semantic features of the entity to be processed in the text, a similarity search is performed in the entity vector database to obtain the second query result.

[0023] Combine the results of the first and second queries to generate supplementary knowledge.

[0024] Preferably, the dynamic text library in the automotive field is a real-time updated database of automotive news articles, and the database is constructed in the following way:

[0025] We extract articles from automotive vertical media platforms that include new car launches, model reviews, and technical analyses.

[0026] Incremental indexing of articles and assigning retrieval weights based on publication time;

[0027] The hybrid knowledge base system is dynamically updated in the following ways:

[0028] Obtain the newly added entities from the secondary recognition output of the large language model;

[0029] When the semantic similarity between a newly added entity and an existing entity in the knowledge base is greater than a preset threshold, the newly added entity will be added to the knowledge base system as a synonym.

[0030] Monitor external automotive data sources, and automatically extract entities and update the knowledge base system when a new model or series of vehicles is detected.

[0031] Preferably, marking candidate entities that cannot be successfully mapped as pending entities includes:

[0032] Calculate the semantic similarity between candidate entities and standard names in the standard knowledge base;

[0033] When the semantic similarity is lower than the dynamic threshold, the mapping is deemed to have failed.

[0034] Among them, the dynamic threshold is adaptively adjusted based on the recognition confidence of candidate entities and the historical mapping success rate;

[0035] Preferably, inputting the undetermined entity and supplementary knowledge into the large language model includes:

[0036] Construct preset prompt words, which include the entity to be determined, the context information of the text to be processed, and the list of candidate entities in the supplementary knowledge;

[0037] Preset prompts are used to guide the large language model to perform disambiguation recognition of the target entity based on contextual information;

[0038] Preset prompts are also used to guide the large language model to output standardized entity representations, which include standard entity names, entity type codes, and confidence scores.

[0039] Preferred options also include:

[0040] The number of entities whose standard names cannot be determined after secondary entity recognition is counted;

[0041] When the number exceeds the tolerance threshold, the corresponding text to be processed and the recognition result are recorded in the manual review queue for optimization of the first recognition model and the hybrid knowledge base system.

[0042] Furthermore, an automotive-domain entity recognition system based on a large language model includes:

[0043] The initial recognition module is used to extract candidate entities from the text to be processed through the first recognition model;

[0044] The mapping and verification module is used to map and match candidate entities with the standard knowledge base in the automotive field, and mark candidate entities that cannot be mapped as pending entities;

[0045] The retrieval enhancement module is used to retrieve relevant documents from a dynamic text library in the automotive field as supplementary knowledge when there is a pending entity.

[0046] The secondary recognition module is used to input the entity to be identified and supplementary knowledge into the large language model for secondary entity recognition and output the final entity recognition result.

[0047] The dynamic update module is used to automatically update the standard knowledge base based on the recognition results;

[0048] The system executes the methods described above.

[0049] Furthermore, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0050] Furthermore, a non-transitory computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the above-described method.

[0051] The beneficial effects of this invention are:

[0052] This invention achieves the following beneficial effects by constructing a closed-loop process of "initial identification - mapping verification - retrieval enhancement - secondary identification - dynamic update":

[0053] Effectively solves the problem of missed identification: For pending entities that the initial identification model cannot map to the standard knowledge base, an automatic retrieval enhancement mechanism is triggered. By introducing a real-time text library in the automotive field as a supplementary knowledge source, the recognition recall rate of uncommon car models, newly released car series, and colloquial expressions is significantly improved, avoiding the defect of termination upon recognition failure in traditional solutions.

[0054] Enhanced disambiguation capability: By employing a large language model combined with contextual information and retrieval candidate entities for disambiguation, compared to existing methods that rely solely on string matching or shallow semantics, it can deeply understand the true intent of the user's query and significantly reduce the misjudgment rate of polysemous words.

[0055] Dynamically adapting to the evolution of domain knowledge: Through a hybrid knowledge base architecture, the system supports incremental updates and periodic alignment. It can automatically learn new entities from processed text and monitor external automotive data sources to achieve automated knowledge base expansion. The update cycle is shortened from months to real time, ensuring a rapid response to market changes.

[0056] Balancing efficiency and accuracy: The initial recognition model quickly extracts candidate entities, and only triggers secondary recognition by the large language model for difficult cases, avoiding the high computational cost of using LLM in its entirety. While maintaining recognition accuracy, the overall processing time is kept within the range of engineering availability.

[0057] Improve the standardization of results: The secondary identification output includes a standardized representation of standard entity names, type codes and confidence scores, which facilitates direct integration with downstream applications such as search ranking and knowledge graph construction, and reduces the cost of data cleaning and format conversion.

[0058] It has self-optimization capabilities: by recording unrecognizable entities to the manual review queue to form a feedback loop, it continuously iterates and optimizes the quality of the initial recognition model and knowledge base. The system's recognition capabilities can evolve with the accumulation of data and maintain high performance in the long term. Attached Figure Description

[0059] Figure 1 This is an overall flowchart of the entity recognition method in the automotive field provided by the present invention;

[0060] Figure 2 This is a module structure diagram of the automotive entity recognition system provided by the present invention; Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0062] Reference Figure 1 The illustrated method for entity recognition in the automotive domain based on a large language model includes the following steps:

[0063] S1. Receive the text to be processed, which contains named entities in the automotive field;

[0064] This step specifically includes: As a pre-processing step in the entity recognition process, this step is not simply data reception, but rather the construction of a high-quality text input assurance system tailored to the characteristics of the automotive field through a multi-layered and intelligent preprocessing mechanism. This step specifically includes the following creative sub-steps:

[0065] S1.1 Establish an adaptive access channel for multi-source heterogeneous data.

[0066] The system first deploys an access gateway with protocol adaptation capabilities, which can handle multiple heterogeneous data sources such as user Queries from the automotive vertical application search box, UGC discussion content from model forums, crawled text of automotive news articles, and customer service conversation logs. According to the characteristics of text from different sources, the gateway automatically identifies the data format and enables the corresponding parsing engine: for user search Queries, the original input's timing information and session ID are mainly retained for subsequent context association tracing; for news articles, metadata such as the title, publication time, and content section are automatically extracted as the weight basis for subsequent retrieval enhancement. The gateway also integrates a traffic shaping mechanism to smooth out the peaks and valleys of instantaneous high-concurrency queries, preventing peak traffic from impacting downstream recognition modules and ensuring the stability of the core recognition service. In particular, for the short texts and colloquial inputs unique to the automotive field, the gateway embeds a length elastic expansion mechanism that allows a single text to be dynamically adjusted within the range of 10 to 500 characters, avoiding the loss of key information such as the full model name due to excessive truncation.

[0067] S1.2. Implement intelligent quality filtering based on automotive field characteristics

[0068] The received original text first enters the quality filtering layer, which not only performs general junk text recognition but also creatively introduces a special invalid input detection model for the automotive field. This model is trained based on historical user behavior data and can accurately identify non-target queries such as random keyboard inputs (e.g., "asdfgh"), meaningless repeated characters (e.g., "BMW BMW BMW"), and general vocabulary beyond the scope of automotive knowledge (e.g., "today's weather"). For text that contains a model name but has an obvious spelling error (e.g., "Tesla"), the filtering layer does not directly discard it but marks it as "suspected incorrect input" and transfers it to the error correction sub-channel, where candidate correct texts are generated through a fuzzy matching algorithm for subsequent reference by the recognition module. In addition, an anti-cheating mechanism is embedded in this layer to identify and intercept maliciously constructed abnormally long texts or inputs containing injection attack features, ensuring the security of the system.

[0069] S1.3. Perform dynamic privacy desensitization and compliance grading review

[0070] All text must undergo privacy desensitization before entering the recognition process. The system employs a hybrid strategy combining named entity recognition and regular expressions to automatically detect and replace sensitive information such as phone numbers, ID card numbers, and license plate numbers in the text. For privacy data specific to the automotive industry, such as vehicle identification numbers (VINs) and location descriptions from dashcam footage, the system provides pre-defined desensitization templates. The desensitized text is then simultaneously classified according to compliance: text containing personal privacy is marked as "private," and such data is prohibited from being used for knowledge base updates in subsequent recognition processes; completely public information is marked as "public," allowing participation in dynamic knowledge base construction. This classification mechanism not only meets data security regulatory requirements but also provides a legitimate basis for subsequent knowledge evolution.

[0071] S1.4, Implement Timeliness Awareness and Intelligent Priority Routing

[0072] Given the rapid pace of knowledge updates in the automotive field (such as new car model releases and brand name changes), the system automatically tags incoming text with timeliness tags. Specifically, by parsing year and model-related terms (such as "2024 model" and "new generation") within the text, and combining this with metadata about the article's publication time, the system calculates the text's timeliness weight. Information related to new car models published on the current day or the previous day is given the highest priority, ensuring that such text receives a higher ranking weight in subsequent search enhancement stages. User search queries containing time-sensitive terms such as "latest" or "newly launched" are also marked as high-timeliness, and the system allocates them to a separate processing queue to shorten recognition response time. Historical articles with low timeliness enter the batch processing channel, undergoing offline recognition and knowledge base updates during system downtime, optimizing the allocation of computing resources.

[0073] S1.5, Load automotive-specific dictionaries and context cache

[0074] At the end of the text receiving phase, the system preloads a specialized automotive dictionary into its memory cache. This dictionary includes diverse entity variants such as official brand names, market slang, historically discontinued models, and Chinese and foreign language names for imported vehicles. For received text, if it contains terms from the dictionary, a "dictionary hit" tag is appended to the text metadata, and the matched standard entity ID is recorded, providing warm-up information for subsequent mapping and verification. Simultaneously, for user search scenarios, the system maintains a session-level context cache: if the same user has multiple related queries within the last 5 minutes (e.g., searching for "Model Y" followed by "range"), previously identified entities are automatically associated with the current text, forming a context entity linked list, which is then passed to the initial recognition model to improve the recognition consistency of consecutive queries.

[0075] S1.6. Generate a standardized text package and inject tracking tags.

[0076] After completing the above processing, the system generates a standardized data packet for each text, containing: anonymized plain text, source marker, timeliness weight, privacy level, dictionary pre-matching information, context entity linked list, and a globally unique tracking ID. This tracking ID is used throughout all subsequent recognition stages to record the complete processing trajectory of each entity from initial recognition to secondary enhancement. When anomalies occur in the recognition result, the original input and intermediate states can be quickly located based on this ID, achieving end-to-end observability. The data packet is finally encapsulated in a unified Protobuf format and asynchronously distributed to the initial recognition model in step S2 via a message queue, achieving efficient decoupling between the receiving and recognition stages.

[0077] S2. Extract candidate entities from the text to be processed using the first recognition model;

[0078] This step specifically includes: As the first core processing step in entity recognition, this step does not rely solely on a general deep learning model for single-pass annotation. Instead, it achieves accurate capture and intelligent preprocessing of candidate entities through a model architecture specific to the automotive field, a hierarchical recognition strategy, and feature design deeply coupled with subsequent retrieval enhancements. This step creatively constructs a three-in-one entity extraction system of "recognition-evaluation-warm-up," specifically including the following sub-steps:

[0079] S2.1 Loading the Automotive Domain Adaptive Model Weights and Dynamic Labeling System

[0080] The system first loads a first recognition model that has been deeply fine-tuned from automotive-related corpora. This model uses a Transformer architecture as its base, but its weight matrix, in addition to general pre-training, further incorporates transfer learning from millions of automotive-specific texts. The fine-tuned corpus not only includes structured text such as vehicle parameters and technical specifications, but also incorporates colloquial expressions from real-world scenarios such as user comments, forum discussions, and customer service dialogues, enabling the model to recognize colloquialisms, abbreviations, and aliases. Notably, the model's tagging system adopts a dynamically scalable design. In addition to standard brand tags such as B-Brand and I-Brand, it also includes secondary tags such as B-Series-NewEnergy and I-Config-ADAS to distinguish automotive-specific entity types such as new energy vehicle series and intelligent driving configurations. The tagging system supports hot updates; when new entity categories are added to the knowledge base, the model's decoding layer can dynamically register new tags, recognizing new types of entities without retraining.

[0081] S2.2, Perform multi-granularity entity recognition and intelligent boundary correction

[0082] After inputting the standardized text package generated in step S1 into the model, the model is not limited to single-granularity entity recognition, but performs three levels of annotation tasks in parallel: coarse-grained recognition to extract car brands and models (such as "BMW", "3 Series"); medium-granularity recognition to capture model year and powertrain version (such as "2024 model", "plug-in hybrid"); and fine-grained recognition to locate configuration parameters and technical terms (such as "LiDAR", "wheelbase 2900mm"). For the entity boundaries initially output by the model, if overlapping or adjacent entities with a gap of less than 2 characters are detected, the boundary intelligent correction component automatically triggers context dependency analysis and adjusts the boundaries according to the combination rules of automotive terminology. For example, when "Tesla Model Y" is identified as two independent entities, the correction component merges them into a single entity according to the combination pattern of "brand + model name without spaces", avoiding fragmentation problems in subsequent mapping verification. This correction process integrates regularization rules and model re-prediction, which has stronger generalization ability compared to pure rule methods.

[0083] S2.3, Implement dynamic confidence level calibration and low-confidence prediction marking.

[0084] The initial confidence score output by the model undergoes secondary calibration driven by automotive domain features. The system maintains prior confidence distributions for different entity types (e.g., brand entities have an average confidence of 0.95, configuration entities 0.82), and performs Bayesian fusion between the original model confidence score and the prior distribution to obtain a calibrated score that better reflects actual reliability. For candidate entities with a calibrated confidence score below 0.7, the system does not discard them directly. Instead, it predicts that they may be new entities not covered by the knowledge base or highly ambiguous entities, proactively adding a "recommended RAG enhancement" tag and increasing their priority in subsequent retrieval ranking. This predictive mechanism ensures that retrieval enhancement is not passively waiting for mapping failures, but rather that pre-preparation begins during the identification stage, significantly reducing overall processing latency.

[0085] S2.4 Extracting domain association features and context embeddings between entities

[0086] While identifying entities, the model extracts automotive-specific association features between entities: if a brand and a car model appear simultaneously in the text, their co-occurrence probability is calculated and marked as "subordinate association"; if "range" and a specific mileage value appear, it is marked as "attribute-value association". These association features are attached to candidate entities in the form of structured metadata and passed to the mapping verification stage in step S3 to help determine the rationality of the standard name mapping. For example, after "BYD Han" and "DM-i" are identified and marked as "model-power technology" association, the mapping verification can prioritize matching the standard name "Han DM-i" rather than treating "Han" and "DM-i" as independent entities. In addition, the model encodes the deep semantics of the context in which the entity is located into a dense vector. This vector is not used for the current identification but is cached in the semantic retrieval channel of step S4 as a query vector, avoiding repeated encoding calculations and realizing feature-level reuse for identification and retrieval.

[0087] S2.5, Constructing a priority queue for candidate entities and preheating the fast mapping process.

[0088] All identified candidate entities are prioritized based on entity type, confidence level, dictionary hit status, and pre-judgment tags. Brand and vehicle series entities receive the highest priority due to their hierarchical criticality and directly enter the mapping verification of step S3; configuration entities have the next highest priority; entities with the "Suggest RAG Enhancement" tag have the lowest priority, but their retrieval preprocessing tasks are asynchronously submitted to the background queue for early execution. Simultaneously, for entities with the "Dictionary Hit" tag, the system initiates a parallel fast mapping preheating request: locking the corresponding standard entity ID from the knowledge base cache and loading it into high-speed memory. When step S3 officially starts, mapping matching can directly hit the cache, reducing I / O wait time. This preheating mechanism reduces the average response time of mapping verification from 15ms to less than 4ms.

[0089] S2.6. Generate a structured candidate entity package and inject a traceable identifier.

[0090] After the above processing, the system encapsulates candidate entities into structured data packets. Each entity contains the following fields: entity text, location information, post-calibration confidence score, entity type label, list of associated entities, context vector fingerprint, prediction tag, priority score, and source model version number. Each entity is assigned a unique tracing identifier, which forms a parent-child relationship with the text tracking ID in step S1, creating a complete lineage chain from the original input to the candidate entity. If the subsequent step S3 determines that an entity mapping has failed, the system can use this tracing identifier to look up the contextual features and confidence details of its identification, providing observability support for error analysis. The data packets are finally sorted by priority and distributed to parallel processing units via an asynchronous message channel. This decoupling of steps S2 and S3 improves system throughput.

[0091] S3. Map the candidate entities to the automotive domain standard knowledge base, and mark the candidate entities that cannot be successfully mapped as pending entities;

[0092] This step specifically includes: As the key bridge connecting initial recognition and retrieval enhancement, this step does not perform simple string comparison or single similarity calculation. Instead, through a hybrid knowledge base architecture specialized for the automotive domain, a hierarchical decision-making mechanism, and forward-looking failure prediction technology, accurate mapping and intelligent marking are achieved. This step creatively constructs a four-level mapping system of "exact - fuzzy - semantic - context" and an intelligent decision-making closed loop of "difficulty prediction - classification marking - priority ranking", specifically including the following sub-steps:

[0093] S3.1. Load the hybrid standard knowledge base for the automotive domain and activate multi-dimensional indexing

[0094] The system first loads the automotive domain standard knowledge base, which adopts a dual-engine hybrid architecture of a graph database and a vector database. The graph database uses automotive brands as root nodes and constructs a hierarchical entity network with a five-layer architecture of "brand - series - model - year - configuration". Each node stores structured attributes such as standard name, official alias, discontinuation mark, and launch time; the vector database stores the semantic vector representations of the above entities, covering non-standardized expressions such as brand common names, model nicknames, and technical parameter abbreviations. When the knowledge base starts, the system automatically activates three types of indexes: The exact index constructs a B+ tree structure based on the entity standard name and alias, supporting millisecond-level literal matching; the fuzzy index uses N-gram tokenization technology to adapt to spelling variants entered by users (such as "Cadillac" and "凯迪拉克"); the semantic index is based on the HNSW graph structure of the vector database to achieve approximate nearest neighbor retrieval in high-dimensional space. In particular, the knowledge base has the ability of real-time hot update. When a new model is released, operators can enter the standard name in the management background, and the system can complete the creation of graph database nodes, vector encoding, and incremental update of the three-level index within 30 seconds through the change stream mechanism, realizing seamless expansion of the knowledge base without restarting the service.

[0095] S3.2. Execute a four-level hierarchical mapping and matching strategy

[0096] For each candidate entity, the system parallelly starts four levels of mapping decision-making processes and executes them in order of priority:

[0097] Exact matching layer: Perform character-level strict matching of the candidate entity text with the standard name, official alias, and historical former names in the knowledge base. If it is completely matched, directly return the standard entity ID and confidence level 1.0, and skip the subsequent levels. This layer specifically processes the common non-space concatenated model names in the automotive domain (such as "Audi A6L 45 TFSI"), and performs intelligent word segmentation and combined matching on it through a preset brand prefix dictionary and regular rules.

[0098] Fuzzy Similarity Layer: For entities that are not precisely matched, the character-level similarity is comprehensively calculated using the edit distance and Jaro-Winkler similarity. This layer incorporates character tolerance rules in the automotive domain. For example, the space difference between "Tesla Model3" and "Tesla Model 3", and the missing letter between "Mercedes-Benz E260" and "Mercedes-Benz E260L" are given a relatively high similarity tolerance. If the similarity exceeds the first threshold (0.9 for brand entities, 0.85 for vehicle series entities, 0.8 for configuration entities), it is determined that the mapping is successful, and the corresponding standard entity ID and similarity are returned as the confidence level.

[0099] Semantic Alignment Layer: For candidate entities with significant character differences (such as mapping the user's common name "Model Y" to "Model Y"), the system extracts the context semantic vectors cached in step S2 and calculates the cosine similarity with the standard entity vectors in the vector database. This layer creatively introduces a timeliness decay factor, which gives a 1.2-fold weight boost to the vector similarity of newly launched vehicle entities (released within 3 months), to address the problem of vector representation deviation caused by insufficient training corpus for new entities. If the semantic similarity exceeds the second threshold (0.75), and there is an attribute intersection between the context of the candidate entity and the attribute description of the standard entity (such as "SUV", "pure electric"), it is determined that the mapping is successful.

[0100] Context Association Layer: For multiple candidate entities identified simultaneously (such as "BYD Han" and "DM-i"), this layer analyzes the co-occurrence relationship between the entities and constructs a domain association graph. If there is a combined entity node "Han DM-i" in the knowledge base, and the text distance between the two candidate entities is less than 5 characters, they are preferentially merged and mapped to the combined entity rather than being mapped independently. This mechanism effectively solves the problem of fragmented recognition of the continuous text of "brand + vehicle series + technology type" input by users, improving the coherence and accuracy of the mapping.

[0101] S3.3. Implement Dynamic Threshold Decision Based on Entity Type and Historical Behavior

[0102] Instead of using a fixed threshold for mapping determination, the system calculates a dynamic adaptive threshold for each candidate entity. This threshold is composed of three weighted parts: the base threshold is set according to the entity type (brand 0.88, vehicle series 0.82, configuration 0.75); the confidence compensation term is linearly adjusted according to the model confidence output in step S2. For entities with high confidence (>0.9), the threshold is reduced by 0.05 to increase the passing rate, and for entities with low confidence (<0.7), the threshold is increased by 0.08 to enhance strictness; the historical success rate feedback term is based on the mapping success statistics of this entity in the recent 30 days. If the historical success rate >95%, the threshold is lowered by 0.03, otherwise it is raised. After the threshold calculation is completed, the system performs cross-validation: if the candidate entity meets the threshold requirements of both the fuzzy similarity layer and the semantic alignment layer at the same time, but the standard entity IDs returned by the two are inconsistent, the conflict arbitration mechanism is triggered. The result of the semantic alignment layer is preferentially adopted (because it integrates context information), and the conflict cases are recorded in the quality analysis log for subsequent threshold tuning.

[0103] S3.4. Conduct forward-looking prediction and soft labeling of mapping failures

[0104] For candidate entities that do not reach the dynamic threshold, the system does not immediately determine them as mapping failures, but starts a forward-looking prediction engine. The engine analyzes multiple dimensional features of this entity: the length of the entity text (whether it is an abbreviation), the confidence score (whether it is in the marginal range of 0.6 - 0.75), the dictionary hit status (whether it matches the fuzzy dictionary), and the context timeliness weight (whether it is associated with newly released content). Based on these features, the system calculates the mapping difficulty coefficient and divides the candidate entities into three categories:

[0105] High-probability mapped type: the difficulty coefficient <0.3, mostly spelling mistakes (such as "Benma" → "Benz"), directly trigger automatic error correction and try mapping again.

[0106] Marginally pending type: the difficulty coefficient is 0.3 - 0.7, marked as "soft pending", reduce its mapping priority but keep it at the end of the candidate queue. If high-confidence evidence is retrieved in step S4, the mapping result can still be reversed.

[0107] High-probability failure type: the difficulty coefficient >0.7, marked as "hard pending", directly determine it as a mapping failure and preferentially push it to the S4 retrieval enhancement channel.

[0108] This prediction mechanism avoids the "one-size-fits-all" mapping failure determination, provides a second chance for boundary entities, and optimizes the system resource allocation.

[0109] S3.5. Construct an intelligent classification and multi-dimensional labeling system for pending entities

[0110] For all entities determined to be mapping failures, the system generates intelligent classification labels and priority scores for them:

[0111] Category tags: Based on the reasons for failure, it is divided into three categories: knowledge deficiency type (no corresponding entity in the knowledge base, such as a brand new brand), strong ambiguity type (corresponds to multiple similar entities, such as "E200" matching both Mercedes-Benz and Lexus), and suspected misidentification type (extremely low confidence and inconsistent with the context semantics, such as misidentifying "driving" as a car model).

[0112] Priority scoring: The score is calculated by weighting the entity's timeliness (new models +3 points), contextual importance (appearing at the beginning of the query +2 points), and business value (brand entities +5 points, configuration entities +1 point). The higher the score, the higher the priority, which determines its retrieval resource allocation and queuing order in step S4.

[0113] After tagging, the system constructs a contextual evidence chain for each entity to be identified, including its original text, location information, related entities, timeliness tags, and traceability ID, packaged as structured data. This evidence chain is passed along with the entity to step S4, enabling enhanced retrieval without repeatedly parsing the text, directly locating the root cause of ambiguity based on the evidence chain, and improving retrieval efficiency.

[0114] S3.6 Generate mapping result data packets and record traceable decision logs.

[0115] After mapping all candidate entities, the system generates a unified result data package, including: successfully mapped entities and their standard IDs, unmapped entities and their classification labels and priorities, and a mapping path record for each entity (which layers of judgment were passed, the similarity scores at each layer, and the final decision basis). Simultaneously, the decision log engine records the complete mapping trajectory: the number of knowledge base queries, threshold configurations at each layer, dynamically adjusted parameters, prediction accuracy, and other performance metrics. The logs are batch-stored into a time-series database using an aggregated writing strategy, supporting subsequent full-link auditing by tracking ID and providing training data for the threshold self-learning algorithm. Finally, the data package is transmitted to step S4 asynchronously, and knowledge base connection resources are released immediately, achieving efficient reuse of system resources.

[0116] S4. When there are entities to be determined, relevant documents are retrieved from the dynamic text library in the automotive field as supplementary knowledge based on the text to be processed.

[0117] This step, as the core execution link of retrieval enhancement, does not simply perform keyword recall or vector similarity queries. Instead, it constructs a high-precision, low-latency, and traceable contextual knowledge supplementation system through the timeliness perception of automotive domain entities, multi-dimensional query intent generalization, and a quality stratification mechanism for retrieval results. This step creatively constructs a four-level retrieval enhancement framework of "query understanding - hybrid retrieval - quality evaluation - knowledge distillation," specifically including the following sub-steps:

[0118] S4.1. Activate the multi-modal indexing engine of the automotive domain dynamic text library

[0119] The system first starts the hybrid indexing engine of the dynamic text library. The underlying storage of this engine is not single text content, but multi-modal document objects. Each document object contains: article title and body text, publication timestamp, data source identifier (such as Autohome, Dongchedi, Xcar), content section classification (new car information, evaluation and shopping guide, technical analysis, user reviews), and pre-computed paragraph-level semantic vectors. The indexing engine maintains three types of retrieval channels simultaneously: The full-text inverted index channel supports character-level fast matching and particularly optimizes the fault tolerance for non-space variants of vehicle model names (such as "BYD Han DM-i" and "BYD Han DMi"); The vector approximate retrieval channel is based on the HNSW graph structure and supports millisecond-level recall of high-dimensional semantic vectors; The time-series partition channel divides documents into hot data area (last 7 days), warm data area (8 - 30 days), and cold data area (more than 30 days) according to the publication time. When retrieving, it preferentially accesses the hot data area to ensure that information related to new vehicle models can be recalled first. When the system detects the release of a new vehicle series from an external data source (such as by monitoring the Ministry of Industry and Information Technology announcements or changes on the manufacturer's official website), the scraping module automatically triggers the construction of incremental indexing, and new articles can participate in retrieval within 30 seconds after scraping, solving the problem of lag in traditional static knowledge bases.

[0120] S4.2. Execute query intention generalization and rewriting based on the difficulty of pending entities

[0121] For the pending entity and its context evidence chain passed in from step S3, the query rewriting module implements a hierarchical generalization strategy. For entities with missing knowledge (such as new brands), the rewriting engine extracts the character features of the entity text (such as the radical similarity between "Zeekr" and "Lynk"), generates a set of phonetically / structurally similar candidate queries, and calls the automotive domain synonym generation model to predict the possible official naming patterns of the new entity and expand the retrieval keywords. For highly ambiguous entities (such as "E200"), the rewriting engine combines the associated entities in the context evidence chain (such as "Mercedes-Benz" appearing above), injects a disambiguation anchor point into the query, and generates a targeted query statement "Mercedes-Benz E200" to preferentially recall documents related to the Mercedes-Benz brand. For entities suspected of misidentification, the rewriting engine reduces their weight in the query and instead extracts the core semantics from other high-confidence entities in the context to reconstruct the query statement, avoiding ineffective retrieval of incorrect entities. In particular, if the text to be processed contains a timeliness marker (such as "2024 model"), the rewriting engine automatically injects the year information into the query to make the retrieval results focus on relevant articles of that year and improve the recall accuracy.

[0122] S4.3. Start three-channel parallel hybrid retrieval and dynamic relevance weighting

[0123] The rewritten query is distributed to three retrieval channels for parallel execution:

[0124] Exact Match Channel: Retrieves documents in the full-text inverted index that are completely identical to the rewritten query or have an edit distance of less than 2, prioritizing high-precision content such as official press releases and vehicle parameter configuration tables. The relevance score of the results from this channel is weighted by the authority of the source (manufacturer's official website > vertical media > user forums).

[0125] Semantic Relevance Channel: This channel uses the contextual semantic vector of the text to be processed as the query vector to retrieve the Top-K related documents from the vector library. This channel creatively introduces brand lineage similarity weighting. If the query involves "BMW iX3", documents of other BMW models (such as "BMW X3") receive a 1.15x weighting in the vector similarity calculation, leveraging the semantic coherence of technical terms within the brand to improve recall quality.

[0126] Timeliness Compensation Channel: Independent of text content, this channel randomly samples documents from high-value time windows based on the timeliness weight of the text to be processed. For example, for a query marked "new car release", this channel forcibly recalls the 5 new articles with the highest confidence in the hot data area to prevent the omission of key information due to low vector similarity.

[0127] After the three-channel results are returned, the dynamic weighted fusion module allocates channel weights based on the priority scores of the entities to be determined: for high-priority entities (such as brands), the exact match channel accounts for 60% of the weight, and the semantic channel accounts for 30%; for low-priority entities (such as configurations), the semantic channel weight is increased to 50%. The fused document list is sorted in descending order of comprehensive relevance score, and duplicate documents are removed.

[0128] S4.4 Implement quality stratification and credibility traceability marking of search results

[0129] The merged documents are then placed into a quality stratification unit and divided into three credibility levels:

[0130] High-credibility layer: Sources include official websites of car manufacturers, announcements from the Ministry of Industry and Information Technology, and officially certified parameter configuration documents, which are directly used as core supplementary knowledge paragraphs.

[0131] Mid-credibility layer: Review articles from automotive vertical media and new car interpretations written by editors need to be filtered for opinion-based content, removing subjective evaluations and speculative statements, and retaining only objective parameter descriptions.

[0132] Low-trust layer: User forum discussions and comment section content, only frequently mentioned entity information is extracted as auxiliary verification, and is not included in the main knowledge package.

[0133] Each document, after being layered, is tagged with a credibility traceability marker, including: source domain credibility score, author authority (certified editors > ordinary users), publication time freshness decay factor (attenuated by antennal attenuation), and co-occurrence frequency with the entity to be identified. This marker is input into the large language model in step S5 along with the document content, providing an interpretable basis for the model to judge the reliability of information.

[0134] S4.5 Intelligent extraction and redundancy elimination of key document segments

[0135] Documents entering the high-confidence and medium-confidence layers require fragment extraction and compression to avoid information overload caused by inputting the entire text into the LLM. The system uses a car entity density index to guide extraction: it scans each paragraph of the document, counts the number and density of car brand, model, vehicle type, configuration, and parameter entities contained therein, and prioritizes extracting the 3-5 paragraphs with the highest entity density. For parameter configuration documents containing tables, the table parsing engine automatically converts the table structure into natural language descriptions (e.g., "The CLTC range of the Model Y long-range version is 688 kilometers"), and labels the data source as "official configuration table". After extraction, cross-document redundancy elimination is performed on the fragments: if multiple fragments contain the same entity information, only the fragment with the most authoritative source and the most recent publication time is retained, and the rest are deleted to avoid duplicate information interfering with the LLM judgment.

[0136] S4.6 Generate a structured supplementary knowledge package and inject a time decay mechanism.

[0137] Finally, all filtered, extracted, and deduplicated document fragments are packaged into a structured supplementary knowledge package. The knowledge package is organized hierarchically: the top layer is a query summary (containing the entities to be identified, the core semantics of the text to be processed, and the retrieval timestamp); the middle layer is a list of documents sorted by credibility (each document includes source, publication time, key fragments, and credibility markers); and the bottom layer is an entity density histogram (showing the frequency distribution of each entity in the supplementary knowledge). Before output, the knowledge package is injected with a time-lapse mechanism: for document fragments published more than 90 days ago, their content credibility markers are automatically multiplied by a decay coefficient of 0.8, and a "historical data warning" is added to the header of the knowledge package, prompting the large language model in step S5 to prioritize recent information during inference. This mechanism ensures the time validity of the supplementary knowledge and prevents outdated information (such as old prices of discontinued car models) from misleading entity recognition.

[0138] S4.7 Implement search cost control and dynamic degradation strategies

[0139] The entire retrieval process is monitored by a cost controller. The system presets a time limit (800ms) and a call limit (maximum access to 10 million documents in the vector library per query). If the three-channel parallel retrieval times out, the cost controller automatically terminates the semantic channel and the time-compensation channel, retaining only the fast path of the exact matching channel and triggering a degradation alarm. If the limit is hit, tiered sampling degradation is initiated: first, the hot data area is searched, and if the number of results is insufficient, it is gradually expanded to the warm data area and the cold data area to avoid the resource waste caused by a full database scan. All degradation events are recorded in the monitoring log for subsequent optimization of retrieval strategies and scaling decisions, ensuring the stability and economy of the system under high-traffic scenarios.

[0140] S5. Input the entity to be identified and the supplementary knowledge into the large language model, perform secondary entity recognition through the large language model to determine the standard name of the entity to be identified, and output the final entity recognition result.

[0141] This step, as the core decision-making process of the retrieval enhancement workflow, does not simply concatenate the pending entity with document fragments and feed them into a large language model. Instead, it constructs an intelligent recognition closed loop with self-doubt and evidence traceability capabilities through automotive-specific prompting engineering, multi-source knowledge fusion control, and reliability verification of the generated results. This step creatively constructs a five-layer reasoning enhancement framework of "role injection - evidence screening - constraint generation - credibility assessment - feedback accumulation," specifically including the following sub-steps:

[0142] S5.1 Implementing deep keyword injection and task deconstruction for automotive expert roles

[0143] The system first constructs hierarchical preset prompts, defining not only the basic tasks of LLM (Limited Learning Manager) but also creatively incorporating the role of automotive experts and the requirement for transparency in reasoning. The top-level prompts explicitly state that LLMs must act as "senior automotive product managers," understanding industry norms such as brand lineage, model generations, and configuration differences. The task description breaks down secondary identification into three sub-tasks executed sequentially: entity existence verification (determining whether the entity truly exists in the supplementary knowledge), standard name inference (deriving a unique standard name based on the chain of evidence), and disambiguation reason generation (outputting original text citations supporting the conclusion). The prompts embed mandatory instructions for the thought chain, requiring LLMs to explicitly demonstrate the step-by-step reasoning process—"first analyze the timeliness of candidate documents → then compare the consistency of entity descriptions → finally eliminate interfering items by comprehensively considering the context"—before outputting the final conclusion. This design transforms black-box generation into white-box decision-making, facilitating subsequent root cause tracing of LLM misjudgments and providing interpretable feedback for optimizing the mapping threshold in step S3.

[0144] S5.2, Implement quality-layered fusion and information density control of supplementary knowledge.

[0145] Faced with multiple supplementary knowledge documents passed in from step S4, the system does not directly input the entire amount into the LLM. Instead, it activates an information density controller. The controller first performs a secondary filtering of knowledge fragments based on document credibility and entity density: retaining only the full content of high-credibility documents, extracting only a three-sentence window containing the context of the entity to be determined from medium-credibility documents, and providing only entity frequency statistics from low-credibility documents. Subsequently, the controller calculates the evidence weight for each piece of information. If multiple documents conflict in their descriptions of the same entity, the most recent and authoritative source is prioritized, and conflicting content is marked with a "contradictory evidence" warning label, requiring the LLM to explicitly address contradictions during reasoning. To avoid long contexts causing LLM attention to become scattered, the system implements dynamic truncation: when the total number of tokens exceeds 80% of the model's effective window, the end of the document with the lowest semantic similarity to the entity to be determined is truncated first, rather than mechanically truncated by time, ensuring that core evidence remains in the LLM's high-responsiveness region.

[0146] S5.3, Initiating Context Disambiguation Reasoning Guided by Entity Association Graph

[0147] When inputting LLM, the system not only provides pending entities and supplementary knowledge, but also injects the entity association graph extracted in step S2 as an auxiliary reasoning structure. The association graph describes the brand-vehicle series affiliation and configuration-model dependency relationships between all identified entities (including successfully mapped ones) in the current query in JSON format. LLM is prompted to use this graph for cross-entity joint reasoning: for example, when the pending entity is "long range," the successfully mapped "Model Y" node in the association graph will guide LLM to prioritize finding evidence of "long range" in the Model Y configuration documents, rather than generalizing the search for "long range" information across all brands. For highly ambiguous entities like "E200," if a "Mercedes-Benz" node exists in the association graph, the contextual signal of the Mercedes-Benz brand is explicitly enhanced in the prompt, guiding LLM to focus its reasoning on the Mercedes-Benz brand by increasing the resampling rate of "Mercedes-Benz" related document fragments. This mechanism transforms the disambiguation of isolated entities into collaborative decision-making among related entities, significantly improving disambiguation accuracy.

[0148] S5.4 Implement adversarial validation and forced formatting of structured output.

[0149] During the generation phase of the LLM, adversarial prompting is used to ensure that the output strictly conforms to the predefined JSON Schema. In the prompt, not only are the field names and types defined, but also anti-hallucination instructions are embedded: "If the standard name cannot be determined from the supplementary knowledge, the confidence field must be set to 0 and an empty string must be returned, and no fabrication is allowed." The system deploys a two-layer formatting guarantee: The first layer is the self-verification of the LLM, which requires it to first output a draft in JSON format and then self-check the field integrity; the second layer is the post-verification of the system, which uses JSONSchema to perform syntax and semantic verification on the output. If the verification fails, an automatic retry is triggered, and the error message is fed back to the LLM for correction. The upper limit of the retry is 3 times. If all retries fail, it is determined as an identification failure to prevent incorrect results from flowing into the downstream. This mechanism improves the accuracy of the output of the vehicle series standard name from 92% in free generation to over 99.5%.

[0150] S5.5. Perform confidence calibration and evidence traceability verification for LLM self-evaluation

[0151] The structured result output by the LLM contains the confidence field for self-evaluation. This score is not randomly generated but is calculated based on the coverage and consistency of the evidence chain within it. The system further calibrates the confidence: If all the documents cited by the LLM are from the high-confidence layer and have the same publication time, the confidence is multiplied by 1.1 for addition; if there are contradictions in the cited documents or the sources include the low-confidence layer, the confidence is multiplied by 0.85 for penalty. At the same time, the system performs evidence traceability verification: Extract the original text that it claims to cite from the reason segment output by the LLM and reverse-check whether this segment actually exists in the supplementary knowledge. If it is found that the LLM fabricates citations (hallucinations), the confidence is forced to be set to 0 and marked as "suspected hallucination", and this result does not participate in the final output. This verification mechanism suppresses the hallucination rate of the LLM from 5% in typical scenarios to below 0.3%.

[0152] S5.6. Make decisions on credibility grading results and multi-path output

[0153] The calibrated confidence score drives the result grading decision:

[0154] High-confidence results (confidence > 0.85): Directly accepted as the standard name, output to the final result, and asynchronously trigger the incremental update of the knowledge base in step S3.

[0155] Medium-confidence results (0.6 < confidence ≤ 0.85): Marked as "requiring manual review", stored in the pending review queue, but allowed to participate in the current identification output. Downstream applications can choose whether to adopt them according to the risk tolerance.

[0156] Low confidence result (confidence≤0.6): The identification is deemed a failure, the entity is not output, and the case along with the supplementary knowledge is pushed to the manual annotation platform to optimize the model and strategy of steps S1-S4.

[0157] For results deemed reasonably reliable, the system triggers an additional secondary verification: a lightweight validation model (a BERT-based binary classifier) ​​is invoked to perform a secondary evaluation of the LLM output. If the validation model also determines the result to be reliable, the confidence level is increased to 0.9; if it determines the result to be unreliable, the level is downgraded to low reliability. This multi-path decision-making mechanism achieves a dynamic balance between strictness and recall.

[0158] S5.7, Generate final recognition results and end-to-end feedback.

[0159] The final recognition results contain three types of entities:

[0160] Entities that are successfully mapped on the first attempt: retain the standard ID and basic confidence level of step S3.

[0161] Successful secondary enhancement entities include the standard name of the LLM inference, the calibrated high confidence level, a list of cited source document IDs, and a summary of the disambiguation inference.

[0162] Identify failed entities: Only retain the entity text, location information, and failure reason classification (knowledge gap / strong ambiguity / insufficient evidence).

[0163] All results are packaged into a fully traceable data block, with added metadata such as processing time, model version, and decision path for each step from S1 to S5. This data block is not only output to downstream applications but also asynchronously written to a feedback repository for daily system performance report generation and error pattern analysis. In particular, for "knowledge-deficient" failure cases, the system automatically triggers a new entity discovery process: the entity text, context, and supplementary knowledge summary are submitted to the operations review platform. After approval, it can be converted into a standard knowledge base entry with one click, achieving continuous closed-loop evolution of system capabilities.

[0164] Preferably, the standard knowledge base in the automotive field is a hybrid knowledge base system, including:

[0165] A structured database used to store standard names, synonyms, and hierarchical relationships for car brands, car series, and car models;

[0166] Entity vector library, used to store semantic representation vectors of standard names and synonyms.

[0167] Preferably, the relevant documents retrieved include:

[0168] Based on the entity name of the entity to be determined, perform an exact match query in the structured database to obtain the first query result;

[0169] Based on the contextual semantic features of the entity to be processed in the text, a similarity search is performed in the entity vector database to obtain the second query result.

[0170] Combine the results of the first and second queries to generate supplementary knowledge.

[0171] Preferably, the dynamic text library in the automotive field is a real-time updated database of automotive news articles, and the database is constructed in the following way:

[0172] We extract articles from automotive vertical media platforms that include new car launches, model reviews, and technical analyses.

[0173] Incremental indexing of articles and assigning retrieval weights based on publication time;

[0174] The hybrid knowledge base system is dynamically updated in the following ways:

[0175] Obtain the newly added entities from the secondary recognition output of the large language model;

[0176] When the semantic similarity between a newly added entity and an existing entity in the knowledge base is greater than a preset threshold, the newly added entity will be added to the knowledge base system as a synonym.

[0177] Monitor external automotive data sources, and automatically extract entities and update the knowledge base system when a new model or series of vehicles is detected.

[0178] Preferably, marking candidate entities that cannot be successfully mapped as pending entities includes:

[0179] Calculate the semantic similarity between candidate entities and standard names in the standard knowledge base;

[0180] When the semantic similarity is lower than the dynamic threshold, the mapping is deemed to have failed.

[0181] Among them, the dynamic threshold is adaptively adjusted based on the recognition confidence of candidate entities and the historical mapping success rate;

[0182] Preferably, inputting the undetermined entity and supplementary knowledge into the large language model includes:

[0183] Construct preset prompt words, which include the entity to be determined, the context information of the text to be processed, and the list of candidate entities in the supplementary knowledge;

[0184] Preset prompts are used to guide the large language model to perform disambiguation recognition of the target entity based on contextual information;

[0185] Preset prompts are also used to guide the large language model to output standardized entity representations, which include standard entity names, entity type codes, and confidence scores.

[0186] Preferred options also include:

[0187] The number of entities whose standard names cannot be determined after secondary entity recognition is counted;

[0188] When the number exceeds the tolerance threshold, the corresponding text to be processed and the recognition result are recorded in the manual review queue for optimization of the first recognition model and the hybrid knowledge base system.

[0189] Reference Figure 2 The illustrated automotive domain entity recognition system includes:

[0190] The initial recognition module is used to extract candidate entities from the text to be processed through the first recognition model;

[0191] The mapping and verification module is used to map and match candidate entities with the standard knowledge base in the automotive field, and mark candidate entities that cannot be mapped as pending entities;

[0192] The retrieval enhancement module is used to retrieve relevant documents from a dynamic text library in the automotive field as supplementary knowledge when there is a pending entity.

[0193] The secondary recognition module is used to input the entity to be identified and supplementary knowledge into the large language model for secondary entity recognition and output the final entity recognition result.

[0194] The dynamic update module is used to automatically update the standard knowledge base based on the recognition results;

[0195] The system executes the methods described above.

[0196] Furthermore, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0197] Furthermore, a non-transitory computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the above-described method.

[0198] This system constructs an intelligent entity recognition system for automotive vertical scenes with self-diagnosis, self-correction, and self-expansion capabilities through a domain-adaptive modular architecture and dynamic knowledge evolution mechanism. Breaking away from the static recognition paradigm of traditional single-unit models, the system creatively adopts a three-layer architecture design of "hierarchical decision-making, collaborative enhancement, and closed-loop feedback." Each module achieves state sharing and policy linkage through a dedicated automotive event bus, forming an organic whole.

[0199] Detailed description of the system's core modules:

[0200] The initial recognition module does not perform static inference using a single model, but instead incorporates a flexible model scheduling engine and a multi-granularity feature pre-extractor. Internally, the module maintains a model family from TinyBERT to BERT-Large, dynamically selecting the optimal model variant based on the latency and accuracy requirements of the business scenario. During recognition, the module not only outputs candidate entities but also creatively extracts three types of pre-computed features: the ambiguity entropy value of entity boundaries, the conflict index of the context, and the weight of timeliness implications. These features are then passed to downstream modules via zero-copy shared memory, achieving one-time computation and global reuse. The module also incorporates a pre-hit mechanism based on an automotive domain dictionary, quickly intercepting and pre-marking common market terms such as "Hongguang MINI" and "Big G," providing prior clues for subsequent mapping verification.

[0201] The mapping and verification module deploys a dual-engine decision-making system based on a hybrid knowledge base in the automotive domain. This knowledge base uses a graph database to store five layers of lineage relationships (brand-model-vehicle series) and a vector database to store semantic representations of millions of entity variants. The two are strongly associated across databases via entity IDs. Matching decisions are not simply threshold judgments, but rather a four-level cascaded decision-making process involving precise matching, fuzzy similarity, semantic alignment, and contextual association. The decision-maker incorporates character error-tolerance rules specific to the automotive domain, enabling intelligent word segmentation and combination mapping for vehicle names like "Audi A6L45TFSI" that do not use spaces. Most innovatively, the module proactively predicts entities that fail to map, calculates the mapping difficulty coefficient, generates a failure feature code, and actively pushes it to the retrieval enhancement module via an event bus. This allows the module to directly obtain retrieval strategy guidance without parsing the text, achieving intent-level collaboration across modules.

[0202] The retrieval enhancement module constructs a multimodal index and strategy-aware retrieval system for a dynamic text library in the automotive field. The text library not only stores article text but also encapsulates multi-dimensional attributes such as publication time, source authority, content section tags, and paragraph-level semantic vectors, forming searchable document objects. The retrieval process abandons single-keyword matching and automatically loads a pre-defined optimal retrieval strategy based on the failure signature codes of the pending entities (knowledge-deficient / strongly ambiguous / misidentified). The module creatively introduces brand lineage similarity weighting and e-commerce scenario awareness capabilities. When searching for "Han DMI," it prioritizes recalling relevant documents from the BYD Dynasty website; when identifying price-sensitive terms such as "bottom fishing," it automatically expands promotional knowledge sources. The retrieval results undergo entity density index calculation and cross-document redundancy elimination, retaining only high-value fragments and adding credibility traceability markers and timeliness attenuation weights to provide a structured chain of evidence for LLM inference.

[0203] The secondary recognition module incorporates a prompt word arrangement engine tailored to automotive industry experts and an adversarial validation framework. The prompt words not only define the task but also infuse the product manager's cognitive and thought process with mandatory instructions, requiring the LLM to explicitly demonstrate the reasoning process. The module performs dual validation on the LLM output: the first layer is self-validation, requiring the model to self-assess its completeness after outputting a JSON draft; the second layer is system-wide source verification, reverse-checking whether the document fragments referenced by the model actually exist in the knowledge package, suppressing the illusion rate to below 0.3%. Most innovatively, the module employs a tiered decision-making process for calibrated confidence levels: high-confidence results are directly stored in the database, medium-confidence results trigger a secondary confirmation by a lightweight validation model, and low-confidence results are rejected and pushed to a manual annotation platform, forming a complete quality control closed loop of "generation-validation-tiering-feedback".

[0204] The dynamic update module enables incremental learning and version-based lifecycle management of the automotive knowledge base. The module performs semantic collision detection on new entities to avoid duplicate synonym additions; it initiates automatic alignment for strongly similar entities, constructing subtype affiliation relationships; and it triggers a cold start acceleration process for new brand-model combinations, automatically filling in technical parameters and setting a temporary confidence decay period. Each update generates a time-series version, with old versions automatically archived, supporting full rollback. The module also monitors external data sources (MIIT announcements, brand websites), and when a new model release event is detected, it completes end-to-end data import within 30 seconds, from data crawling, entity extraction, relationship construction to index publishing, shortening the traditional manual operation cycle from several weeks to real-time.

[0205] Inter-module collaboration mechanisms and global optimization:

[0206] The system's modules achieve loosely coupled collaboration through an automotive-grade event bus. When a mapping failure event from the mapping verification module is broadcast with a signature code, the retrieval enhancement module can immediately adjust its retrieval weights. The recall rate metric from the retrieval enhancement module is fed back to the dynamic update module in real time, which then optimizes the knowledge base update frequency accordingly. A hallucination error event from the secondary recognition module triggers the automatic revision of the prompt word template. This event-driven architecture enables the system to adapt its strategies, allowing for online adjustments to module behavior without requiring a restart.

[0207] The feedback engine aggregates cross-module logs daily and performs global parameter self-learning. If it finds that a certain type of entity frequently fails in step S3 but succeeds with a high probability in step S5, the self-learning algorithm automatically lowers the dynamic threshold for that entity in mapping verification, making it easier for it to enter the retrieval enhancement path. Conversely, if the retrieval module's recall quality remains consistently low, the semantic retrieval weight is increased and the synonym rewriting rules are expanded. The self-learning process is verified using A / B testing to ensure that adjustments have no negative impact, allowing the overall system performance to continuously evolve over time.

[0208] Electronic device implementation:

[0209] The storage system employs a tiered storage architecture for automotive data. Hot data (user queries and knowledge base updates over the past 7 days) is stored on NVMe SSDs, enabling microsecond-level access via memory-mapped files; warm data (identification logs over the past 30 days) is stored on SATA SSDs, supporting rapid batch analysis; cold data (historical versions of the knowledge base) is automatically migrated to object storage and loaded on demand via the S3 interface. The storage system also includes a feature cache, which caches the pre-calculated features extracted in step S2 in binary serialized format for 12 hours, allowing downstream modules to reuse them and avoiding redundant calculations.

[0210] When the processor executes a computer program, it activates a dynamic scheduling engine for heterogeneous computing resources. The initial identification module and lightweight verification model are deployed on a GPU inference cluster, achieving a query throughput of 2000 queries per second using TensorRT optimization. Large language model inference is exclusively handled by the AI ​​accelerator card cluster, employing model parallelism and dynamic batch processing technologies to keep the average response latency within 800ms. The knowledge base retrieval service runs on in-memory computing nodes, achieving cross-node vector retrieval latency of less than 5ms via RDMA networking. The processor also integrates a circuit breaker and degradation strategy: when the GPU load exceeds 90%, the initial identification model is automatically switched to a lightweight version running on the CPU; when the AI ​​accelerator card fails, the secondary identification module automatically switches to the fast verification mode of the local BERT model, rejecting complex inference while ensuring the availability of basic functions. This design ensures business continuity of electronic equipment under resource fluctuations and failure scenarios.

[0211] Non-transitory computer-readable storage media:

[0212] The computer program stored on the storage medium not only contains the system's runtime code but also a versioned snapshot of the automotive domain knowledge base. Encoded in Protocol Buffers format, the snapshot includes the knowledge base's graph structure, vector representation, temporal version number, and prior distribution of entity difficulty. It can be loaded within 10 seconds during a system cold start, accelerating the recovery process. The program code adopts a plug-in architecture, with each recognition module compiled as an independent dynamic link library, supporting runtime hot updates. When a dynamically updated module discovers a new entity recognition pattern, it can automatically download the A / B-tested and verified new version of the module and replace the old version without restarting the entire service.

[0213] The storage medium also records an immutable sequence of full decision logs. Each log entry records the complete processing trajectory from step S1 to S5, the module version number, dynamic parameter configuration, and the basis for the final decision. A Merkle tree structure is used to ensure the logs are tamper-proof and meet data compliance auditing requirements. Log files are sharded by time and compressed for storage. Logs from the past 180 days can be retrieved and traced back within seconds, providing full lifecycle data support for system error diagnosis and performance optimization.

[0214] Example 1: Complete Recognition Process in the Automotive Vertical Search Scenario

[0215] This example describes the end-to-end processing process when a user searches for "2024 Model Y Long Range On-road Price" through the Autohome APP.

[0216] After the system receives this search query, the text reception module first identifies it as a highly time-sensitive user query, marks the privacy level as "public", and loads the context cache (if the user's previous query was "Tesla maintenance cost", then cache the "Tesla" brand entity). The initial recognition module runs the BERT-CRF model fine-tuned with automotive corpus and outputs candidate entities: "Model Y" (confidence 0.68), "Long Range" (confidence 0.71), "On-road Price" (non-entity). After querying the standard knowledge base, the mapping verification module finds that the semantic similarity between "Model Y" and "毛豆Y" is only 0.58, lower than the dynamic threshold of 0.75, and there is a strong ambiguity (possibly referring to the edamame plant), so it determines that the mapping fails, marks it as a pending entity, and attaches a "knowledge missing type" feature code.

[0217] After receiving the feature code, the retrieval enhancement module activates the "knowledge missing type" retrieval strategy: combines "Model Y" with "Tesla" (from the cached context) as a targeted query "特斯拉 毛豆Y", performs an exact retrieval in the dynamic text library, hits the user discussion post "What kind of car is the Model Y that everyone is talking about" on the Autohome forum, and simultaneously recalls多篇文章 such as "Model Y Review" in the semantic channel. After the hybrid retrieval is fused, a hierarchical knowledge package containing user colloquial expression explanations and official parameters is generated.

[0218] The secondary recognition module constructs an expert role prompt to guide the DeepSeek model to analyze the content of the knowledge package: "Model Y" is the nickname given by the user to the Tesla Model Y, and "Long Range" corresponds to the official configuration "Long Range AWD". The model outputs the standardized result: "Model Y Long Range AWD", confidence 0.82. The dynamic update module detects that "Model Y" is a newly added synonym, automatically associates it with the "Model Y" entity and stores it in the database. Finally, the system returns the standard entity set for the search ranking module to use, and the recall rate is improved from 68% in the initial recognition to 97% finally.

[0219] Example 2: Construction and Update Mechanism of the Hybrid Knowledge Base System

[0220] This example details the daily operation and maintenance process of the automotive domain standard knowledge base.

[0221] In the initial knowledge base construction phase, the system extracted data from the Autohome vehicle database, the Ministry of Industry and Information Technology product announcements, and the technical specifications pages of various brand websites. After entity alignment and conflict resolution, a graph database containing 487 brands, 3260 vehicle series, and 15800 vehicle models was constructed. Each "vehicle model" node stores 12 attributes, including standard name, launch date, energy type, and body level. Simultaneously, the Sentence-BERT model was used to encode all entity names, aliases, and parameters into 768-dimensional vectors, which were then stored in the Milvus vector library to support semantic retrieval.

[0222] The dynamic update module performs regular alignment every morning at midnight: monitoring the RSS feed of the Ministry of Industry and Information Technology's announcement website, parsing the product model, manufacturer, and energy type of newly released car models, automatically creating corresponding nodes and populating attributes. For technical terms such as "plug-in hybrid" appearing in the announcements, it automatically maps them to existing standard entries in the knowledge base. For incremental updates, the system monitors the high-reliability output of the secondary recognition module in real time. When it detects that "Z9GT" (an abbreviation for Denza Z9GT) has been successfully recognized by LLM but does not exist in the knowledge base, it automatically calculates the semantic similarity between "Z9GT" and the "Denza" brand node. If the similarity reaches 0.89, exceeding the threshold, "Z9GT" is automatically added to the database as a synonym for "Denza Z9GT," and the source is marked as "LLM self-learning." If the similarity is below the threshold, it is submitted to the manual operation platform for review. Each knowledge base update generates a new version number, and old versions are retained for 30 days for retrospective queries.

[0223] Example 3: Real-time e-commerce promotion scenario enhanced by dual-channel retrieval

[0224] This embodiment describes the recognition of comment text on a car model's promotional page during the "Double Eleven" shopping festival.

[0225] The system batch-processes the comment "Bought the Han DMI at rock-bottom prices, its price-performance ratio crushes the Model 3." Initially, it outputs "Han DMI" (confidence 0.74) and "Model 3" (confidence 0.92). During mapping verification, "Model 3" directly and accurately matches the standard library. "Han DMI," through fuzzy matching, shows a similarity of 0.71 to "Han DM-i," but since this model is a 2024 model, the dynamic threshold is adjusted to 0.68 due to timeliness weighting, thus the mapping is considered successful. However, the system detects that the phrase "rock-bottom prices" implies price sensitivity. The retrieval enhancement module proactively marks the comment text as an "e-commerce promotion scenario," activating the e-commerce knowledge base retrieval channel: in addition to the regular text library, it additionally searches recent promotional articles and price trend posts, supplementing contextual knowledge such as "Double Eleven discounts" and "price reduction." After secondary recognition, the model not only confirms the vehicle model but also identifies the user intent tag "price focus" corresponding to "rock-bottom prices," providing input for subsequent financial plan recommendations. In this scenario, the retrieval enhancement module expands its knowledge source through scene awareness, enabling a leap from simple entity recognition to intent understanding.

[0226] Example 4: Edge-Cloud Collaborative Recognition in Edge Computing Scenarios

[0227] This embodiment illustrates the recognition implementation of in-vehicle voice assistants in weak network environments.

[0228] The in-vehicle device uses the lightweight TinyBERT model as the initial recognition module. The model weights are compressed to 8MB after knowledge distillation, allowing it to run in real-time on the in-vehicle NPU. When the user voice-inputs "Navigate to a charging station near the Li Auto L9 Max," the edge device recognizes "Li Auto L9 Max." However, the mapping verification fails because the in-vehicle local knowledge base only contains basic brand information, and it is marked as an entity to be identified. Due to network latency exceeding 500ms in a tunnel, the retrieval enhancement module initiates cloud-based collaborative degradation: the edge device packages the entity to be identified along with the vehicle's GPS location information into a minimum query unit and sends it to the cloud via the MQTT protocol at QoS=1. The cloud-based retrieval enhancement module prioritizes searching for "Li Auto L9 Max" test drive articles published by dealers in the area, quickly returning supplementary knowledge. After generating a standard name using the cloud-based LLM, it is sent to the in-vehicle device via the data channel. The edge device's secondary recognition module uses a lightweight verification model to confirm the cloud result. Upon successful verification, it announces "A charging station near the Li Auto L9 Max has been found." If the network is completely down, the edge device initiates a local fallback strategy: it uses cached historical recognition results for fuzzy matching, or prompts the user "The network is poor, please provide a more accurate vehicle model name." This embodiment ensures availability in resource-constrained and weak network environments through model compression, protocol optimization, and degradation mechanisms.

[0229] Example 5: Dynamic Threshold Self-Learning Adversarial Input

[0230] This embodiment describes the adaptive process by which the system defends against malicious user queries.

[0231] A user repeatedly entered "Mercedes-Benz E260", "Mercedes-Benz E260L", and "Mercedes-Benz E260Li" to probe the system's threshold. Initially, the recognition module output high confidence for each. During mapping verification, "E260" failed because it is similar to multiple models within the "Mercedes-Benz E-Class" series, resulting in a mapping failure with an initial threshold of 0.82. However, the system feedback engine detected three consecutive failed mappings of similar entities and automatically triggered threshold self-learning: analyzing the patterns of the failed entities revealed they were all variations of "brand + model + letter suffix," consistent with Mercedes-Benz model naming conventions. The self-learning algorithm dynamically lowered the threshold for vehicle series entities in this pattern to 0.75 and temporarily increased the user's query priority. Subsequent input of "Mercedes-Benz E300L" resulted in successful mapping verification due to the threshold adjustment. Simultaneously, the system recorded the user's behavior pattern. If more than 20 probing queries were detected within 24 hours, the request was routed to a sandbox environment to limit its impact on the main knowledge base updates, protecting system stability. This mechanism enables the system to proactively adapt and defend against adversarial input.

[0232] Example 6: Cold Start of New Entities and Rapid Expansion of Knowledge Base

[0233] This example describes the cold start of the knowledge base when a new brand launches its first model.

[0234] A new brand, "Interstellar Automotive," launched its first model, "Interstellar Voyage," in April 2025. Within 24 hours of the launch, numerous search queries for "Interstellar Voyage price" and "Interstellar Voyage range" appeared. The initial recognition module, due to its training corpus not covering the brand, identified "Interstellar Voyage" as a common word, resulting in a failed mapping verification. After the retrieval enhancement module retrieved multiple media reports, the secondary recognition module successfully identified the entity as a highly reliable result. The dynamic update module detected that this entity was a combination of a new brand and a new model, representing a new addition to the knowledge base. It immediately triggered a cold start acceleration process: automatically creating a brand root node for the entity, automatically filling in the vehicle class and energy type attributes based on the description of "mid-to-large-sized pure electric SUV" in the articles, and automatically associating the initially identified synonyms "Interstellar Voyage" and "Voyage SUV." Due to the noise and confusion surrounding the initial launch of a new car, the system implemented a temporary confidence decay: within 7 days of the entity's inclusion in the database, its mapping threshold was automatically increased by 0.1 to prevent knowledge base contamination due to early recognition errors. After 7 days, based on the success rate statistics, if the success rate is >90%, the attenuation is lifted and the entity is converted into a formal entity. This process enables the rapid improvement of the knowledge base from scratch, shortening the new entity entry cycle that traditionally required several weeks of manual operation to within 2 hours.

[0235] Example 7: Knowledge Isolation and Sharing in Multi-Tenant Scenarios

[0236] This embodiment describes a scenario where multiple sub-brands within an automotive group share a common identification system.

[0237] A certain automotive group comprises "Main Brand A" and "High-end Sub-brand B." While their vehicle data is business-isolated, their technical terminology is shared. During system deployment, the dynamic update module creates a multi-tenant knowledge base view: the Main Brand A's knowledge base is a complete view, containing all brand entities within the group; the Sub-brand B's knowledge base is a restricted view, containing only the B brand and shared technical entities. The initial identification module automatically loads the corresponding view based on the tenant ID of the query source, ensuring that user queries do not involve cross-brand mapping. During mapping verification, shared technical terms (such as "ternary lithium battery" and "LiDAR") can be found in both views, but brand-specific entities are strictly isolated. When Sub-brand B releases a new model, the dynamic update module adds the entity to the restricted view and automatically synchronizes the entity's technical parameter nodes to the shared area of ​​the Main Brand A view, achieving "data isolation, knowledge sharing." The retrieval enhancement module performs partitioned retrieval based on the view ID, and the expert role prompts generated by the secondary identification module also vary by tenant: users querying Sub-brand B use the "Luxury Brand Product Manager" role, while users querying Main Brand A use the "Mainstream Brand Product Manager" role. This architecture maximizes knowledge reuse and reduces overall group operating costs while ensuring data security.

[0238] Example 8: Generation of an interpretability report for identification results

[0239] This embodiment illustrates the diagnostic and analysis functions provided by the system for operators.

[0240] When operations personnel discover a decline in the recognition accuracy of the "Denza D9" model within a certain time period, they can specify the time range and entity ID through the diagnostic interface. The system extracts all relevant recognition records for that period from the feedback repository and automatically generates an interpretability report. The report includes three parts: first, attribution analysis, which statistically analyzes the distribution ratio of "outdated knowledge base version," "colloquial user input," and "insufficient contextual information" in recognition failure cases; second, evidence chain visualization, which highlights the source document fragments cited and the confidence calculation process for each LLM decision of the secondary recognition module; and third, optimization suggestions, which suggest supplementing data sources from specific vertical media if the knowledge base deficiency is the main cause, and adjusting the dynamic threshold baseline for a certain type of entity if the threshold setting is too strict. The report also includes an impact assessment, which calculates the quantitative impact of the recognition error on downstream business such as search CTR and lead conversion rate, helping operations personnel prioritize high-value issues. This function transforms the black-box recognition process into an observable, interventionable, and optimizable white-box system.

[0241] By adopting the above-disclosed technical solution of this invention, the following beneficial effects are obtained:

[0242] Effectively solves the problem of missed identification: For pending entities that the initial identification model cannot map to the standard knowledge base, an automatic retrieval enhancement mechanism is triggered. By introducing a real-time text library in the automotive field as a supplementary knowledge source, the recognition recall rate of uncommon car models, newly released car series, and colloquial expressions is significantly improved, avoiding the defect of termination upon recognition failure in traditional solutions.

[0243] Enhanced disambiguation capabilities: By employing a large language model combined with contextual information and retrieval candidate entities for disambiguation, it can deeply understand the true intent of the user's query and significantly reduce the misjudgment rate of polysemous words.

[0244] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying entities in the field of automobiles based on a large language model, characterized in that, Includes the following steps: Receive the text to be processed, which contains named entities in the automotive domain; Candidate entities are extracted from the text to be processed using a first recognition model; The candidate entities are mapped and matched with the standard knowledge base in the automotive field, and candidate entities that cannot be successfully mapped are marked as pending entities; When the undetermined entity exists, relevant documents are retrieved from the dynamic text library in the automotive field as supplementary knowledge based on the text to be processed. The entity to be identified and the supplementary knowledge are input into a large language model. Secondary entity recognition is performed through the large language model to determine the standard name of the entity to be identified, and the final entity recognition result is output.

2. The method of claim 1, wherein, The automotive industry standard knowledge base is a hybrid knowledge base system, including: A structured database used to store standard names, synonyms, and hierarchical relationships for car brands, car series, and car models; An entity vector library is used to store the semantic representation vectors of the standard names and synonyms.

3. The method of claim 2, wherein, The relevant documents retrieved include: Based on the entity name of the entity to be determined, an exact match query is performed in the structured database to obtain the first query result; Based on the contextual semantic features of the entity to be determined in the text to be processed, a similarity search is performed in the entity vector database to obtain a second query result. The supplementary knowledge is generated by merging the first query result and the second query result.

4. The method of claim 2, wherein, The aforementioned dynamic text library in the automotive field is a real-time updated database of automotive news articles, which is constructed in the following manner: We extract articles from automotive vertical media platforms that include new car launches, model reviews, and technical analyses. Incremental indexing is performed on the articles, and retrieval weights are assigned based on publication time; The hybrid knowledge base system is dynamically updated in the following ways: Obtain the newly added entities from the secondary recognition output of the large language model; When the semantic similarity between the newly added entity and an existing entity in the knowledge base is greater than a preset threshold, the newly added entity is added to the knowledge base system as a synonym. The system monitors external automotive data sources and automatically extracts entities and updates the knowledge base system when a new car model or series is detected.

5. The method of claim 1, wherein, Marking candidate entities that cannot be successfully mapped as pending entities includes: Calculate the semantic similarity between the candidate entity and the standard name in the standard knowledge base; When the semantic similarity is lower than the dynamic threshold, the mapping is deemed to have failed. The dynamic threshold is adaptively adjusted based on the recognition confidence of the candidate entity and the historical mapping success rate.

6. The method of claim 1, wherein, Inputting the undetermined entity and the supplementary knowledge into the large language model includes: Construct preset prompt words, which include the entity to be determined, the context information of the text to be processed, and the list of candidate entities in the supplementary knowledge; The preset prompt words are used to guide the large language model to perform disambiguation recognition of the entity to be determined based on the context information; The preset prompt words are also used to guide the large language model to output a normalized entity representation, which includes a standard entity name, entity type encoding, and confidence score.

7. The method of claim 1, wherein, Also includes: The number of entities whose standard names cannot be determined after the aforementioned secondary entity recognition is counted. When the number exceeds the tolerance threshold, the corresponding text to be processed and the recognition result are recorded in the manual review queue for the purpose of optimizing the first recognition model and the hybrid knowledge base system.

8. A large language model-based automobile field entity recognition system, characterized in that, include: The initial recognition module is used to extract candidate entities from the text to be processed through the first recognition model; The mapping verification module is used to map and match the candidate entities with the standard knowledge base in the automotive field, and mark the candidate entities that cannot be mapped as pending entities; The retrieval enhancement module is used to retrieve relevant documents from the dynamic text library in the automotive field as supplementary knowledge when the entity to be identified exists. The secondary recognition module is used to input the entity to be identified and the supplementary knowledge into the large language model for secondary entity recognition and output the final entity recognition result. The dynamic update module is used to automatically update the standard knowledge base based on the recognition results; The system performs the method as described in any one of claims 1 to 7.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the computer program comprises instructions for: When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7. 10.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images