Iron and steel e-commerce platform customer information identification method and system based on bert and large model

Abstract

This invention relates to a method and system for identifying customer information on a steel e-commerce platform based on BERT and a large-scale model. The method includes the following steps: acquiring existing end-customer data; generating filtering conditions based on the existing end-customer data using a large-scale model; obtaining a list of potential end-customer companies based on the filtering conditions; processing the list of potential end-customer companies using a pre-trained and fine-tuned BERT model to obtain company classification results and address resolution results; adding company type tags and address tags to each record in the list of end-customer companies; constructing multiple end-customer clusters based on the company type tags and address tags; obtaining the main business of the end-customer clusters using a large-scale model; and generating a steel demand profile for the corresponding end-customer clusters by combining the steel knowledge base. Compared with existing technologies, this invention has the advantages of high efficiency and high accuracy.

Description

Technical Field

[0001] This invention relates to the field of steel e-commerce customer data processing technology, and in particular to a method and system for identifying customer information on steel e-commerce platforms based on BERT and large models. Background Technology

[0002] Traditional e-commerce platforms primarily rely on user behavior data (such as click-through rates and browsing history) for marketing and customer acquisition, using user profiling and recommendation algorithms to achieve precise marketing. However, steel, as a typical bulk commodity, has trading platforms targeting corporate clients in the steel trading industry. The supply chain involves multiple entities, including end-user companies, traders, steel mills, processing centers, and storage and transportation companies. Among these, end-user companies, as the final users of steel, have purchasing behavior strongly correlated with actual production needs, exhibiting planning and stability, making them the core value customers of steel e-commerce platforms. Traders' purchases are more short-term, while steel mills and storage and transportation companies are supporting links in the supply chain. Therefore, steel e-commerce platforms need to establish a precise targeting mechanism for end-user companies to achieve sustainable business development.

[0003] Existing technologies, such as patent application CN119648330A, propose a method and system for agent interaction on a steel e-commerce platform based on a large model. First, user, product, and market information is collected, cleaned, structured, and enhanced to construct a knowledge graph integrating user characteristics and steel product attributes. Second, prompts for the large model are designed and fine-tuned across multiple scenarios using LoRA. Finally, intelligent shopping guidance is implemented based on the knowledge graph and the fine-tuned large model: rewriting user questions, calling the agent to obtain recommendation results, and proactively generating guiding questions to refine user needs. This method aims to improve the efficiency and industry adaptability of personalized shopping guidance in steel e-commerce. However, this method requires multiple calls to a large language model (LLM) in the question rewriting, agent scheduling, and guiding message generation stages. While large language models possess powerful text understanding capabilities, they have limitations in terms of computational efficiency and cost-effectiveness. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a highly efficient and accurate method and system for identifying customer information on steel e-commerce platforms based on BERT and large models.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for identifying customer information on a steel e-commerce platform based on BERT and a large model includes the following steps: Acquire existing end-customer data, generate filtering conditions based on the existing end-customer data using a large model, and obtain a list of potential end-customer enterprises based on the filtering conditions. The potential terminal enterprise list is processed using a pre-trained and finely tuned BERT model to obtain enterprise classification results and address resolution results. Enterprise type labels and address labels are added to each record in the terminal enterprise list. Based on the enterprise type tags and address tags, the list of terminal enterprises is constructed into multiple terminal clusters; The main business of the terminal cluster is obtained through a large model, and a steel demand profile for the corresponding terminal cluster is generated by combining it with a steel knowledge base.

[0006] Furthermore, the filtering criteria include at least a list of cities and keywords related to the scope of business.

[0007] Furthermore, a list of potential end-user companies is obtained by combining the aforementioned screening criteria and fixed criteria.

[0008] Furthermore, the enterprise types include end-user, trading, storage and transportation, non-steel, and unspecified; The address label contains information including at least one of the following: province, city, district, town, industrial cluster, road, and house number.

[0009] Furthermore, the acquisition of training data for the pre-training fine-tuning of the BERT model includes: Obtain manually labeled data; Based on the manually labeled data and random input data, a large model is used to process the random input data to obtain the processing result. Based on the processing result, the manually labeled data and random input data are balanced to obtain fine-tuned sample data, which is used as the training data.

[0010] Furthermore, during the pre-training fine-tuning process of the BERT model, the training data is adjusted and optimized based on the recall performance of the BERT model after fine-tuning.

[0011] Furthermore, the adjustment and optimization of the training data specifically includes: Based on the recall situation, samples with confidence levels within the set range but containing errors are selected as candidates for difficult-to-bear examples. The hard-negation candidate examples are added to the training data, and data augmentation is performed simultaneously. Retrain using the training data with the added data.

[0012] Furthermore, the steel knowledge base includes at least a variety of materials such as steel mill standard documents, steel mill product manuals, grade dictionary manuals, steel quality dispute case analyses, and popular science articles on steel.

[0013] Furthermore, the existing terminal customer data is dynamically updated.

[0014] This invention also provides a customer information identification system for a steel e-commerce platform based on BERT and a large model, comprising: The potential information filtering module is used to acquire existing end-customer data, generate filtering conditions based on the existing end-customer data through a large model, and obtain a list of potential end-customer enterprises based on the filtering conditions. The enterprise information coarse screening module is used to process the potential terminal enterprise list using a pre-trained and finely tuned BERT model to obtain enterprise classification results and address resolution results, and to add enterprise type labels and address labels to each record in the terminal enterprise list; The terminal cluster construction module is used to construct multiple terminal clusters from the list of terminal enterprises based on the enterprise type label and address label; The enterprise information analysis module is used to obtain the main business of the terminal cluster through a large model, and generate a steel demand profile for the corresponding terminal cluster by combining it with a steel knowledge base.

[0015] This invention innovatively adopts a hierarchical collaborative architecture of BERT and LLM. It utilizes the BERT model to efficiently complete structured processing tasks such as enterprise classification and address resolution, and then leverages LLM and domain knowledge augmentation (RAG) to achieve semantic reasoning and cluster profile generation. This significantly improves system efficiency while ensuring analytical accuracy, providing a technologically advanced and commercially viable solution for customer mining in the steel e-commerce industry. Compared with existing technologies, this invention has the following beneficial effects: (1) Collaborative application of BERT and large models: Traditional methods rely on manual rules or single models. This invention innovatively employs a funnel-shaped processing architecture combining BERT coarse screening and LLM fine analysis. It utilizes a finely tuned BERT model to handle address resolution and classification tasks for massive numbers of enterprises, only calling the large model for profile generation on a small number of selected core clusters. This invention addresses the efficiency limitations of large models through a collaborative model architecture and the accuracy limitations of general models through knowledge augmentation technology, thus forming a complete technical solution with industrial application value.

[0016] Using a purely large-scale model approach, processing a single enterprise classification takes approximately 1 second; however, using the collaborative architecture of this invention, the average processing time for 100 data entries is only 1.7 seconds. This data demonstrates that this solution improves processing efficiency by nearly 60 times while maintaining accuracy, solving the timeliness problem of large-scale models being unable to be implemented in industrial-grade massive data processing, and improving data mining efficiency and accuracy.

[0017] (2) Dynamic terminal filtering conditions: This invention analyzes existing terminal data through a large model, automatically generates dynamic filtering conditions (such as city range and business scope keywords), and feeds them back to the database for filtering. This technical process of semantic feature extraction, structured query, and result feedback enables adaptive updating of mining rules. Unlike traditional technologies, which rely on fixed rules that cannot adapt to market changes, this invention avoids the limitations of fixed rules.

[0018] (3) Intelligent prediction of steel demand: This invention connects a large-scale intelligent model to a steel knowledge base, automatically matching manufacturing products with steel requirements to generate accurate cluster profiles and assist in manual research. The invention features an external steel knowledge base (including steel mill standards, grade dictionaries, and quality cases). Through keyword engineering, it forces the large model to generate technical parameter tables (such as yield strength and elongation) solely based on the retrieved knowledge base content, and strictly labels the data sources. This invention's combination of professional database retrieval and logical reasoning effectively avoids the illusion of industrial parameters inherent in general-purpose large models, ensuring that the output steel grades and process parameters conform to industrial standards.

[0019] (4) Reduce customer acquisition costs: This invention reduces manual intervention, shortens the customer acquisition cycle, reduces the workload of manual research and data analysis, lowers the marginal cost of customer acquisition for enterprises, and improves the accuracy of customer acquisition through automated screening and classification.

[0020] (5) The model has strong scalability: This invention can migrate the solution to other customer mining scenarios by adjusting the knowledge base and model fine-tuning the data, such as mining other types of customers like machining centers, or mining specific product customers like gear steel and bearing steel. Attached Figure Description

[0021] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0023] Example 1 This embodiment provides a method for identifying customer information on a steel e-commerce platform based on BERT and a large model, such as... Figure 1 As shown, it includes the following steps: Step S1: Obtain existing end-customer data, generate filtering conditions based on the existing end-customer data using the large model LLM, obtain a list of potential end-customer companies based on the filtering conditions, and form a pool of companies to be explored. Step S2: Process the potential terminal enterprise list using the pre-trained and fine-tuned BERT (Bidirectional Encoder Representations from Transformers) model to obtain enterprise classification results and address resolution results, and add enterprise type labels and address labels to each record in the terminal enterprise list; Step S3: Construct the terminal enterprise list into multiple terminal clusters based on the enterprise type label and address label; Step S4: Obtain the main business of the terminal cluster through the large model, and generate a steel demand profile for the corresponding terminal cluster by combining the steel knowledge base.

[0024] The above method adopts a collaborative architecture of BERT and LLM (such as the GPT series). It uses a fine-tuned BERT model to handle the address resolution and classification tasks of massive enterprises, and only calls the large model to generate profiles for a small number of core clusters, which solves the efficiency defects of the large model. At the same time, it combines a steel knowledge base and solves the accuracy defects of the general model through knowledge augmentation technology, which can efficiently and accurately realize customer mining of steel e-commerce platforms.

[0025] The specific explanation of the above method is as follows: (1) Pre-screening of the pool of companies to be explored Input: Existing end-customer data.

[0026] Specifically, existing end-customer data may include business registration information (fields such as company name, company address, business scope, etc.), transaction information (fields such as transaction type, transaction amount, etc.), and marketing information (content used to assist in judgment, such as visit records and manual judgment conclusions).

[0027] Processing flow: a. Large-scale model analysis of existing end-customer data: Extract city lists, business scope keywords, and other conditions to generate filtering conditions (dynamic conditions), and cross-validate them with the purchased products.

[0028] b. Database query: Combine the filtering conditions generated by the large model to filter the list of companies from the enterprise database.

[0029] The enterprise database includes the enterprise's business registration information, such as enterprise name, operating status, registration date, enterprise address, business scope, and contact information.

[0030] Output: A list of potential end-user companies, including company name, type, legal representative, address, registration date, business scope, contact number, and other information.

[0031] The generated filter criteria should include at least a list of cities and keywords related to the scope of business.

[0032] In another embodiment, a list of potential end-user companies is obtained by combining the aforementioned screening criteria and fixed criteria. The fixed criteria include the end-user industry, years of operation, and operating status. In this embodiment, the years of operation are 5 or more, and the operating status is normal, such as existing or in operation.

[0033] (2) Terminal type determination and address resolution Input: Company name, company registered address.

[0034] Processing flow: a. Enterprise type determination: The fine-tuned BERT (bert_base_chinese) model is used to classify enterprises. Enterprise types include end-user, trade, storage and transportation, non-steel, and unknown.

[0035] End-user enterprises: Enterprises that use finished products from steel mills, whose products are significantly different from the original output products of steel mills. This includes, but is not limited to, steel material manufacturers, project owners and construction units, and processing centers with the ability to cut and grind steel coils and plates.

[0036] Traders: Trading companies that do not use steel mill products but aim to profit from reselling steel.

[0037] Storage and transportation companies: Enterprises whose main business is cargo storage and transportation.

[0038] Unknown: The company's business description is related to steel and metals, but it is not easy to determine its position in the steel trade.

[0039] Non-steel enterprises: Enterprises whose main business does not require the purchase of steel, other than those mentioned above.

[0040] b. Enterprise Address Resolution: Using a fine-tuned BERT model, entities are identified from enterprise addresses, and the administrative division (province, city, district, town, street, village, community, etc.), industrial zone (development zone, pilot zone, etc.), road, and house number are labeled. In the address strings of the fine-tuned samples, related words (XX development zone, XX industrial zone, XX pilot zone) are labeled as the same category.

[0041] Output: Enterprise type label (terminal, trade, storage and transportation, non-steel, unknown), enterprise address label (province, city, district, town, industrial cluster, road, house number, etc.).

[0042] The acquisition of training data for the pre-training and fine-tuning of the BERT model mentioned above includes: Obtain manually labeled data; Based on the manually labeled data and random input data, a large model is used to process the random input data to obtain the processing result. Based on the processing result, the manually labeled data and random input data are balanced to obtain fine-tuned sample data, which is used as the training data.

[0043] Preferably, during the pre-training fine-tuning of the BERT model, the training data is adjusted and optimized based on the recall performance of the fine-tuned BERT model. The adjustment methods include: (1) increasing the number of samples for difficult or easily confused examples by means of oversampling, synonym supplementation, etc.; (2) correcting incorrectly labeled examples.

[0044] Furthermore, adjusting and optimizing the training data specifically includes: Based on the recall situation, samples with confidence levels within the set range but containing errors are selected as candidates for difficult-to-bear examples. The difficult negative example candidates are added to the training data, and data augmentation is performed simultaneously, including synonym replacement, specification value transformation, etc. Retrain using the expanded training data, reassess the recall performance, and confirm whether readjustment and optimization are necessary.

[0045] Through the fine-tuning described above, the BERT model can improve its processing performance in specific scenarios, enhancing both customer acquisition efficiency and accuracy. Taking enterprise type classification as an example, the acquisition of training data for fine-tuning includes the following steps: a. Based on business visit experience, mark the types of companies with the names of about 200 companies. The marking methods include: (1) judging from the visit and survey records of marketing personnel, and (2) judging by business experts. b. Build an intelligent agent, set the context to the standard definition of enterprise type and 200 manually labeled data, and ask the large model to analyze the input enterprise name and determine the enterprise type through prompt words; c. Randomly select 10,000 company names and use an intelligent agent to determine the company type; d. Perform sample balancing on the data generated by the agent's judgment to finally obtain the fine-tuned sample data for BERT. The sample balancing adopts oversampling and appropriately reduces the number of non-steel enterprises. e. Based on the recall performance after BERT fine-tuning, adjust and optimize the training data, with a recall target of 90%.

[0046] (3) Terminal cluster determination Input: Enterprise type label, enterprise address label (administrative division below the county level, industrial zone established in urban construction).

[0047] Processing flow: a. Filter companies whose company type label is "terminal".

[0048] b. Use administrative divisions below the county level (such as towns, streets, villages, communities, etc.) and industrial zones (development zones, pilot zones) as cluster names, and summarize the number of enterprises by name.

[0049] Output: List of enterprise clusters (cluster name, number of enterprises).

[0050] Note: Based on industry experience, the clustering of manufacturing enterprises can be broadly divided into two categories: naturally formed and government-guided. a. Naturally formed: Manufacturing enterprises often cluster together due to reasons such as partnerships between relatives and neighbors, resulting in economies of scale. They are generally clustered within village or town-level administrative divisions in second-tier cities. These clusters use the administrative division name (e.g., XX town, XX village) as the cluster name, and clusters with more than 300 enterprises under the same tag will be selected.

[0051] b. Government-guided: When the government is developing cities, it attracts manufacturing companies to set up factories by planning economic development zones and industrial zones, based on considerations of economies of scale. These clusters use the name of the industrial zone as the cluster name, and clusters with more than 150 companies under the same tag will be selected.

[0052] (4) Generate cluster profiles from large models Input: Terminal cluster list (company name, business scope).

[0053] Processing flow: a. Enterprise sampling analysis: Randomly select no more than 100 enterprises from the terminal cluster for analysis.

[0054] b. Extraction of main business: By analyzing enterprise information through a large model intelligent agent, including enterprise name, national economic classification, and business scope, the main business of the enterprise is extracted and its manufactured products are determined.

[0055] c. Steel demand matching: Combine the steel knowledge base to query the material selection requirements for manufacturing products.

[0056] d. Cluster Profile Generation: Integrate and analyze manufactured products and required steel grades to generate cluster profiles and provide marketing guidance, such as recommending steel grades.

[0057] Outputs: Cluster profile report (manufactured products, steel demand, marketing suggestions), cluster enterprise directory (enterprise business registration information, manufactured products, and contact information).

[0058] In a preferred embodiment, existing end-customer data is updated in real time, thereby dynamically updating the screening criteria, avoiding the limitations of fixed rules, and meeting the requirements of real-time market changes.

[0059] If the above methods are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0060] Example 2 This embodiment provides a customer information identification system for a steel e-commerce platform based on BERT and a large model, such as... Figure 2 As shown, it includes: Potential information filtering module 1 is used to acquire existing end-customer data, generate filtering conditions based on the existing end-customer data through a large model, and obtain a list of potential end-customer enterprises based on the filtering conditions. The enterprise information coarse screening module 2 is used to process the potential terminal enterprise list using a pre-trained and finely tuned BERT model to obtain enterprise classification results and address resolution results, and to add enterprise type labels and address labels to each record in the terminal enterprise list; Terminal cluster construction module 3 is used to construct multiple terminal clusters from the list of terminal enterprises based on the enterprise type label and address label; The Enterprise Information Analysis Module 4 is used to obtain the main business of the terminal cluster through a large model, and generate a steel demand profile for the corresponding terminal cluster by combining it with the steel knowledge base.

[0061] The rest is the same as in Example 1.

[0062] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for identifying customer information on a steel e-commerce platform based on BERT and a large model, characterized in that, Includes the following steps: Acquire existing end-customer data, generate filtering conditions based on the existing end-customer data using a large model, and obtain a list of potential end-customer enterprises based on the filtering conditions. The potential terminal enterprise list is processed using a pre-trained and finely tuned BERT model to obtain enterprise classification results and address resolution results. Enterprise type labels and address labels are added to each record in the terminal enterprise list. Based on the enterprise type tags and address tags, the list of terminal enterprises is constructed into multiple terminal clusters; The main business of the terminal cluster is obtained through a large model, and a steel demand profile for the corresponding terminal cluster is generated by combining it with a steel knowledge base.

2. The method for identifying customer information on a steel e-commerce platform based on BERT and a large model as described in claim 1, characterized in that, The filtering criteria include at least a list of cities and keywords related to the scope of business.

3. The method for identifying customer information on a steel e-commerce platform based on BERT and a large model as described in claim 1, characterized in that, A list of potential end-user companies is obtained by combining the aforementioned screening criteria and fixed criteria.

4. The method for identifying customer information on a steel e-commerce platform based on BERT and a large model as described in claim 1, characterized in that, The types of enterprises mentioned include end-user, trading, storage and transportation, non-steel, and unspecified. The address label contains information including at least one of the following: province, city, district, town, industrial cluster, road, and house number.

5. The method for identifying customer information on a steel e-commerce platform based on BERT and a large model according to claim 1, characterized in that, The acquisition of training data for the pre-training fine-tuning of the BERT model includes: Obtain manually labeled data; Based on the manually labeled data and random input data, a large model is used to process the random input data to obtain the processing result. Based on the processing result, the manually labeled data and random input data are balanced to obtain fine-tuned sample data, which is used as the training data.

6. The method for identifying customer information on a steel e-commerce platform based on BERT and a large model according to claim 1, characterized in that, During the pre-training fine-tuning process of the BERT model, the training data is adjusted and optimized based on the recall performance of the BERT model after fine-tuning.

7. The method for identifying customer information on a steel e-commerce platform based on BERT and a large model as described in claim 6, characterized in that, The adjustment and optimization of the training data specifically includes: Based on the recall situation, samples with confidence levels within the set range but containing errors are selected as candidates for difficult-to-bear examples. The hard-negation candidate examples are added to the training data, and data augmentation is performed simultaneously. Retrain using the training data with the added data.

8. The method for identifying customer information on a steel e-commerce platform based on BERT and a large model according to claim 1, characterized in that, The steel knowledge base includes at least a variety of documents from steel mills, such as standard documents, product manuals, grade dictionary manuals, case analyses of steel quality disputes, and popular science articles on steel.

9. The method for identifying customer information on a steel e-commerce platform based on BERT and a large model according to claim 1, characterized in that, The existing terminal customer data is dynamically updated.

10. A customer information identification system for a steel e-commerce platform based on BERT and a large model, characterized in that, include: The potential information filtering module is used to acquire existing end-customer data, generate filtering conditions based on the existing end-customer data through a large model, and obtain a list of potential end-customer enterprises based on the filtering conditions. The enterprise information coarse screening module is used to process the potential terminal enterprise list using a pre-trained and finely tuned BERT model to obtain enterprise classification results and address resolution results, and to add enterprise type labels and address labels to each record in the terminal enterprise list; The terminal cluster construction module is used to construct multiple terminal clusters from the list of terminal enterprises based on the enterprise type label and address label; The enterprise information analysis module is used to obtain the main business of the terminal cluster through a large model, and generate a steel demand profile for the corresponding terminal cluster by combining it with a steel knowledge base.

Images