Enterprise industry chain identification method and system based on large language model

Abstract

The invention relates to an enterprise industry chain identification method and system based on a big language model, and the method comprises the steps: obtaining target enterprise information based on a big data analysis technology; dividing a plurality of relatively independent industry chain links for the industry to be analyzed; for each divided industrial chain link, obtaining a descriptive text; based on text matching, labeling the target enterprise information with an affiliated industry chain link label to obtain enterprise industry chain link affiliation labeling information; constructing a training data set and a verification data set based on the descriptive text of each industry chain link, enterprise industry chain link affiliation labeling information and task cues determined according to task requirements and a model optimization strategy; and inputting the training data set and the verification data set into a generative large language model for model training to obtain an exclusive large language model for enterprise industry chain identification. The implementation of the method can realize accurate identification of the enterprise industry chain.

Description

Technical Field

[0001] The present invention relates to the technical field of large language models and natural language processing, and in particular to a method and system for identifying an enterprise industrial chain based on a large language model. Background Art

[0002] With the acceleration of globalization and the increasingly fierce market competition, enterprises need to understand their position in the industrial chain more accurately in order to formulate effective market competition strategies. In order to reduce costs and improve efficiency, enterprises need to manage the supply chain in a refined manner. Therefore, industrial chain identification is an important prerequisite for achieving supply chain optimization. At present, the methods used to achieve industrial chain identification include traditional industry analysis reports, market research, enterprise interviews, etc. Although these methods can provide a certain degree of industrial chain information, they can help enterprises to initially understand their position in the industrial chain and their main competitors. However, these methods often rely on regularly released market reports or industry analysis, which will lead to slow information updates and make it difficult to meet the needs of enterprises for rapid decision-making. With the rapid development of big data and artificial intelligence technologies, it is necessary to develop and apply industrial chain identification methods and systems based on big data and artificial intelligence technologies to achieve real-time, comprehensive, and accurate identification of industrial chain information. Summary of the invention

[0003] In order to achieve real-time, comprehensive, and accurate recognition of industrial chain information, the purpose of the present invention is to provide a method and system for enterprise industrial chain recognition based on a large language model. The technical solutions adopted are as follows: In a first aspect, the present application discloses a method for identifying an enterprise industrial chain based on a large language model, the method comprising: S1. Obtain target enterprise information based on big data analysis technology; S2. For the industry to be analyzed, divide the industry chain into multiple relatively independent links; S3. Obtain corresponding descriptive text for each segment of the industrial chain; S4. Based on text matching, label the target enterprise information with the industrial chain link label to which it belongs, and obtain the enterprise industrial chain link attribution labeling information; S5. Construct a training data set and a validation data set based on the descriptive text of each industrial chain link, the attribution annotation information of the enterprise industrial chain link, and the task prompt words determined according to the task requirements and model optimization strategy; S6. Input the training data set and the verification data set into a generative large language model for model training to obtain a dedicated large language model for identifying the enterprise industrial chain.

[0004] Furthermore, in step S1, the target enterprise information includes at least one of the industry classification to which the enterprise belongs, the business scope of the enterprise, the abstract of the core patents held by the enterprise, the name of the enterprise's marketing activities or projects, the name of the software works owned by the enterprise, the job titles owned by the enterprise, and the product introduction of the enterprise.

[0005] Furthermore, in step S1, the acquisition of target enterprise information based on big data analysis technology includes: S11. Obtain initial enterprise information based on big data analysis technology; S12, based on a preset data cleaning rule, when it is determined that the initial enterprise information contains at least one abnormal data of missing data, noise data, and duplicate data, performing data cleaning processing on the initial enterprise information to obtain enterprise cleansing information; S13, performing standardization processing based on the enterprise cleaning information to obtain enterprise standardized information; S14. Based on natural language processing technology, remove text-irrelevant information and / or redundant description noise in the enterprise standardized information to obtain target enterprise information.

[0006] Furthermore, in step S5, the task prompt words include: The task background is set according to the task requirements; Reference examples set based on the annotated information on the industry chain links of the annotated enterprises; A description of the priorities set based on the importance of key business attributes and industry regulations; Whitelists and blacklists are set based on the accuracy of model recognition and the need to avoid misjudgment.

[0007] Furthermore, in step S5, the training data set and the verification data set are both lists in JSON array format, each element in the list is a JSON object, and each JSON object contains a first key-value pair corresponding to the task prompt word, a second key-value pair corresponding to the expected input of the model, and a third key-value pair corresponding to the expected output of the model.

[0008] Furthermore, in step S6, during the model training process, the method further includes: S61. In the initial parameter setting of the model, set the learning rate to 5e-5, the batch size to 100, the regularization coefficient to 0.01, and the number of model fine-tuning layers to the last two layers; S62, based on the initial parameter settings and the training data set, perform the first round of iterative training, and during the training process, monitor the performance indicators of the model, and after the training is completed, when it is determined based on the verification data set that the recognition accuracy of the model for any type of enterprise industrial chain is lower than a preset threshold, adjust the learning rate by reducing it by one tenth, adjust the regularization coefficient by increasing it by two times, and increase one or two layers of model fine-tuning layers according to the specific structure and performance feedback of the model, and start the second round of iterative training; S63. In each round of iteration, the process of training, evaluation, and parameter adjustment is repeated. Through multiple rounds of iterative training, the parameter settings of the model are continuously optimized so that the model focuses on the task of identifying the enterprise industrial chain.

[0009] Furthermore, based on the model parameter setting, the method further includes: S64, collecting target cases predicted incorrectly by the model, and recording features and labels of the target cases; S65. Based on the characteristics and labels of the target case, when it is determined that the data has category imbalance, noisy data or missing values, oversampling or undersampling techniques are used to balance data categories, fill missing values, use data enhancement techniques to increase data diversity, and optimize prompt word design to help the model gradually understand and master task requirements; S66. When it is determined that the model has an overfitting or underfitting problem based on the characteristics and labels of the target case, the model parameter settings are optimized to ensure that the accuracy and robustness of the model are improved, and after completing the parameter fine-tuning, the model weights are saved.

[0010] Furthermore, before fully automated reasoning is performed on millions of pieces of enterprise information based on the exclusive large language model, the method further includes: S7. Based on the specified industry classification standards, preliminary industry classification and matching are performed on the information of the millions of enterprises to determine the industry category to which each enterprise belongs; S8. Based on the industry classification and matching results, taking into account at least one key factor including enterprise size, business scope, and market position, select enterprise information that is representative and can reflect the characteristics of the industry as sample data for inference analysis.

[0011] In the second aspect, the present application discloses a system for identifying an enterprise industrial chain based on a large language model, the system comprising an enterprise information acquisition module, an industrial chain segment division module, an industrial chain segment descriptive text generation module, an enterprise information annotation module, a training data construction module, and a model training module, wherein: The enterprise information acquisition module is used to acquire target enterprise information based on big data analysis technology; The industrial chain segment division module is used to divide the industry to be analyzed into multiple relatively independent industrial chain segments; The industrial chain link descriptive text generation module is used to obtain corresponding descriptive text for each divided industrial chain link; The enterprise information labeling module is used to label the target enterprise information with the industry chain link label based on text matching to obtain the enterprise industry chain link attribution labeling information; The training data construction module is used to construct a training data set and a verification data set based on the descriptive text of each industrial chain link, the attribution labeling information of the enterprise industrial chain link, and the task prompt words determined according to the task requirements and the model optimization strategy; The model training module is used to input the training data set and the verification data set into the generative large language model for model training, so as to obtain an exclusive large language model for identifying the enterprise industrial chain.

[0012] In a third aspect, the present application discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the enterprise industrial chain identification method based on a large language model.

[0013] The present invention has the following beneficial effects: 1) Based on the descriptive text of the industrial chain links, the labeling information of the enterprise industrial chain links, and the task prompt words, the training data set and the verification data set are constructed. Such a data set not only contains rich semantic information, but also is optimized according to the task requirements, which is helpful for training high-performance models; 2) Input the training data set and the validation data set into the generative large language model for training, and finally obtain a dedicated large language model for enterprise industrial chain identification. This model can accurately understand the relationship between enterprise information and industrial chain links, and has a strong generation capability, which can provide personalized industrial chain identification services for enterprises; 3) Dividing the industrial chain into several relatively independent links based on its actual operation will help to gain a deeper understanding.

[0014] The composition and operation mechanism of the industrial chain provide a clear framework for subsequent enterprise information annotation and model training. In addition, descriptive texts are obtained for each industrial chain link. These texts contain key features, upstream and downstream relationships, functions and other information of the industrial chain links, providing rich semantic and contextual information for model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 A method flow chart of a method for identifying an enterprise industrial chain based on a large language model provided by an embodiment of the present invention; Figure 2 A system structure diagram of an enterprise industry chain identification system based on a large language model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the enterprise industry chain identification method and system based on a large language model proposed by the present invention, its specific implementation method, structure, features and effects, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0018] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0019] The specific scheme of the enterprise industrial chain identification method and system based on a large language model provided by the present invention is described in detail below with reference to the accompanying drawings.

[0020] See also Figure 1 , which shows a method flow chart of a method for identifying an enterprise industrial chain based on a large language model provided by an embodiment of the present invention, the method comprising: Step S1, obtaining target enterprise information based on big data analysis technology.

[0021] Specifically, this application collects corporate public reports, social media data, and industry analysis reports, and uses big data analysis technology to extract key corporate information, such as the industry classification to which the company belongs, the company's business scope, the summary of the core patents held by the company, the name of the company's marketing activities or projects, the name of the software works owned by the company, the job titles owned by the company, and the company's product introduction, thereby forming a comprehensive and accurate information database.

[0022] Step S2, dividing the industry to be analyzed into multiple relatively independent links of the industrial chain.

[0023] Specifically, this application will first analyze the overall industrial chain structure of the industry in which the enterprise is located, and identify key operational links such as raw material supply, production and processing, and sales. Then, based on the logical relationship between the key operational links, multiple relatively independent industrial chain links will be divided.

[0024] Step S3, obtaining corresponding descriptive text for each divided industrial chain link.

[0025] Specifically, this application will write corresponding descriptive texts for each link in the industrial chain, which will cover the main functions, core technologies, market competition landscape, etc. of that link.

[0026] In one of the embodiments, the present application will also maintain the flexibility of the definition of industrial chain links by dynamically adjusting the division standards and descriptive texts of the industrial chain links so as to enable rapid adjustment and optimization according to changes in demand.

[0027] Step S4, based on text matching, annotate the target enterprise information with the industrial chain link label to which it belongs, and obtain the enterprise industrial chain link attribution annotation information.

[0028] Specifically, this application will compare the target enterprise information with the descriptive text of each industrial chain link to find the most matching link and mark it with the corresponding industrial chain link label.

[0029] Step S5, constructing a training data set and a verification data set based on the descriptive text of each industrial chain link, the attribution labeling information of the enterprise industrial chain link, and the task prompt words determined according to the task requirements and the model optimization strategy.

[0030] Specifically, this application will determine the feature vector representation of the industrial chain link based on the descriptive text of the industrial chain link and the attribution annotation information of the enterprise industrial chain link. These feature vector representations will serve as the basis for model training to help the model learn the distinction and correlation between the industrial chain links. On the other hand, appropriate task prompt words (such as "Which industrial chain link does the enterprise belong to?") will be determined based on task requirements (such as recognition accuracy, recall rate, etc.) and model optimization strategies (such as avoiding overfitting, improving training efficiency, etc.). Finally, the corresponding training data set and verification data set are constructed based on the descriptive text of each industrial chain link, the attribution annotation information of the enterprise industrial chain link, and the task prompt words.

[0031] Step S6: input the training data set and the verification data set into a generative large language model for model training to obtain a dedicated large language model for enterprise industrial chain identification.

[0032] Specifically, before model training, it is necessary to select or build a generative large language model (such as GPT, BERT, etc.) according to task requirements. The previously constructed training dataset and verification dataset will be input into the model for training and verification respectively.

[0033] During the training process, the model's parameter settings are continuously adjusted to improve the model's accuracy and efficiency in identifying the links in the industrial chain. At the same time, this application also uses a validation data set to evaluate the model's performance to ensure the stability and reliability of the model in practical applications. Finally, after the training is completed, a dedicated large language model is obtained that can accurately identify the links in the industrial chain to which the enterprise belongs.

[0034] As can be seen from the above, the enterprise industrial chain identification method based on the large language model disclosed in this application is based on the descriptive text of the industrial chain link, the enterprise industrial chain link attribution annotation information and the task prompt words to construct a training data set and a verification data set. Such a data set not only contains rich semantic information, but also optimizes according to the task requirements, which helps to train a high-performance model; the training data set and the verification data set are input into the generative large language model for training, and finally an exclusive large language model for enterprise industrial chain identification is obtained. The model can accurately understand the relationship between enterprise information and industrial chain links, and has a strong generation ability, and can provide personalized industrial chain identification services for enterprises; according to the actual operation of the industrial chain, the industrial chain is divided into multiple relatively independent links, which helps to deeply understand the composition and operation mechanism of the industrial chain, and provides a clear framework for subsequent enterprise information annotation and model training; in addition, descriptive texts are obtained for each industrial chain link, which contain key features, upstream and downstream relationships, functions and other information of the industrial chain link, and provide rich semantic and contextual information for model training.

[0035] In one embodiment, in step S1, the target enterprise information includes at least one of the industry classification to which the enterprise belongs, the business scope of the enterprise, the abstract of the core patents held by the enterprise, the name of the enterprise's marketing activities or projects, the name of the software works owned by the enterprise, the job titles owned by the enterprise, and the product introduction of the enterprise.

[0036] In one embodiment, in step S1, obtaining target enterprise information based on big data analysis technology includes: Step S11, based on big data analysis technology, obtain initial enterprise information.

[0037] Step S12, based on preset data cleaning rules, when it is determined that the initial enterprise information contains at least one abnormal data of missing data, noise data, and duplicate data, data cleaning processing is performed on the initial enterprise information to obtain enterprise cleansing information.

[0038] Specifically, for missing data, this application will use interpolation or a method based on similar data to推算 to complete the data. For noise data, this application will identify and filter out outliers and anomalies from the dataset by setting thresholds and performing cluster analysis. For duplicate data, this application will identify and delete the duplicate data based on a content similarity detection method to ensure the uniqueness of the data.

[0039] Step S13: Perform standardization processing based on the enterprise cleaning information to obtain enterprise standardization information.

[0040] Specifically, when performing standardization processing, this application will unify the field formats in the enterprise cleaning information. Then, according to the preset standardization rules, map enterprise information in different formats to unified standardized fields to achieve unified data formats.

[0041] Step S14: Remove text-irrelevant information and / or redundant description noise from the enterprise standardization information based on natural language processing technology to obtain target enterprise information.

[0042] Specifically, in the process of applying natural language processing technology, this application will first perform word segmentation on the text content in the enterprise standardization information to split it into individual words. Then, remove stop words in the segmented text, such as common but meaningless words like "de" (的) and "shi" (是), to reduce redundant information. Finally, use the keyword extraction algorithm in natural language processing technology to extract key information in the text, such as enterprise names, business scopes, etc., and integrate the text after word segmentation, filtering, extraction, and simplification processing to obtain the target enterprise information with text-irrelevant information and / or redundant description noise removed.

[0043] In the above embodiments, on the one hand, applying data cleaning processing can identify and correct missing data, noise data, and duplicate data in the initial enterprise information, significantly improving the accuracy and integrity of the data. On the other hand, the application of natural language processing technology can effectively remove text-irrelevant information and redundant description noise from the enterprise standardization information and extract key and valuable information. This helps to improve the accuracy and practicality of the information and provides a more accurate data basis for decision-making support.

[0044] In one of the embodiments, in step S5, the task prompt words include: the task background set according to the task requirements; the reference examples set according to the labeled enterprise industrial chain link attribution annotation information; the priority description set according to the importance of key business attributes and industry regulations; the white list and black list set according to the requirements of model recognition accuracy and avoiding misjudgment.

[0045] Specifically, for the task background, it is an exclusive identity set for the model, and it is clear that the target task is to answer the link of the industrial chain to which the enterprise belongs. For the reference example, it refers to the enterprise information text of the industrial chain that is manually annotated, and cites typical enterprise information and corresponding outputs for each link of the industrial chain. For the priority description, it corresponds to the scope limitation of each key business attribute, the priority of the content of each key business attribute itself (for example: the first business item in the business scope has a higher priority than the subsequent business items), and the priority between different key business attributes (for example: the analysis results of the patent abstract take precedence over the business scope, etc.). For the whitelist and blacklist, the whitelist lists the key attributes and keywords that should be considered to ensure that the model can pay attention to these important information during identification; the blacklist lists the key attributes and keywords that should be excluded to prevent the model from misjudgment.

[0046] In the above embodiment, the setting of task prompt words can clarify the task orientation, provide intuitive reference, optimize the processing sequence, and improve the model recognition accuracy, and can further determine the performance boundary and optimization direction of the model.

[0047] In one embodiment, in step S5, the training data set and the verification data set are both lists in JSON array format, each element in the list is a JSON object, and each JSON object includes a first key-value pair corresponding to the task prompt word, a second key-value pair corresponding to the expected input of the model, and a third key-value pair corresponding to the expected output of the model.

[0048] Specifically, the first key-value pair in each JSON object is used to store task prompts, which describe in detail key information such as task background, reference examples, and priority descriptions, providing clear guidance for model learning. The second key-value pair stores the input data expected by the model, which is preprocessed and labeled data to ensure that the model can learn effective features and information from it. The third key-value pair stores the output results expected by the model to indicate the correct industrial chain segment or mark it as "other industries." These results can be classification labels, regression values, etc., depending on the requirements of the task.

[0049] In one embodiment, in step S6, during the model training process, the method further includes: Step S61, in the initial parameter setting of the model, set the learning rate to 5e-5, the batch size to 100, the regularization coefficient to 0.01, and the number of model fine-tuning layers to the last two layers.

[0050] Specifically, the setting of the initial parameters of the model is mainly determined based on empirical values, among which the learning rate determines the amplitude of the model parameter update in each iteration. Batch processing affects the speed and memory usage of the model training. Therefore, it is necessary to set a moderate batch size according to the device configuration to ensure that the model can learn enough features while ensuring training efficiency. The regularization coefficient is used to control the complexity of the model and prevent the model from overfitting. The number of model fine-tuning layers determines which layers' parameters will be retrained based on the pre-trained model to adapt to the current task.

[0051] Step S62, based on the initial parameter settings and the training data set, perform the first round of iterative training, and during the training process, monitor the performance indicators of the model, and after the training is completed, when it is determined based on the verification data set that the recognition accuracy of the model for any type of enterprise industrial chain is lower than a preset threshold, adjust the learning rate by reducing it by one tenth, adjust the regularization coefficient by doubling it, and add one or two layers of model fine-tuning layers according to the specific structure and performance feedback of the model, and start the second round of iterative training.

[0052] Specifically, since a smaller learning rate can enable the model to learn data features more carefully during training, it helps to improve the stability and accuracy of the model. Therefore, in this embodiment, if the initial learning rate is set to 5e-5, then after adjusting it by one-tenth of the ratio, the adjusted learning rate is 4.5e-5. At this time, by slowing down the update speed of the parameters, the model can be prevented from overfitting.

[0053] Specifically, since increasing the regularization coefficient can strengthen the penalty for model complexity, it helps to improve the generalization ability of the model. Therefore, in this embodiment, if the initial regularization coefficient is set to 0.01, the regularization coefficient is adjusted according to the ratio of two times, and the adjusted regularization coefficient is 0.02, so as to further prevent the model from overfitting.

[0054] Specifically, since increasing the number of fine-tuning layers can make the model more flexible to meet the needs of specific tasks and improve the recognition ability of the model, in this embodiment, if the initial number of fine-tuning layers is set to the last two layers, it can be increased to the last three or four layers as needed to enhance the expression ability of the model.

[0055] Step S63, in each round of iteration, the process of training, evaluation, and parameter adjustment is repeated. Through multiple rounds of iterative training, the parameter settings of the model are continuously optimized so that the model focuses on the task of enterprise industrial chain identification.

[0056] Specifically, this application will monitor the loss function value, accuracy and other evaluation indicators during the training process in real time, as well as the evaluation results on the validation data set, so as to promptly discover and solve problems that arise during the training process. Furthermore, this application will also flexibly adjust the measurement of iterative training based on the performance and evaluation results of the model. For example, if the model performs well in a certain round of iterations, you can consider reducing the learning rate to further stabilize the performance; if the model has an overfitting problem, increase the regularization coefficient or increase the training data. Furthermore, this application will also terminate the training process in advance when it is determined that the performance on the validation data set has not improved for multiple consecutive rounds of iterations. This will help retain the parameter configuration of the model in the optimal state while saving computing resources.

[0057] In one embodiment, based on the model parameter setting, the method further includes: Step S64, collecting target cases where the model predicts errors, and recording the features and labels of the target cases.

[0058] Specifically, when collecting target cases where the model predicts errors, this application will set a standard for error judgment. For example, cases with a prediction accuracy lower than a certain set value (such as 80%) can be regarded as target cases with prediction errors. Subsequently, these incorrectly predicted cases are screened out from the test set through automated scripts (i.e., data screening and analysis tools). For each target case detected, this application will record its features (such as the various dimensions of the input data) and the corresponding labels in detail.

[0059] In one of the embodiments, in addition to recording the features and labels of the target case, the present application will also record some additional information, such as the confidence level at the time of prediction, other relevant information of the model output, etc., to facilitate subsequent analysis.

[0060] Step S65, based on the characteristics and labels of the target case, when it is determined that the data has category imbalance, noisy data or missing values, oversampling or undersampling technology is used to balance data categories, fill missing values, use data enhancement technology to increase data diversity, and optimize prompt word design to help the model gradually understand and master task requirements.

[0061] Specifically, for optimizing the prompt word design, this application will clarify the task requirements to ensure that the prompt words can clearly convey the task requirements and avoid ambiguity or vagueness. And design a variety of different prompt words for the same task to increase the diversity of data and improve the generalization ability of the model. Finally, according to the specific application scenario and context, adjust the expression of the prompt words to make it more in line with actual needs. For the application of data enhancement technology, this application will generate new samples through synonym replacement, sentence reorganization, random insertion or deletion of words, etc. to increase data diversity. For applications using oversampling or undersampling technology, for minority class samples, existing minority class samples will be repeated to increase their number, thereby balancing the category distribution; for majority class samples, this application will use clustering methods to select representative samples from majority class samples to reduce their number and achieve category balance. Filling missing values ​​belongs to the prior art, and this application does not limit this.

[0062] Step S66, based on the characteristics and labels of the target case, when it is determined that the model has overfitting or underfitting problems, the model parameter settings are optimized to ensure that the accuracy and robustness of the model are improved, and after completing the parameter fine-tuning, the model weights are saved.

[0063] Specifically, for the overfitting problem, this application will increase the weight of the regularization term, or reduce the complexity of the model (such as reducing the number of layers of the neural network, the number of neurons, etc.) to reduce the overfitting of the model to the training data. For the underfitting problem, this application will increase the complexity of the model (such as increasing the number of layers of the neural network, the number of neurons, etc.), or extend the training time so that the model has enough time to learn the characteristics of the data. At the same time, you can also adjust hyperparameters such as learning rate and batch size to find the best parameter combination to improve the accuracy and robustness of the model. After completing the parameter fine-tuning, the weight of the model will be further saved (this application does not limit its specific retention form) for subsequent applications.

[0064] In the above-mentioned embodiment, the present application can accurately locate the specific scenarios in which the model performs poorly by collecting target cases in which the model predicts errors, and provide a clear direction for subsequent analysis and model improvement. In addition, by recording the characteristics and labels of the target cases, it is helpful to explore and utilize the potential value in these data, such as discovering abnormal patterns, wrong annotations, etc. in the data, and provide a basis for subsequent parameter fine-tuning. In addition, by dealing with problems such as class imbalance, noisy data and missing values, the quality and diversity of training data can be improved, thereby enhancing the generalization ability and robustness of the model. By optimizing the prompt word design, the model can be helped to better understand and process the input data, thereby significantly improving the performance of the model.

[0065] In one embodiment, before fully automated reasoning is performed on millions of enterprise information based on the dedicated large language model, the method further includes: Step S7, based on the specified industry classification standard, preliminary industry classification and matching are performed on the million-level enterprise information to determine the industry category to which each enterprise belongs.

[0066] Specifically, this application conducts preliminary industry classification and matching of the millions of corporate information based on the national economic industry classification standards and in combination with the existing industry chain segmentation links.

[0067] Step S8, based on the industry classification and matching results, and taking into account at least one key factor of enterprise scale, business scope, and market position, select enterprise information that is representative and can reflect the characteristics of the industry as sample data for reasoning analysis.

[0068] Specifically, this application will conduct a detailed assessment of the information of each enterprise, including its scale (such as the number of employees, turnover, etc.), business scope (such as product lines, service areas, etc.) and market position (such as market share, brand influence, etc.). Based on the evaluation results, enterprises that meet the screening criteria are screened out. Afterwards, representative enterprises are selected from the screened enterprises. These enterprises have significant industry characteristics, such as industry leaders or enterprises with special competitive advantages. Finally, the enterprise information of the selected enterprises is organized into structured sample data, and these sample data will be used as input for the exclusive large language model reasoning analysis to achieve fully automated reasoning of millions of enterprise information, wherein the reasoning results (i.e., the link in the industrial chain to which the enterprise belongs) are directly output in a tabular format.

[0069] In the above embodiment, on the one hand, by performing preliminary industry classification and matching of millions of corporate information, the originally disorganized data can be made orderly and easy to manage. This helps to quickly locate and efficiently process the data in subsequent steps. On the other hand, selecting representative corporate information as sample data can enable the model to learn more commonalities and characteristics of industries during the training process. This helps to enhance the generalization ability of the model, so that it can also perform accurate reasoning and analysis when faced with new corporate information.

[0070] Please refer to Figure 2 The present application discloses a system for identifying an enterprise industrial chain based on a large language model, the system comprising an enterprise information acquisition module, an industrial chain segment division module, an industrial chain segment descriptive text generation module, an enterprise information annotation module, a training data construction module, and a model training module, wherein: The enterprise information acquisition module is used to acquire target enterprise information based on big data analysis technology.

[0071] The industrial chain segment division module is used to divide the industry to be analyzed into multiple relatively independent industrial chain segments.

[0072] The module for generating descriptive texts for industrial chain links is used to obtain corresponding descriptive texts for each divided industrial chain link.

[0073] The enterprise information labeling module is used to label the target enterprise information with the industrial chain link label based on text matching, and obtain the enterprise industrial chain link attribution labeling information.

[0074] The training data construction module is used to construct a training data set and a verification data set based on the descriptive text of each industrial chain link, the attribution labeling information of the enterprise industrial chain link, and the task prompt words determined according to the task requirements and the model optimization strategy.

[0075] The model training module is used to input the training data set and the verification data set into the generative large language model for model training, so as to obtain an exclusive large language model for identifying the enterprise industrial chain.

[0076] In one embodiment, the above modules are also used to implement the enterprise industrial chain identification method based on a large language model as described in any of the above method embodiments, which is not limited in this application.

[0077] As can be seen from the above, the enterprise industrial chain identification system based on the large language model disclosed in this application is based on the descriptive text of the industrial chain link, the enterprise industrial chain link attribution annotation information and the task prompt words to construct a training data set and a verification data set. Such a data set not only contains rich semantic information, but also optimizes according to the task requirements, which is helpful to train a high-performance model; the training data set and the verification data set are input into the generative large language model for training, and finally an exclusive large language model for enterprise industrial chain identification is obtained. The model can accurately understand the relationship between enterprise information and industrial chain links, and has a strong generation ability, and can provide personalized industrial chain identification services for enterprises; according to the actual operation of the industrial chain, the industrial chain is divided into multiple relatively independent links, which helps to deeply understand the composition and operation mechanism of the industrial chain, and provides a clear framework for subsequent enterprise information annotation and model training; in addition, descriptive texts are obtained for each industrial chain link, which contain key features, upstream and downstream relationships, functions and other information of the industrial chain link, and provide rich semantic and contextual information for model training.

[0078] The present application discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the enterprise industrial chain identification method based on a large language model is implemented.

[0079] As can be seen from the above, a computer-readable storage medium disclosed in the present application constructs a training data set and a verification data set based on the descriptive text of the industrial chain link, the attribution annotation information of the enterprise industrial chain link and the task prompt words. Such a data set not only contains rich semantic information, but also is optimized according to the task requirements, which is helpful to train a high-performance model; the training data set and the verification data set are input into the generative large language model for training, and finally an exclusive large language model for enterprise industrial chain identification is obtained. The model can accurately understand the relationship between enterprise information and industrial chain links, and has a strong generation ability, and can provide personalized industrial chain identification services for enterprises; according to the actual operation of the industrial chain, the industrial chain is divided into multiple relatively independent links, which helps to deeply understand the composition and operation mechanism of the industrial chain, and provides a clear framework for subsequent enterprise information annotation and model training; in addition, descriptive texts are obtained for each industrial chain link, which contain key features, upstream and downstream relationships, functions and other information of the industrial chain link, and provide rich semantic and contextual information for model training.

[0080] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0081] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A method for identifying an enterprise industrial chain based on a large language model, characterized in that: The method comprises: S1. Obtain target enterprise information based on big data analysis technology; S2. For the industry to be analyzed, divide the industry chain into multiple relatively independent links; S3. Obtain corresponding descriptive text for each segment of the industrial chain; S4. Based on text matching, label the target enterprise information with the industrial chain link label to which it belongs, and obtain the enterprise industrial chain link attribution labeling information; S5. Construct a training data set and a validation data set based on the descriptive text of each industrial chain link, the attribution annotation information of the enterprise industrial chain link, and the task prompt words determined according to the task requirements and model optimization strategy; S6. Input the training data set and the verification data set into a generative large language model for model training to obtain a dedicated large language model for identifying the enterprise industrial chain.

2. The method according to claim 1, characterized in that In step S1, the target enterprise information includes at least one of the industry classification to which the enterprise belongs, the business scope of the enterprise, the abstract of the core patents held by the enterprise, the name of the enterprise's marketing activities or projects, the name of the software works owned by the enterprise, the job titles owned by the enterprise, and the product introduction of the enterprise.

3. The method according to claim 1, characterized in that In step S1, the target enterprise information is obtained based on big data analysis technology, including: S11. Obtain initial enterprise information based on big data analysis technology; S12, based on a preset data cleaning rule, when it is determined that the initial enterprise information contains at least one abnormal data of missing data, noise data, and duplicate data, performing data cleaning processing on the initial enterprise information to obtain enterprise cleansing information; S13, performing standardization processing based on the enterprise cleaning information to obtain enterprise standardized information; S14. Based on natural language processing technology, remove text-irrelevant information and / or redundant description noise in the enterprise standardized information to obtain target enterprise information.

4. The method according to claim 1, characterized in that In step S5, the task prompt words include: The task background is set according to the task requirements; Reference examples set based on the annotated information on the industry chain links of the annotated enterprises; A description of the priorities set based on the importance of key business attributes and industry regulations; Whitelists and blacklists are set based on the accuracy of model recognition and the need to avoid misjudgment.

5. The method according to claim 1, characterized in that In step S5, the training data set and the verification data set are both lists in JSON array format, each element in the list is a JSON object, and each JSON object contains a first key-value pair corresponding to the task prompt word, a second key-value pair corresponding to the expected input of the model, and a third key-value pair corresponding to the expected output of the model.

6. The method according to claim 1, characterized in that In step S6, during the model training process, the method further includes: S61. In the initial parameter setting of the model, set the learning rate to 5e-5, the batch size to 100, the regularization coefficient to 0.01, and the number of model fine-tuning layers to the last two layers; S62, based on the initial parameter settings and the training data set, perform the first round of iterative training, and during the training process, monitor the performance indicators of the model, and after the training is completed, when it is determined based on the verification data set that the recognition accuracy of the model for any type of enterprise industrial chain is lower than a preset threshold, adjust the learning rate by reducing it by one tenth, adjust the regularization coefficient by increasing it by two times, and increase one or two layers of model fine-tuning layers according to the specific structure and performance feedback of the model, and start the second round of iterative training; S63. In each round of iteration, the process of training, evaluation, and parameter adjustment is repeated. Through multiple rounds of iterative training, the parameter settings of the model are continuously optimized so that the model focuses on the task of identifying the enterprise industrial chain.

7. The method according to claim 6, characterized in that Based on the model parameter setting, the method further includes: S64, collecting target cases predicted incorrectly by the model, and recording features and labels of the target cases; S65. Based on the characteristics and labels of the target case, when it is determined that the data has category imbalance, noisy data or missing values, oversampling or undersampling techniques are used to balance data categories, fill missing values, use data enhancement techniques to increase data diversity, and optimize prompt word design to help the model gradually understand and master task requirements; S66. When it is determined that the model has an overfitting or underfitting problem based on the characteristics and labels of the target case, the model parameter settings are optimized to ensure that the accuracy and robustness of the model are improved, and after completing the parameter fine-tuning, the model weights are saved.

8. The method according to claim 1, characterized in that Before fully automated reasoning of millions of enterprise information based on the exclusive large language model, the method further includes: S7. Based on the specified industry classification standards, preliminary industry classification and matching are performed on the information of the millions of enterprises to determine the industry category to which each enterprise belongs; S8. Based on the industry classification and matching results, taking into account at least one key factor including enterprise size, business scope, and market position, select enterprise information that is representative and can reflect the characteristics of the industry as sample data for inference analysis.

9. An enterprise industry chain identification system based on a large language model, characterized in that: The system includes an enterprise information acquisition module, an industrial chain segment division module, an industrial chain segment descriptive text generation module, an enterprise information annotation module, a training data construction module, and a model training module, wherein: The enterprise information acquisition module is used to acquire target enterprise information based on big data analysis technology; The industrial chain segment division module is used to divide the industry to be analyzed into multiple relatively independent industrial chain segments; The module for generating descriptive texts of industrial chain links is used to obtain corresponding descriptive texts for each divided industrial chain link; The enterprise information labeling module is used to label the target enterprise information with the industry chain link label based on text matching to obtain the enterprise industry chain link attribution labeling information; The training data construction module is used to construct a training data set and a verification data set based on the descriptive text of each industrial chain link, the attribution labeling information of the enterprise industrial chain link, and the task prompt words determined according to the task requirements and the model optimization strategy; The model training module is used to input the training data set and the verification data set into the generative large language model for model training, so as to obtain an exclusive large language model for identifying the enterprise industrial chain.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the enterprise industrial chain identification method based on a large language model is implemented.

Images