A method and related device for obtaining affiliated companies based on enterprise and institution information

Abstract

The present application relates to the technical field of information processing technology, and discloses a method and related device for obtaining related companies based on enterprise and institution information. The method includes: obtaining raw data, cleaning and confidence calculation of enterprise and institution information, filtering low-confidence data, and obtaining high-confidence data; performing semantic analysis and confidence verification on high-confidence data to filter out trusted related data; inputting the trusted related data into a stable node processing language model to obtain the final related result, wherein the stable node processing model and the sensitive node processing model are mutually trained, the stable node processing model is used to train common nodes and high-confidence samples, and the sensitive node processing model is used to train new nodes or low-confidence samples. The present application has the advantage of balancing the accuracy of high-confidence samples and the adaptability of dynamic nodes.

Description

Technical Field

[0001] The present application relates to the technical field of information processing technology, and in particular to a method for obtaining affiliated companies based on information of enterprises and institutions. Background Art

[0002] In modern society, obtaining information about related businesses and institutions is crucial. Analyzing the relationships between businesses allows us to provide users with more comprehensive and accurate services, which plays an indispensable role in business collaboration, risk assessment, and industry chain optimization.

[0003] Existing techniques typically use a single model to process input material to identify the affiliated companies of a target enterprise or institution. This approach establishes matching relationships between enterprises and affiliated entities through textual analysis and semantic understanding of enterprise information. However, while a single model performs well for high-confidence samples (such as readily apparent affiliated companies), its adaptability to dynamic nodes (such as newly emerging enterprise information or unseen node types) is limited. In particular, when a single model needs to process both common and dynamic nodes, the complexity of the task often leads to overfitting on high-confidence samples, reducing its generalization ability to new nodes and low-confidence samples. Summary of the Invention

[0004] In order to balance the accuracy of high-confidence samples and the adaptability of dynamic nodes, the present application provides a method and related devices for obtaining affiliated companies based on enterprise and institution information.

[0005] In the first aspect, this application provides a method for obtaining affiliated companies based on information of enterprises and institutions, using the following technical solutions:

[0006] A method for obtaining affiliated companies based on enterprise and institution information, comprising the following steps:

[0007] S1. Obtain raw data, clean and calculate confidence levels for enterprise and institution information, filter low-confidence data, and obtain high-confidence data;

[0008] S2. Perform semantic analysis and confidence verification on high-confidence data to screen out credible related data;

[0009] S3. Input the trusted association data into the stable node processing language model to obtain the final association result, wherein the stable node processing model and the sensitive node processing model are trained with each other. The stable node processing model is used to train common nodes and high-confidence samples, and the sensitive node processing model is used to train new nodes or low-confidence samples. The stable node processing model and the sensitive node processing model share the same input data, and the sensitive node processing model introduces an attention mechanism to strengthen the feature learning of new nodes; the stable node processing model and the sensitive node processing model share a comprehensive loss function, and the results of the stable node processing model and the sensitive node processing model are interactively fed back, and the parameters are adjusted according to the joint loss function.

[0010] Optionally, the structural design step of the stable node processing model includes:

[0011] The embedding layer is used to convert the input text into a context-sensitive high-dimensional vector representation using BERT or RoBERTa. The model input is the concatenated text to be processed, which includes enterprise information and node descriptions. The enterprise information includes the enterprise name, enterprise profile, and a list of main business keywords. The node description includes the node name and node details.

[0012] The feature extraction layer, composed of multiple layers of Transformers, is used to capture the semantic relationship between enterprise information and node descriptions;

[0013] The output layer is used to output the classification head and association probability. The classification head uses BERT's output [CLS] tag as the overall semantic representation, and the association probability passes through a fully connected layer and a softmax function.

[0014] Optionally, the structural design step of the sensitive node processing model includes:

[0015] An embedding layer, which uses BERT or RoBERTa to convert input text into a context-sensitive high-dimensional vector representation and uses a dynamic rule embedding module to embed the dynamic features of new nodes into the model. The model input is samples whose probability output by the stable node processing model is below a certain threshold, including enterprise information and node descriptions. The enterprise information includes the enterprise name, enterprise profile, main business keyword list, and historical confidence information. The node description includes the new node name, new node details, and dynamic keywords. The historical confidence information is the predicted distribution of low-confidence samples obtained from the output of the second language model.

[0016] The feature extraction layer, composed of multiple layers of Transformers, is used to strengthen the semantic association between enterprise information and new node descriptions by introducing a multi-head attention mechanism. The association between dynamic keywords and enterprise information is processed by a dedicated attention head.

[0017] The output layer includes a classification task head and a dynamic rule generation head. The classification task head is used to determine whether an enterprise or institution belongs to a new node, and the dynamic rule generation head is used to generate dynamic features that can be used for learning by the stable node processing model.

[0018] Optionally, the step of strengthening the semantic association between enterprise and institution information and new node description by introducing a multi-head attention mechanism includes the following steps:

[0019] Split the input embedding into multiple subspaces, i.e. multiple heads;

[0020] Each head independently calculates attention weights to capture different semantic features;

[0021] The outputs of all heads are concatenated and linearly transformed.

[0022] Optionally, the joint training step of the sensitive node processing model and the stable node processing model includes:

[0023] S301. Separately train the sensitive node processing model and the stable node processing model;

[0024] S302. Input the current sample into the stable node processing model for prediction and calculate the classification loss;

[0025] S303. The low confidence samples are transferred to the sensitive node processing model to generate dynamic rule characteristics;

[0026] S304. Calculate the difference between the predicted probability and the true label based on the joint loss function;

[0027] S305. Based on the joint loss function, calculate the gradient through back propagation;

[0028] S306. The parameters of the sensitive node processing model and the stable node processing model are updated simultaneously through the optimizer, and the parameter sharing part is synchronously adjusted between the two.

[0029] Optionally, the joint loss function is , is a hyperparameter used to control the contribution of each part loss to training;

[0030] is the classification loss, which is used to optimize the classification task of whether enterprises and institutions belong to nodes. ;in, is the true label, indicating whether the enterprise or institution belongs to a certain target node; is the probability output predicted by the model; N is the number of samples;

[0031] To stabilize the consistency loss of the node processing model and the sensitive node output model, ,in, To stabilize the output distribution of the node processing model, Processing model output distribution for sensitive nodes; is the Kullback–Leibler divergence, which measures the difference between two probability distributions;

[0032] Supplement the loss for dynamic rules, ,in, To stabilize the node feature embedding of the node processing model, Handle dynamic feature embedding of models for sensitive nodes.

[0033] Optionally, the step S304 includes the following sub-steps:

[0034] S3041. Combine the output of the sensitive node processing model and compare it with the results of the stable node processing model to calculate the consistency loss;

[0035] S3042. Using dynamic rules to supplement features to optimize feature generation of stable node processing models;

[0036] S3043. Dynamically adjust the weights of consistency loss and dynamic rule supplement loss according to the model training progress.

[0037] Optionally, S1 includes the following sub-steps:

[0038] S11. Obtaining raw data and performing preprocessing;

[0039] S12. Use the first pre-trained model to perform semantic analysis, map the enterprise information and industrial chain into the same semantic space, and determine the initial association result based on the semantic similarity;

[0040] S13. Perform confidence verification on the initial association results to screen out credible association data.

[0041] Secondly, this application provides a system for obtaining affiliated companies based on information of enterprises and institutions, which adopts the following technical solutions:

[0042] A system for acquiring related companies based on enterprise and institution information includes a processor running a program of any one of the above methods for acquiring related companies based on enterprise and institution information.

[0043] In a third aspect, the present application provides a storage medium that adopts the following technical solution:

[0044] A storage medium stores a program for any one of the above methods for obtaining related companies based on enterprise and institution information.

[0045] In summary, this application includes at least one of the following beneficial technical effects:

[0046] First, by combining the stable node processing model with the sensitive node processing model, the present invention conducts targeted training and processing on high-confidence samples and new nodes or low-confidence samples respectively. It can not only ensure high classification accuracy on known nodes, but also quickly adapt to newly emerging node types, thereby effectively improving the generalization ability and adaptability of the overall system.

[0047] Second, the present invention introduces a multi-head attention mechanism and a dynamic rule supplementation mechanism, enabling the model to fully explore hidden semantic connections and continuously learn key features of new nodes when processing the associations between enterprise and institution information and industrial chain nodes. By comprehensively considering classification loss, consistency loss, and dynamic rule supplementation loss in a common loss function, the system can continuously absorb new feature information while maintaining overall stability, further enhancing the accuracy and scalability of association analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flowchart of a method for obtaining affiliated companies based on enterprise and institution information in a certain embodiment of the present application. DETAILED DESCRIPTION

[0049] Embodiments of the present application are described in detail below, examples of which are illustrated in the accompanying drawings.

[0050] Throughout this specification, reference to the terms "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with the embodiment or example is included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0051] The embodiment of the present application discloses a method for obtaining related companies based on enterprise and institution information, referring to Figure 1 , including the following steps S1-S3.

[0052] S1. Obtain raw data, clean and calculate confidence levels for enterprise and institution information, filter low-confidence data, and obtain high-confidence data.

[0053] Optionally, the S1 includes the following sub-steps S11-S13.

[0054] S11. Obtaining raw data and performing preprocessing;

[0055] S12. Use the first pre-trained model to perform semantic analysis, map the enterprise information and industrial chain into the same semantic space, and determine the initial association result based on the semantic similarity;

[0056] S13. Perform confidence verification on the initial association results to screen out credible association data.

[0057] This step first requires obtaining raw company information and industry chain data, performing preliminary preprocessing and calculating confidence. Company information typically includes the company name, main products, a company profile, and other data reflecting the company's core business and scale. Through cleaning and statistical analysis, obviously incomplete or logically incorrect information can be removed, and entries with low confidence levels can be filtered out. Confidence, as used here, can be understood as a quantitative assessment of data reliability, measured using metrics such as historical records, source credibility, and internal consistency. For example, for a newly established company, its core business information may be incomplete and its source lacks authoritativeness, resulting in a low confidence level. After completing this filtering step, a relatively reliable and high-quality dataset is often left, laying the foundation for subsequent semantic analysis. This process ensures that the model is less susceptible to noise when encountering new data, resulting in more accurate results in subsequent correlation analysis.

[0058] After preprocessing, semantic analysis of enterprise information and industry chain descriptions is performed using a stable node processing model. "Semantic analysis" involves processing text using large-scale pre-trained language models (such as BERT and GPT), mapping it into high-dimensional vector representations to capture implicit semantic relationships. Pre-training on a large-scale general corpus allows the model to learn common semantic distribution patterns of words, phrases, and sentences. Subsequently, fine-tuning on a smaller, industry-specific dataset allows the model to better adapt to the characteristics of specific industry text. Converting both enterprise information and industry chain descriptions into vectors and mapping them to the same semantic space allows the model to measure similarity between them using the same semantic metric. This "semantic space" can be understood as a multi-dimensional vector space: Textual content such as company names, company profiles, and product keywords are individually converted into vectors in the model's embedding layer. Industry chain node descriptions are also converted into vectors using the same embedding method. The relative distances or directions between these vectors reflect the semantic similarities and differences between the two.

[0059] In this semantic space, by calculating the semantic similarity between vectors (common methods include cosine similarity or Euclidean distance), the correlation between a company's information and a specific industry chain node can be determined. Specifically, if a company's name or core business keywords show a high degree of similarity with a node description, the model will preliminarily determine that the two are closely related. This calculation process leverages the pre-trained model's ability to capture contextual semantics: the model goes beyond simply matching keywords. Instead, based on linguistic knowledge learned from a large corpus, it can understand the semantic similarities between company names and different representations of industry chain nodes. For example, if a company's core business is "smart home," and the industry chain node description uses the term "smart home devices," the two are not exactly the same in literal sense. However, because the model captures their shared relevance to "home automation," the distance between them in the semantic space will be closer, leading to a preliminary association.

[0060] After completing the above-mentioned vector mapping and similarity calculation, the system will determine which companies and which industrial chain nodes have significant associations based on the set threshold, and regard these matching results as initial association results. The purpose of this preliminary screening is to ensure that the data volume is controllable and to include the most likely related company-industry chain combinations in the subsequent processing scope. After this step, users can use this part of the preliminary association data to carry out further feature extraction or detailed analysis, realize the visualization and precise mining of the connection between enterprises and industrial chains, and achieve the effect of improving efficiency and accuracy in actual business cooperation and risk assessment. Combined with the good adaptability of the stable node processing model to common, high-confidence samples, this step can capture routine business connections to the greatest extent in the initial stage, preparing for subsequent dynamic node adaptation and more complex sensitive node processing.

[0061] S2. Perform semantic analysis and confidence verification on high-confidence data to filter out credible related data.

[0062] After completing the initial semantic analysis and association matching, the resulting association results need to be confidence verified. Confidence verification is a process that assesses the reliability of the results. Based on factors such as semantic similarity, multi-source data support, and historical records, the generated association results are assigned a score between 0 and 1, known as a confidence score. When this score exceeds a pre-set threshold, the result is considered credible association data. Otherwise, it is marked as low-confidence or uncertain data, and may require manual review or dynamic learning by the sensitive node processing model for further processing.

[0063] In scenarios where information about enterprises and institutions is associated with industrial chain nodes, the implementation of confidence verification usually includes the quantitative calculation of semantic similarity and the evaluation of the authority and consistency of the data source. For example, if the output result of the semantic similarity based on the stable node processing model is higher than a threshold value widely used in practice (such as 0.75 or 0.9), it can be determined that there is a significant association between the name of the enterprise and institution and the description of the industrial chain node at the semantic level; at the same time, if the association result is cross-validated by multiple authoritative data sources or public data from relevant departments, then its confidence will be further improved. This comprehensive calculation process helps the system give a relatively objective confidence score for each associated record, and then, combined with user-defined industry needs, selects an appropriate threshold to distinguish between credible data and suspicious data.

[0064] In order to make the confidence score more accurate, a variety of scoring factors are introduced in practice. The most direct factor is the semantic similarity score, which is usually output by the model and normalized. Secondly, the authority of the data source itself cannot be ignored: if the identified association comes only from anonymous online text, the score will be correspondingly lowered; if it comes from the official website of an enterprise or institution or a professional database, the score will be higher. In addition, the model sometimes also counts historical matching records. If an enterprise or institution has often been correctly associated with a certain type of node in previous analyses, then historical performance will also be a major factor in the scoring. For different fields or different business scenarios, the specific threshold settings may also be different: in high-risk or high-value fields (such as scientific research institutions, intellectual property, etc.), higher thresholds are usually used to ensure the accuracy of the results; in general manufacturing or trade fields, the threshold can be appropriately lowered to avoid filtering out a large number of potential association opportunities.

[0065] This confidence verification mechanism not only allows the system to screen reliable association data for subsequent analysis or display, but also effectively identifies association results with high uncertainty, preparing for further dynamic processing. If certain association scores fall below the threshold but are not completely invalid, they are considered potential low-confidence samples, often requiring further learning and correction by sensitive node processing models.

[0066] S3. Input the trusted association data into the stable node processing language model to obtain the final association result, wherein the stable node processing model and the sensitive node processing model are trained with each other. The stable node processing model is used to train common nodes and high-confidence samples, and the sensitive node processing models are trained with each other to train new nodes or low-confidence samples. The stable node processing model and the sensitive node processing model share the same input data, and the sensitive node processing model introduces an attention mechanism to strengthen the feature learning of new nodes; the stable node processing model and the sensitive node processing model share a comprehensive loss function, and the results of the stable node processing model and the sensitive node processing model are interactively fed back, and the parameters are adjusted according to the joint loss function.

[0067] Specifically, the structural design steps of the stable node processing model include an embedding layer, a feature extraction layer and an output layer.

[0068] The embedding layer is used to convert the input text into a context-sensitive high-dimensional vector representation using BERT (Bidirectional Encoder Representations from Transformers) or RoBERTa (Robustly optimized BERT approach). The model input is the concatenated text to be processed, which includes enterprise information and node descriptions. The enterprise information includes the enterprise name, enterprise profile, and a list of main business keywords. The node description includes the node name and node details.

[0069] The feature extraction layer is composed of multiple layers of Transformer stacking, which is used to capture the semantic association between enterprise information and node descriptions.

[0070] The output layer is used to output the classification head and association probability. The classification head uses BERT's output [CLS] tag as the overall semantic representation, and the association probability passes through a fully connected layer and a softmax function.

[0071] In this solution, the stable node processing model assumes the core function of efficiently processing common nodes and high-confidence samples. Its overall structure is based on a pre-trained Transformer framework (such as BERT or RoBERTa), primarily comprising an embedding layer, a feature extraction layer, and an output layer. By concatenating enterprise and institution information with industry chain node descriptions into a single input sequence, it achieves a joint understanding of the semantics of both parts of the text. To better address various business scenarios, the model is fine-tuned based on industry characteristics after pre-training, making it more accurate and stable when processing common nodes or high-confidence text.

[0072] In the model's embedding layer, all input text must first be converted into a context-sensitive, high-dimensional vector representation. This is accomplished by concatenating the enterprise's information (such as its name, profile, and list of key business keywords) with the industry chain node description (such as the node name and details), and inserting a special marker at the beginning to generate the overall semantic vector. For example, if an enterprise is called a "green energy agency" and its main businesses include "photovoltaic equipment" and "wind turbines," and the industry chain node description includes information about "new energy equipment manufacturing," the model will serialize all of this text and input it into the pre-trained embedding layer. Because BERT or RoBERTa can learn rich contextual semantics from massive corpora, the model can not only identify the specific industries these keywords refer to, but also capture their inherent semantic connections with concepts such as "new energy" and "manufacturing," thereby constructing a more accurate vector representation in a high-dimensional space.

[0073] After embedding, the model enters the feature extraction layer, comprised of multiple layers of Transformers. This layer's multi-head attention mechanism calculates the dependencies between each position in the sequence and other positions, extracting deeper semantic interactions between the text. For example, if an enterprise's main business and the description of a specific industry chain node share similar or complementary keywords, the multi-head attention mechanism accurately captures this correspondence and highlights it in the subsequent feature representation. This allows the model to distinguish subtle differences between concepts like "photovoltaic" and "wind power," understanding that they belong to higher-level semantic categories like "new energy equipment." Because stable node processing models primarily target common nodes and high-confidence data, they often focus on providing highly accurate classification results within known business domains while minimizing misclassification or neglect of such data. The deep learning capabilities of multi-layer Transformers precisely meet this requirement, enabling stable association judgments across massive amounts of enterprise and industry chain node text.

[0074] After feature extraction is complete, the model will concentrate the contextual representation of the entire sequence into the output layer, using a fully connected layer and a softmax function to output the association probability. At this point, the hidden state vector labeled [CLS] will be passed into the classification head as a semantic aggregation vector representing the entire input sequence (including enterprise information and industry chain node descriptions), and then calculate whether the current input combination has a high degree of association. For example, if the comprehensive semantic similarity between "photovoltaic equipment manufacturer" and "new energy equipment manufacturing" exceeds the threshold, the model will determine that they have a high probability of belonging to the same industry chain node. In addition, in many application scenarios, a more detailed association degree score can be obtained by calculating the cosine similarity between the enterprise text and the node text, providing a more multi-dimensional reference indicator for subsequent confidence verification and business decision-making.

[0075] Because the model is pre-trained on a large-scale corpus and fine-tuned with industry data, it can provide relatively stable and accurate association results in most common situations. As a result, when the system processes high-confidence samples, the stable node processing model is often able to make accurate judgments immediately and directly feedback the results to users or enter downstream processes. For dynamic node data with lower confidence or containing new concepts, the sensitive node processing model can further perform reinforcement learning or refined processing, achieving a global balance between high accuracy and good adaptation to new nodes.

[0076] Specifically, the structural design steps of the sensitive node processing model include an embedding layer, a feature extraction layer and an output layer.

[0077] The embedding layer is used to convert the input text into a context-sensitive high-dimensional vector representation using BERT or RoBERTa, and utilizes the dynamic rule embedding module to embed the dynamic features of new nodes into the model. The model input is samples whose probability output by the stable node processing model is lower than a certain threshold, including enterprise information and node descriptions. The enterprise information includes the enterprise name, enterprise profile, main business keyword list, and historical confidence information. The node description includes the new node name, new node details, and dynamic keywords. The historical confidence information is the predicted distribution of low-confidence samples obtained from the output of the second language model.

[0078] The feature extraction layer is composed of multiple layers of Transformers. It uses a multi-head attention mechanism to strengthen the semantic association between enterprise information and new node descriptions. The association between dynamic keywords and enterprise information is processed by a dedicated attention head.

[0079] The output layer contains a classification task head and a dynamic rule generation head. The classification task head is used to determine whether an enterprise or institution belongs to a new node, and the dynamic rule generation head is used to generate dynamic features that can be used for learning by the stable node processing model.

[0080] The sensitive node processing model primarily targets new, unseen industry chain nodes and low-confidence samples. It uses dynamic learning and attention mechanisms to supplement or enhance the analysis of this uncertain data. Architecturally, it is also based on a pre-trained Transformer model, but with more specialized processing methods tailored to the dynamic characteristics of new nodes in input design and feature extraction, enabling rapid adaptation to enterprise and institution information and unknown or uncommon node descriptions.

[0081] During the input phase, the model receives low-confidence samples from the previous phase (i.e., output from the stable node processing model) as its primary data source. These samples often have low confidence due to significant semantic differences, relatively new business areas, or a lack of high-quality historical records. Enterprise and institution information includes a name, a brief introduction, a list of key business keywords, and optional historical confidence information. Nodes may contain new node names, descriptive text about the node, and dynamic keywords (typically extracted from recent industry reports, news, or the latest research). These nodes are called "sensitive nodes" because they may contain new technologies, new industry directions, or features not yet incorporated into the existing rule set. Without separate processing, they can easily be misjudged or ignored by traditional models.

[0082] In the embedding layer, the model uses BERT, RoBERTa, or similar pre-trained language models to convert text sequences into context-sensitive high-dimensional vector representations. Unlike conventional embedding methods, it introduces an additional dynamic rule embedding module to learn new node features and embed them into the model. Simply put, in addition to conventional text vectors, the model also generates specific "new node vectors" based on dynamic keywords, industry hotspots, and possible historical confidence features, and integrates them with the text vectors of enterprises and institutions. This approach can help the model capture emerging or not-yet-popular node concepts more quickly, so that subsequent semantic interaction processes are no longer limited to known areas.

[0083] In the feature extraction layer, multi-layer Transformer stacking provides the ability to learn deep semantic associations between different text fragments. Since the sensitive node processing model focuses on the relationship between the new node and the enterprise, the attention mechanism design will perform separate attention head processing on dynamic keywords and enterprise profiles and main business keywords to ensure that potential new technologies or business points can be highlighted. For example, if the dynamic keyword of the new node is "hydrogen fuel cell", and the enterprise profile mentions words such as "hydrogen production process" and "gas purification equipment", the model can capture signs of potential intersection between the two in the industrial chain through a specialized attention mechanism.

[0084] Finally, at the output layer, the model typically consists of two parts: a classification task head, which determines whether the current enterprise or institution belongs to a given new node and outputs an association probability; and a dynamic rule generation head, which generates dynamic characteristics or rules related to the new node based on new industry information captured by the model during training. These dynamic rules can be used to improve the description of similar new nodes and also feed back into the stable node processing model, allowing the overall system to gradually enrich the known node rule base in subsequent iterations, thereby improving the processing efficiency and accuracy of similar new nodes. By collaborating with the stable node processing model, the sensitive node processing model can effectively share the analysis work of difficult or undefined node data.

[0085] Optionally, the step of strengthening the semantic association between enterprise and institution information and new node description by introducing a multi-head attention mechanism includes the following steps ac.

[0086] a. Split the input embedding into multiple subspaces, i.e. multiple heads.

[0087] b. Each head independently calculates the attention weight to capture different semantic features; the attention weight is: Q is the query vector, representing the feature representation of the enterprise information; K is the key vector, representing the feature representation of the new node description; V is the value vector, representing the semantic information of the new node description; d is the dimension of the vector, and the scaling factor is used to balance the calculation. T represents the transpose of the matrix.

[0088] c. Concatenate and linearly transform the outputs of all heads, i.e.

[0089] .in, ; is an independent parameter matrix for each head, is the output transformation matrix.

[0090] When semantically interacting with enterprise and institution information and new node descriptions, complex and ever-changing contextual dependencies are often encountered. As new nodes continue to emerge, there may be implicit or cross-sentence connections between them and text such as the enterprise introduction or main business. Relying solely on a single attention head often cannot fully capture all the details. By having multiple attention heads work in parallel, the model can focus on multi-dimensional information such as keywords, syntactic structures, and upstream and downstream semantics from different angles, and give appropriate attention to multiple aspects of the same input, making the overall semantic mapping more refined. In this process, the model can adaptively adjust the weights of each head, so that when faced with different types of input, it can flexibly highlight key points and weaken noise.

[0091] In specific implementations, both the stable node processing model and the sensitive node processing model will split the input text into several vector sequences. Enterprise information is treated as the query vector Q, while the new node description and its dynamic keywords serve as the key vector K and value vector V for the model to calculate attention. The core of multi-head attention is to perform a dot product operation on Q and K and scale them, then obtain the attention distribution through the softmax function, and then multiply it with V to generate the context vector. Because the parameters of each head are independent of each other, different heads will gradually learn to capture different semantic features during training. For example, the first head may focus on direct keyword matching, the second head may study syntactic dependencies or temporal information in the context, and the third head may pay more attention to long-distance cross-references. In this way, the same enterprise information and new node description will generate several sub-features in different heads, which will eventually be concatenated and linearly transformed to output a more comprehensive contextual representation.

[0092] For example, a company's profile mentions "focusing on energy storage system R&D," while a new node's description mentions "new energy storage equipment." When the multi-head attention algorithm begins, the model first maps "energy storage system R&D" into multiple vector segments in Q, while "new energy storage equipment" and other text segments are entered separately into K and V. Each head independently calculates the degree of match between them, using cosine similarity or other metrics to determine the semantic relevance of the two segments. It then allocates attention to the most relevant segments. For example, if the attention mechanism recognizes that "energy storage system" and "energy storage equipment" share similar industry terminology, it assigns a higher weight to highlight this connection. After all heads have determined their attention distributions, the model merges their outputs into a unified vector representation, which is then passed to the subsequent classification layer or dynamic generation module to determine the probability of association between the company and the new node and to discover new industry characteristics.

[0093] On the one hand, multi-head attention can effectively capture the correspondence between keywords, helping the system emphasize semantic points that can reflect the potential for business overlap or technical collaboration; on the other hand, it can track contextual information across multiple sentences, allowing the model to recognize that concepts scattered across multiple parts of the company profile and new node descriptions are closely related.

[0094] Optionally, the joint training step of the sensitive node processing model and the stable node processing model includes:

[0095] S301. Separately train the sensitive node processing model and the stable node processing model;

[0096] S302. Input the current sample into the stable node processing model for prediction and calculate the classification loss;

[0097] S303. The low confidence samples are transferred to the sensitive node processing model to generate dynamic rule characteristics;

[0098] S304. Calculate the difference between the predicted probability and the true label based on the joint loss function;

[0099] S305. Based on the joint loss function, calculate the gradient through back propagation;

[0100] S306. The parameters of the sensitive node processing model and the stable node processing model are updated simultaneously through the optimizer, and the parameter sharing part is synchronously adjusted between the two.

[0101] This joint training process begins with separate training of the stable and sensitive node processing models, each achieving relatively mature initial weights for its respective tasks. The stable node processing model primarily learns common nodes and high-confidence samples, focusing on ensuring classification accuracy and model stability for these data. The sensitive node processing model, on the other hand, dynamically adapts to new nodes or low-confidence samples, aiming to improve generalization capabilities in unknown scenarios through mechanisms such as active learning and rule generation.

[0102] After independent training is complete, the training process enters the joint training phase (S302-S306). In this phase, the stable node processing model first predicts the current sample and calculates the classification loss. If any low-confidence results are found, these samples are passed to the sensitive node processing model to generate new dynamic rule features. The purpose of this step is to allow the sensitive node processing model to focus on samples that the stable node processing model cannot accurately identify or that have insufficient confidence. By mining these data for more subtle semantic features or new node information, it can generate dynamic features that are instructive for the new domain. Subsequently, the system uses a joint loss function to measure the deviation between the current predicted probability and the true label. Backpropagation is then performed using a combination of a consistency loss and a dynamic rule supplementation loss. In this joint loss function, the classification loss primarily ensures overall prediction performance, the consistency loss constrains the output distributions of the stable and sensitive node processing models from diverging completely, and the dynamic rule supplementation loss helps the stable node processing model gradually incorporate new node features captured by the sensitive node processing model, thereby achieving improvements for low-confidence samples in the next round of training.

[0103] As the loss backpropagates, the parameters of the two models are updated synchronously through the optimizer and the shared parameters are adjusted in conjunction. This means that the experience gained by the stable node processing model when facing common node data will complement the adaptability of the sensitive node processing model to new nodes. On the other hand, the dynamic rule features extracted by the sensitive node processing model will also be reflected in the shared part, allowing the system to continuously integrate new knowledge and adapt to more changes. For example, suppose a new "hydrogen energy storage" node appears in the manufacturing field. The stable node processing model may produce low-confidence predictions due to the lack of relevant prior information, but the sensitive node processing model can identify and extract keywords such as "hydrogen fuel cell" or "hydrogen production equipment", and then pass these dynamic characteristics back to the stable node processing model, so that it can gradually learn to make more accurate judgments on this type of new node.

[0104] Specifically, the joint loss function is , is a hyperparameter used to control the contribution of each part loss to training;

[0105] is the classification loss, which is used to optimize the classification task of whether enterprises and institutions belong to nodes. ;in, is the true label, indicating whether the enterprise or institution belongs to a certain target node; is the probability output predicted by the model; N is the number of samples;

[0106] To stabilize the consistency loss of the node processing model and the sensitive node output model, ,in, To stabilize the output distribution of the node processing model, Processing model output distribution for sensitive nodes; Kullback–Leibler divergence, referred to as KL divergence, is used to measure the difference between two probability distributions; minimize The goal is to make the outputs of the two models consistent in semantic classification.

[0107] Supplement the loss for dynamic rules, ,in, To stabilize the node feature embedding of the node processing model, Handle dynamic feature embedding of models for sensitive nodes.

[0108] In order to better integrate the collaborative training effects of the stable node processing model and the sensitive node processing model, the present invention introduces a joint loss function L consisting of three parts of loss, and uses a linear weighted form to integrate the priorities of different targets. The first item is the classification loss L1, which is used to measure the difference between the current model prediction and the true label, thereby optimizing the classification accuracy of whether the enterprise or institution belongs to the target node. This part usually takes the form of cross entropy, which improves the model's ability to distinguish between positive and negative samples by minimizing the negative log-likelihood on the training set. In practical applications, for example, when we want to determine whether a "new energy organization" belongs to the "photovoltaic industry chain", the predicted probability given by the model will be compared with the true label. If the classification is wrong or the uncertainty is high, a larger loss value will be generated, which will prompt the model to adjust the parameters in the back propagation, in order to reduce similar classification errors in subsequent training.

[0109] The second term, L2, is the consistency loss. It uses KL divergence to measure the difference between the output distribution P1 of the stable node processing model and the output distribution P2 of the sensitive node processing model. Its core purpose is to encourage the two models to maintain a certain degree of consistency in their output distributions for the same input sample. This is especially true in edge cases between high and low confidence. If the predictions of the association between the stable and sensitive node processing models differ significantly, L2 will increase. By adjusting the hyperparameter β to increase or decrease the constraint on consistency, the synergy between the two models can be strengthened in the early stages of training, or the system can be made more flexible by allowing each to leverage its strengths in the later stages. For example, if the stable node processing model achieves high accuracy on a common node such as the "hydrogen energy industry chain," and the sensitive node processing model agrees with this result, the output probability distributions of the two models will converge. However, if the sensitive node processing model reaches a different conclusion based on new features or new semantic analysis, L2 will increase, forcing the two models to learn from each other during the backpropagation phase.

[0110] The third item is the dynamic rule supplement loss L3, which mainly focuses on the difference in feature representation between the stable node processing model and the sensitive node processing model, and uses vector distance to measure this difference. Specifically, It is the vector representation of the node feature latent part of the stable node processing model. It is the dynamic feature embedding generated by the sensitive node processing model. The squared Euclidean distance between the two reflects the degree of deviation of the learned node semantics in the high-dimensional space. By minimizing this distance during training, the stable node processing model can gradually absorb the new node information or special features of low-confidence samples provided by the sensitive node processing model, helping it to better adapt to unseen scenarios in subsequent predictions. For example, when the sensitive node processing model finds that "hydrogen production equipment" and "hydrogen fuel cell" are strongly correlated features, it will incorporate the corresponding dynamic keywords into the model. ; If the stable node handles the model If there is still a lack of sensitivity to these features, the loss value will increase, driving the system to update the weights during iteration so that the representations of the two gradually converge.

[0111] In general, the losses in these three parts have different functions: the classification loss L1 ensures the accuracy of the core task, the consistency loss L2 promotes mutual learning of the output distributions of the two models, and the dynamic rule supplement loss L3 helps the stable node processing model to continuously absorb new information. By setting different hyperparameters α, β, and γ to regulate the contribution weights of the three in training, you can flexibly switch according to the focus of the business scenario: if you need to enhance overall stability, you can increase the proportion of classification loss; if new nodes appear frequently, you need to increase the proportion of L2 and L3 to maximize the use of dynamic rules generated by the sensitive node processing model. Ultimately, this multi-objective joint optimization method can significantly improve the adaptability of the system in practical applications, ensuring performance on high-confidence data while capturing low-confidence data or new node information in a timely manner.

[0112] Furthermore, the S304 includes the following sub-steps S3041-S3043.

[0113] S3041. Combine the output of the sensitive node processing model and compare it with the results of the stable node processing model to calculate the consistency loss.

[0114] S3042. Use dynamic rules to supplement features to optimize feature generation for stable node processing models.

[0115] S3043. Dynamically adjust the weights of consistency loss and dynamic rule supplement loss according to the model training progress.

[0116] To better achieve joint optimization between the stable and sensitive node processing models, the outputs of the two models need to be combined and mutually corrected at different levels. First, by comparing the predictions of the sensitive node processing model for the same input sample with the classification results of the stable node processing model, it is possible to measure whether there are significant differences in classification or association judgment between the two. If the difference is large, it indicates that there is a significant divergence in the interpretation of the sample between the two models, and further convergence of the output distributions through consistency loss is necessary. For example, when the stable node processing model has high stability when processing traditional manufacturing nodes, while the sensitive node processing model captures recently updated industry information and outputs a different probability distribution, it is necessary to calculate the KL divergence (i.e., consistency loss) of their outputs and appropriately adjust the parameters during backpropagation to gradually achieve relatively consistent judgments on similar samples between the two models.

[0117] After completing the calculation of the consistency loss, the dynamic rule features generated by the sensitive node processing model can be used to optimize the internal representation of the stable node processing model. Especially for low-confidence samples with new industry trends or unknown field characteristics, the sensitive node processing model often extracts additional semantic clues from dynamic keywords or new business descriptions. This information will be supplemented to the stable node processing model in the form of embedded vectors or dynamic rules, thereby helping the latter to improve the recognition accuracy of similar new nodes in future predictions. For example, if the sensitive node processing model learns certain key phrases or feature vectors from a batch of "new energy materials" related samples, these features can be integrated into the node feature latent part of the stable node processing model, allowing the model to gradually master the ability to understand emerging directions such as "hydrogen energy storage" or "solid-state batteries."

[0118] During the training process, the weights of consistency loss and dynamic rule supplementation loss can also be dynamically adjusted according to the progress of training. If the system focuses on the collaboration between models in the early stages, the proportion of consistency loss can be increased to ensure that the output distribution of the two models is as consistent as possible, so as to quickly build a globally stable classification or association framework. When the training enters the middle and late stages, as the understanding of the two models gradually converges, the proportion of consistency loss can be gradually reduced, and the dynamic rule supplementation loss can be strengthened instead, so that the stable node processing model has more room to receive and integrate new features, thereby achieving more significant improvements in refined adaptation or new node expansion. Through such an adaptive adjustment mechanism, the system can eventually strike a balance between stability and scalability, and maintain a high level of prediction accuracy and association discovery capabilities even when rapid changes occur in different industrial fields or new nodes continue to emerge.

[0119] An embodiment of the present application further discloses a system for acquiring related companies based on enterprise and institution information, comprising a processor running a program of any one of the above-described methods for acquiring related companies based on enterprise and institution information.

[0120] An embodiment of the present application further discloses a storage medium storing a program for any one of the above-mentioned methods for obtaining affiliated companies based on enterprise and institution information.

[0121] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for obtaining affiliated companies based on enterprise and institution information, characterized in that: The following steps are involved: S1. Obtain raw data, clean and calculate confidence levels for enterprise and institution information, filter low-confidence data, and obtain high-confidence data; S2. Perform semantic analysis and confidence verification on high-confidence data to screen out credible related data; S3. Input the trusted association data into the stable node processing language model to obtain the final association result. The stable node processing model and the sensitive node processing model are mutually trained. The stable node processing model is used to train common nodes and high-confidence samples, while the sensitive node processing model is used to train new nodes or low-confidence samples. The stable node processing model and the sensitive node processing model share the same input data. The sensitive node processing model introduces an attention mechanism to enhance the learning of the characteristics of new nodes. The stable node processing model and the sensitive node processing model share a joint loss function, and the results of the stable node processing model and the sensitive node processing model are interactively fed back, and the parameters are adjusted according to the joint loss function. The mutual training step of the sensitive node processing model and the stable node processing model includes: S301. Separately train the sensitive node processing model and the stable node processing model; S302. Input the current sample into the stable node processing model for prediction and calculate the classification loss; S303. The low confidence samples are transferred to the sensitive node processing model to generate dynamic rule characteristics; S304. Calculate the difference between the predicted probability and the true label based on the joint loss function; S305. Based on the joint loss function, calculate the gradient through back propagation; S306. The parameters of the sensitive node processing model and the stable node processing model are updated simultaneously by the optimizer, and the parameter sharing part is adjusted synchronously between the two; The joint loss function is , is a hyperparameter used to control the contribution of each part loss to training; is the classification loss, which is used to optimize the classification task of whether enterprises and institutions belong to nodes. ;in, is the true label, indicating whether the enterprise or institution belongs to a certain target node; is the probability output predicted by the model; N is the number of samples; To stabilize the consistency loss of the node processing model and the sensitive node output model, ,in, To stabilize the output distribution of the node processing model, Processing model output distribution for sensitive nodes; is the Kullback–Leibler divergence, which measures the difference between two probability distributions; Supplement the loss for dynamic rules, ,in, To stabilize the node feature embedding of the node processing model, Handle dynamic feature embedding of models for sensitive nodes.

2. The method for obtaining related companies based on enterprise and institution information according to claim 1, characterized in that: The structural design steps of the stable node processing model include: The embedding layer is used to convert the input text into a context-sensitive high-dimensional vector representation using BERT or RoBERTa. The model input is the concatenated text to be processed, which includes enterprise information and node descriptions. The enterprise information includes the enterprise name, enterprise profile, and a list of main business keywords. The node description includes the node name and node details. The feature extraction layer, composed of multiple layers of Transformers, is used to capture the semantic relationship between enterprise information and node descriptions; The output layer is used to output the classification head and association probability. The classification head uses BERT's output [CLS] tag as the overall semantic representation, and the association probability passes through a fully connected layer and a softmax function.

3. The method for obtaining related companies based on enterprise and institution information according to claim 2, characterized in that: The structural design steps of the sensitive node processing model include: An embedding layer, which uses BERT or RoBERTa to convert input text into a context-sensitive high-dimensional vector representation and uses a dynamic rule embedding module to embed the dynamic features of new nodes into the model. The model input is samples whose probability output by the stable node processing model is below a certain threshold, including enterprise information and node descriptions. The enterprise information includes the enterprise name, enterprise profile, main business keyword list, and historical confidence information. The node description includes the new node name, new node details, and dynamic keywords. The historical confidence information is the predicted distribution of low-confidence samples obtained from the output of the second language model. The feature extraction layer, composed of multiple layers of Transformers, is used to strengthen the semantic association between enterprise information and new node descriptions by introducing a multi-head attention mechanism. The association between dynamic keywords and enterprise information is processed by a dedicated attention head. The output layer includes a classification task head and a dynamic rule generation head. The classification task head is used to determine whether an enterprise or institution belongs to a new node, and the dynamic rule generation head is used to generate dynamic features that can be used for learning by the stable node processing model.

4. The method for obtaining related companies based on enterprise and institution information according to claim 3, characterized in that: The step of strengthening the semantic association between enterprise and institution information and new node description by introducing a multi-head attention mechanism includes the following steps: Split the input embedding into multiple subspaces, i.e. multiple heads; Each head independently calculates attention weights to capture different semantic features; The outputs of all heads are concatenated and linearly transformed.

5. The method for obtaining related companies based on enterprise and institution information according to claim 1, characterized in that: The S304 includes the following sub-steps: S3041. Combine the output of the sensitive node processing model and compare it with the results of the stable node processing model to calculate the consistency loss; S3042. Using dynamic rules to supplement features to optimize feature generation of stable node processing models; S3043. Dynamically adjust the weights of consistency loss and dynamic rule supplement loss according to the model training progress.

6. The method for obtaining related companies based on enterprise and institution information according to claim 5, characterized in that: The S1 comprises the following sub-steps: S11. Obtaining raw data and performing preprocessing; S12. Use the first pre-trained model to perform semantic analysis, map the enterprise information and industrial chain into the same semantic space, and determine the initial association result based on the semantic similarity; S13. Perform confidence verification on the initial association results to screen out credible association data.

7. A system for obtaining related companies based on enterprise and institution information, characterized in that: The method comprises a processor in which a program for obtaining related companies based on enterprise and institution information as described in any one of claims 1 to 6 is run.

8. A storage medium, characterized in that: A program is stored for the method of obtaining related companies based on enterprise and institution information as described in any one of claims 1 to 6.

Images