Enterprise division method and device

By segmenting and clustering corporate text data, building an industrial map, and combining models and rules, the problem of low corporate classification accuracy is solved, and higher accuracy and coverage is achieved.

CN119066492APending Publication Date: 2024-12-03PEKING UNIV +1

Patent Information

Application Number
CN202411131035.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-16
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing technology is difficult to accurately divide enterprises into corresponding industrial chain nodes, especially in short text classification processing, manual classification costs are high, expert rules are insufficient coverage, and model-based methods fail to consider business or empirical nature. personalized processing leads to low classification accuracy.

Method used

By obtaining multi-dimensional enterprise text data, performing word segmentation processing and clustering, building an industrial map, using pre-trained enterprise classification model and graph neural network for analysis, and combining regular matching and similarity matching rules, the enterprise classification results are determined.

Benefits of technology

The accuracy and coverage of enterprise classification are improved, combined with models and rules, higher accuracy and flexibility are achieved, and the problem of difficulty in accurately dividing enterprise classification into industrial chain nodes is solved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066492A_ABST
    Figure CN119066492A_ABST
Patent Text Reader

Abstract

The invention discloses an enterprise division method and device. The method comprises the steps of obtaining multiple groups of enterprise text data corresponding to multiple target enterprises associated with a target industry chain; performing word segmentation processing on each group of enterprise text data to obtain multiple groups of target word segmentation sets; constructing a target industry map according to the multiple groups of target word segmentation sets, and analyzing the target industry map by using a pre-trained enterprise classification model to obtain a first classification result of each target enterprise; analyzing the multiple groups of target word segmentation sets according to a preset enterprise classification rule to obtain a second classification result of each target enterprise; target classification results of all the target enterprises are determined according to the first classification results and the second classification results, and all the classification results are used for indicating the corresponding industry nodes of the target enterprises in the target industry chain. The technical problem that a related classification scheme is difficult to accurately divide enterprise related data to corresponding industrial chain nodes according to the enterprise related data is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of text processing technology, and in particular to a method and device for dividing an enterprise. Background Art

[0002] Industrial chain analysis refers to the analysis and research of the upstream and downstream of a specific industry to understand the industrial structure, competitive situation and industrial value. It is crucial for enterprises themselves, management departments and service agencies. For enterprises themselves, industrial chain analysis can better grasp market dynamics, formulate competitive strategies and control risks, and achieve cost reduction and efficiency improvement; for management departments, using industrial chain analysis to understand the current status of industrial development can provide support for the formulation of industrial policies and industrial development strategies, and achieve sustainable development; for service agencies, such as financial institutions, with the help of industrial chain analysis to gain insight into the market, they can discover potential investment opportunities and improve profitability.

[0003] The basis of industrial chain analysis is to accurately classify enterprises into industrial chain nodes. Since industrial chain nodes are a summary of the same type of products, and a company usually produces and operates multiple products, in principle, a company can be classified into multiple industrial chain nodes. Classifying enterprises into industrial chain nodes involves the processing of text data, especially short texts, so the enterprise classification task can be regarded as a short text multi-classification task.

[0004] Common short text classification methods include manual classification, expert rule-based classification, and model-based classification. Manual classification of short texts is time-consuming and time-consuming; the expert rule-based method can customize the matching logic, especially the business or empirical processing, but the disadvantage is that the constraint space of the rule definition is relatively limited, resulting in insufficient coverage of text classification; the model-based classification method makes full use of the prior knowledge learned from large-scale data, but the disadvantage is that it does not consider business or empirical personalized processing, resulting in the difficulty of meeting the accuracy requirements of text classification.

[0005] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0006] The embodiments of the present application provide a method and device for dividing an enterprise, so as to at least solve the technical problem that it is difficult for relevant classification schemes to accurately divide the enterprise into corresponding industrial chain nodes based on relevant enterprise data.

[0007] According to one aspect of an embodiment of the present application, a method for enterprise segmentation is provided, including: obtaining multiple groups of enterprise text data corresponding to multiple target enterprises associated with a target industrial chain, wherein each group of enterprise text data includes text data of different dimensions for describing the industries involved in the enterprise; performing word segmentation processing on each group of enterprise text data to obtain multiple groups of target word segmentation sets; constructing a target industry map based on the multiple groups of target word segmentation sets, and analyzing the target industry map using a pre-trained enterprise classification model to obtain a first classification result for each target enterprise; analyzing the multiple groups of target word segmentation sets according to preset enterprise classification rules to obtain a second classification result for each target enterprise, wherein the type of enterprise classification rule includes at least one of the following: regular matching rule, similarity matching rule; determining a target classification result for each target enterprise based on the first classification result and the second classification result, wherein the target classification result is used to indicate the industry node corresponding to the target enterprise in the target industrial chain.

[0008] Optionally, multiple groups of enterprise text data corresponding to multiple target enterprises associated with the target industrial chain are obtained, including: determining the industry classification identifier associated with the target industrial chain in the preset economic industry classification standard; obtaining multiple groups of enterprise text data corresponding to multiple target enterprises matching the industry classification identifier, wherein each target enterprise corresponds to a group of enterprise text data, and the text dimensions in each group of enterprise text data include at least one of the following: business scope, software copyright title, patent title, and patent abstract.

[0009] Optionally, each group of enterprise text data is segmented to obtain multiple groups of target segmented word sets, including: for the business scope text in each group of enterprise text data, the punctuation marks in the business scope text are adjusted according to the preset punctuation mark usage standard to obtain a first text; the first text is segmented based on other punctuation marks except the first type of punctuation marks to obtain a first segmented word set, and the first type of punctuation marks include: colon and semicolon; determine all the segmented words in the first segmented word set that match the preset first keyword to form a second segmented word set, and the first keyword is a keyword associated with the target industrial chain ; Segment the words in the second word segmentation set based on the first type of punctuation marks to obtain a third word segmentation set; delete the verbs in each word in the third word segmentation set to obtain the first type of target word segmentation set corresponding to the enterprise text data; and / or, for each set of software title text in the enterprise text data, delete the enterprise name, region name and second keyword in the software title text to obtain the second type of target word segmentation set corresponding to the enterprise text data, wherein the second keyword includes: a, based on, of; and / or, for each set of patent title text in the enterprise text data, delete the design patent name other fields in the title except the product name; for utility model patent titles whose title length is not greater than the first preset length, delete the third keyword in the utility model patent title, the third keyword includes: of, a kind of, applicable to, used for, with, with; for utility model patent titles whose title length is greater than the first preset length, use the TextRank algorithm to extract keywords in the utility model patent title; for invention patent titles whose title length is not greater than the first preset length and includes the fourth keyword, delete the third keyword in the invention patent title, the fourth keyword includes: method, comma; for invention patent titles whose title length is greater than the first preset length or includes the fourth keyword, use the TextRank algorithm to extract keywords in the invention patent title; use the processing result as the third type of target word segmentation set corresponding to the enterprise text data; and / or, for each set of patent abstract texts in the enterprise text data, delete the design patent abstract and the utility model patent abstract; use the preset word segmentation dictionary to perform word segmentation processing on the invention patent abstract text to obtain a fourth word segmentation set; delete the stop words in each word segmentation in the fourth word segmentation set to obtain a fourth type of target word segmentation set corresponding to the enterprise text data.

[0010] Optionally, the process of establishing the word segmentation dictionary includes: obtaining multiple historical invention patent abstract texts, deleting stop words in the historical invention patent abstract texts, and performing word segmentation processing on the historical invention patent abstract texts with the stop words deleted according to punctuation marks to obtain a fifth word segmentation set; determining that the word segmentations in the fifth word segmentation set whose length is less than a first preset length form a sixth word segmentation set, and extracting the word segmentations in the sixth word segmentation set based on a sliding window of a second preset length to construct a first word graph; deleting the word segmentations of non-target parts of speech in the first word graph to obtain a second word graph, wherein the target parts of speech include: nouns, verbs, gerunds, adjectives, English words, and subject-related vocabulary; for each pair of adjacent word segmentations in the second word graph, calculating the degree of aggregation between the two adjacent word segmentations, if the degree of aggregation is greater than a preset degree of aggregation threshold, the two adjacent word segmentations are combined into a first phrase; obtaining multiple second phrases from a preset source node, wherein the type of the source node includes at least one of the following: books, official websites, and term libraries; and constructing a word segmentation dictionary using the multiple first phrases and second phrases.

[0011] Optionally, calculating the degree of aggregation between two adjacent participles includes: calculating the degree of aggregation between two adjacent participles using the following formula: In the formula, a and b represent two adjacent participles, |a| represents the length of participle a, |b| represents the length of participle b, and cohesion a→b represents the degree of aggregation between participle a and participle b, represents the probability of word a and word b appearing at the same time, count(a→b) represents the frequency of word a pointing to word b, and count(all) represents the total number of word segments in the second word graph. represents the degree of aggregation of word a, n represents the number of words pointed to by word a, count(a→w i ) indicates that participle a points to participle w i The frequency of represents the indexing aggregation of word b, m represents the number of words pointing to word b, count(w i →b) represents the participle w i The frequency of pointing to participle b.

[0012] Optionally, a target industry map is constructed based on multiple groups of target word segmentation sets, including: clustering all the words in multiple groups of first-category target word segmentation sets corresponding to multiple groups of enterprise text data using a density-based clustering algorithm, and taking each obtained word segmentation set as a business scope node; clustering all the words in multiple groups of second-category target word segmentation sets corresponding to multiple groups of enterprise text data using software copyright classification numbers, and taking each obtained word segmentation set as a software copyright title node; clustering all the words in multiple groups of third-category target word segmentation sets corresponding to multiple groups of enterprise text data using international patent classification numbers, and taking each obtained word segmentation set as a patent title node; clustering all the words in multiple groups of fourth-category target word segmentation sets corresponding to multiple groups of enterprise text data using international patent classification numbers, and taking each obtained word segmentation set as a patent abstract node; taking each target enterprise as an enterprise node, and constructing a target industry map based on the association between each enterprise node and each business scope node, software copyright title node, patent title node, and patent abstract node.

[0013] Optionally, a pre-trained enterprise classification model is used to analyze the target industry map to obtain a first classification result for each target enterprise, including: using the GloVe algorithm to encode each business scope node, software title node, and patent title node, and using a bidirectional long short-term memory algorithm to encode each patent abstract node to obtain a vectorized representation of the target industry map; using the graph neural network module in the enterprise classification model to analyze the vectorized representation of the target industry map to obtain enhanced word vectors corresponding to each target enterprise; using the classification module in the enterprise classification model to analyze each enhanced word vector to obtain a first classification result corresponding to each target enterprise, wherein the classification module performs classification based on a probability conversion function, and the type of the probability conversion function includes at least one of the following: sigmoid, softmax, and logistic.

[0014] Optionally, the training process of the enterprise classification model includes: constructing an initial model, wherein the initial model includes a graph neural network module and a classification module; obtaining multiple historical industry maps as training samples, and determining the classification results corresponding to the enterprises in each historical industry map as sample labels; using the training samples to iteratively train the initial model, wherein, when each training batch is completed, the F1 value of each enterprise is calculated based on the classification results and sample labels of each enterprise output by the model, and the average F1 value of all enterprises is calculated, and the model parameters are adjusted based on the average F1 value.

[0015] Optionally, multiple groups of target segmentation word sets are analyzed according to preset enterprise classification rules to obtain second classification results for each target enterprise, including: for each type of target segmentation word set corresponding to each target enterprise, the segmentations in the target segmentation word set of this type are sorted according to the frequency of occurrence, and the segmentations of a first preset proportion sorted in front form a segmentation word set to be tested; a standard segmentation word set corresponding to the target segmentation word set of this type is determined, wherein the standard segmentation word set includes preset standard segmentations corresponding to industrial nodes in the target industrial chain; for each segmentation word to be tested in the segmentation word set to be tested, a standard segmentation word matching the segmentation word to be tested is determined from the standard segmentation word set according to regular matching rules and / or similarity matching rules, and the standard segmentation word is used as the sub-classification result corresponding to the segmentation word to be tested; and multiple sub-classification results corresponding to multiple segmentations to be tested corresponding to the target enterprise are unified as the second classification result of the target enterprise.

[0016] Optionally, determining a standard participle that matches the participle to be detected from a standard participle set according to a regular matching rule and / or a similarity matching rule includes: for each standard participle in the standard participle set, determining a subset of the standard participle and the participle to be detected, and calculating the ratio of the length of the subset to the length of the participle to be detected; if the largest ratio is greater than a preset ratio threshold, determining that the standard participle corresponding to the largest ratio matches the participle to be detected; and / or, for each standard participle in the standard participle set, determining the similarity between the standard participle and the participle to be detected, the type of similarity including at least one of the following: Jaccard similarity, Dice similarity, overlap, cosine similarity, correlation coefficient, Hamming distance; if the largest similarity is greater than a preset similarity threshold, determining that the standard participle corresponding to the largest similarity matches the participle to be detected.

[0017] Optionally, the first classification result and the second classification result both include multiple sub-classification results, and the target classification results of each target enterprise are determined based on the first classification result and the second classification result, including: determining that the sub-classification results that exist in both the first classification result and the second classification result are target sub-classification results; for sub-classification results that only exist in the first classification result, sorting each sub-classification result according to the probability value corresponding to each sub-classification result predicted by the enterprise classification model, and determining that the sub-classification results with a second preset proportion ranked first are the target sub-classification results; for sub-classification results that only exist in the second classification result, sorting each sub-classification result according to the frequency of occurrence of each sub-classification result, and determining that the sub-classification results with a third preset proportion ranked first are the target sub-classification results; and unifying all target sub-classification results as the target classification result.

[0018] According to another aspect of an embodiment of the present application, there is also provided an enterprise classification device, including: an acquisition module, used to acquire multiple groups of enterprise text data corresponding to multiple target enterprises associated with a target industrial chain, wherein each group of enterprise text data includes text data of different dimensions for describing the industries involved in the enterprise; a word segmentation module, used to perform word segmentation processing on each group of enterprise text data to obtain multiple groups of target word segmentation sets; a first classification module, used to construct a target industry map based on the multiple groups of target word segmentation sets, and use a pre-trained enterprise classification model to analyze the target industry map to obtain a first classification result for each target enterprise; a second classification module, used to analyze the multiple groups of target word segmentation sets according to preset enterprise classification rules to obtain a second classification result for each target enterprise, wherein the type of enterprise classification rule includes at least one of the following: regular matching rule, similarity matching rule; a determination module, used to determine the target classification result of each target enterprise based on the first classification result and the second classification result, wherein the target classification result is used to indicate the industry node corresponding to the target enterprise in the target industrial chain.

[0019] According to another aspect of an embodiment of the present application, a computer program product is further provided, the computer program product comprising: a computer program, wherein the above-mentioned enterprise division method is implemented when the computer program is executed by a processor.

[0020] According to another aspect of an embodiment of the present application, an electronic device is further provided, the electronic device comprising: a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to execute the above-mentioned enterprise division method through the computer program.

[0021] In the embodiments of the present application, in terms of data usage and processing, different from the prior art that performs modeling and analysis on a single type of text, the present application obtains enterprise text data information in multiple dimensions and adopts different data preprocessing mechanisms according to the characteristics of different texts, which helps to improve the accuracy of subsequent classification; in the specific classification process, the model classification method can utilize the prior knowledge learned from large-scale data, and has high flexibility, while the rule-based classification method is simple and easy, and the matching logic can be personalized, especially the business or empirical processing can be customized, and the accuracy is high. The present application integrates model classification and rule-based classification solutions, which can effectively combine the advantages of the two methods, increase the coverage of enterprise classification, and improve the accuracy of enterprise classification, and effectively solve the technical problem that the relevant classification schemes are difficult to accurately divide the enterprise-related data into corresponding industrial chain nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0023] Figure 1 is a flowchart of an optional enterprise division method according to an embodiment of the present application;

[0024] Figure 2 is a schematic diagram of an optional construction of a word graph according to an embodiment of the present application;

[0025] Figure 3 It is a schematic diagram of an optional construction of an industry map according to an embodiment of the present application;

[0026] Figure 4 is a schematic diagram of an optional complete process of enterprise division according to an embodiment of the present application;

[0027] Figure 5 is a schematic structural diagram of an optional enterprise division device according to an embodiment of the present application;

[0028] Figure 6 It is a schematic diagram of the structure of an optional electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0030] It should be noted that the terms "first", "second", etc. in the specification, claims and drawings of the present application are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0031] In order to better understand the embodiments of the present application, some nouns or terms that appear in the description of the embodiments of the present application are first translated and explained as follows:

[0032] TextRank is a graph-based ranking algorithm for keyword extraction and document summarization. It is an improvement on the PageRank algorithm, a web page importance ranking algorithm. This method uses the co-occurrence information between words within a document to extract keywords and keyword groups, and uses an extractive automatic summarization method to extract the key sentences of the text.

[0033] GloVe (Global Vectors for Word Representation) is an algorithm that represents words as vectors. It combines the two methods of LSA (Latent Semantic Analysis) and Word2Vec, making use of both global features and local context features.

[0034] LSA,represents documents as a linear combination of a low-dimensional,semantic space.

[0035] Word2Vec is a method for generating word vectors. Its basic idea is that adjacent words have similar word vectors based on the context, including using context words to predict target words, and using target words to predict context words, thereby learning the semantic relationship between words.

[0036] DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a density-based clustering algorithm. Its basic idea is to find points with higher density and then connect sample points with similar higher density to form multiple clusters.

[0037] GNN (Graph Neural Networks) is a machine learning model used to process graph structures and is suitable for scenarios such as social analysis, bioinformatics, and social networks.

[0038] Softmax is the activation function of the multi-classification output layer, which can convert an m-dimensional real number vector into an m-dimensional probability distribution vector.

[0039] Example 1

[0040] According to an embodiment of the present application, a method for dividing an enterprise is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0041] Figure 1is a flow chart of a method for dividing an enterprise according to an embodiment of the present application, such as Figure 1 As shown, the method comprises the following steps:

[0042] Step S102, obtaining multiple groups of enterprise text data corresponding to multiple target enterprises associated with the target industrial chain, wherein each group of enterprise text data includes text data of different dimensions for describing the industry involved in the enterprise;

[0043] Step S104, performing word segmentation processing on each group of enterprise text data to obtain multiple groups of target word segmentation sets;

[0044] Step S106, constructing a target industry map based on multiple sets of target word segments, and analyzing the target industry map using a pre-trained enterprise classification model to obtain a first classification result for each target enterprise;

[0045] Step S108, analyzing the multiple target word segmentation sets according to the preset enterprise classification rules to obtain the second classification results of each target enterprise, wherein the type of enterprise classification rule includes at least one of the following: regular matching rule, similarity matching rule;

[0046] Step S110, determining a target classification result for each target enterprise based on the first classification result and the second classification result, wherein the target classification result is used to indicate the industrial node corresponding to the target enterprise in the target industrial chain.

[0047] The following describes each step of the enterprise segmentation method in conjunction with the specific implementation process.

[0048] As an optional implementation, when obtaining multiple groups of enterprise text data corresponding to multiple target enterprises associated with the target industrial chain, the industry classification identifier associated with the target industrial chain in the preset economic industry classification standard can be first determined; then multiple groups of enterprise text data corresponding to multiple target enterprises matching the industry classification identifier are obtained, wherein each target enterprise corresponds to a group of enterprise text data, and the text dimensions in each group of enterprise text data include at least one of the following: business scope, software copyright title, patent title, and patent abstract.

[0049] Among them, the above-mentioned economic industry classification standard can adopt the national economic industry classification. The national economic industry classification can be used to screen the data to control the relevance of the included enterprise text data and the target industrial chain to be built, and improve the accuracy of matching enterprises with industrial chain nodes. By obtaining enterprise text data information in multiple dimensions, the accuracy of subsequent classification can be further improved.

[0050] When building a specific industrial chain, multiple national economic industry classifications can be selected. For example, for the construction of the automobile industry chain, you can screen according to the national economic industry classifications such as automobile manufacturing, automobile wholesale and retail, and on this basis, divide the screened enterprises into the corresponding industrial nodes of the automobile industry chain. The industrial node can also be referred to as the standard classification in this application. In this application, whether it is the first classification result, the second classification result, or the final target classification result obtained during the processing process, they all correspond to the standard classification, that is, they indicate the corresponding industrial node of the target enterprise in the target industrial chain.

[0051] Table 1 takes the automobile industry as an example and shows a set of enterprise text data corresponding to a certain enterprise.

[0052] Table 1

[0053]

[0054] It can be understood that each set of enterprise text data includes text data of different dimensions. In order to ensure the accuracy of subsequent classification, different word segmentation methods need to be used to process them. The specific implementation methods are as follows.

[0055] Optionally, for the business scope text in each group of enterprise text data, the punctuation marks in the business scope text are adjusted according to a preset punctuation mark usage standard to obtain a first text; the first text is segmented based on punctuation marks other than the first category of punctuation marks to obtain a first segmentation set, and the first category of punctuation marks includes: colon and semicolon; all segmentations in the first segmentation set that match a preset first keyword are determined to form a second segmentation set, and the first keyword is a keyword associated with the target industrial chain; the segmentations in the second segmentation set are segmented based on the first category of punctuation marks to obtain a third segmentation set; the verbs in each segmentation in the third segmentation set are deleted to obtain a first category of target segmentation set corresponding to the enterprise text data.

[0056] Among them, the standardized processing of punctuation marks is to ensure the reliability of subsequent word segmentation. For example, all half-width symbols are replaced with corresponding full-width symbols, and semicolons are uniformly replaced with commas, etc.; then, the standardized first text is initially segmented according to other punctuation marks excluding colons and semicolons, and the word segmentations related to the target industrial chain are retained, such as retaining the word segmentations containing keywords such as manufacturing, processing, and production in the automotive manufacturing industry; then the retained word segmentations are further segmented according to colons and semicolons, and finally the verb expressions are deleted.

[0057] Optionally, for each set of software title text in corporate text data, delete the corporate name, region name and second keyword in the software title text to obtain a second type of target word segmentation set corresponding to the corporate text data, where the second keyword includes: a, based on, and.

[0058] Specifically, considering that the commonly used expressions of software copyright titles are: "application scenario / name + software series", or "based on application scenario + software series", or "company name + application scenario + software series", such as: parking lot automatic license plate recognition management system, VR panoramic virtual exhibition hall 3D multi-person real-time interactive display platform, Internet-based hot pot base online sales platform, etc., they generally end with platform, system, software, engine. In order to avoid the interference of the company name in the software copyright title on the matching of the company and the industrial chain node, the largest common substring of the company name and the software copyright title can be taken and deleted; in order to avoid the influence of the regional name on product identification, the regional entity recognition can be used for the remaining software copyright titles, and the regional name can be deleted; it is also possible to delete the "one kind" in the software copyright title that starts with "a kind", and delete the "based on" and the "of" in the software copyright title that starts with "based on".

[0059] Optionally, for the patent title text in each group of enterprise text data, delete other fields in the design patent title except the product name; for the utility model patent title whose title length is not greater than the first preset length, delete the third keyword in the utility model patent title, and the third keyword includes: of, a kind of, applicable to, used for, with, with; for the utility model patent title whose title length is greater than the first preset length, use the TextRank algorithm to extract the keywords in the utility model patent title; for the invention patent title whose title length is not greater than the first preset length and includes the fourth keyword, delete the third keyword in the invention patent title, and the fourth keyword includes: method, comma; for the invention patent title whose title length is greater than the first preset length or includes the fourth keyword, use the TextRank algorithm to extract the keywords in the invention patent title; and use the processing result as the third category target word segmentation set corresponding to the enterprise text data.

[0060] Specifically, considering that the title of a design patent is usually "product name + product type / product use", that is, the length of the words related to the product in the title is short, mostly between 2-10, so only the product name in the title of the design patent can be retained. The title of a utility model patent usually appears in the form of "product name + product use" or "product name", involving physical products, and the granularity is close to the standard product name, and will not contain specific technologies or methods. Therefore, for the title of the utility model patent with a word length of no more than 10, delete the "of", and delete the opening words such as "a kind", "applicable to", "used for", "with", "with"; otherwise, use the keyword extraction method, such as the TextRank method in the jieba library. The title of an invention patent generally contains "product name", "method", "device", "system", etc. Therefore, for the title with a word length greater than 10 or the title contains "method" or ",", the keyword extraction method is used; otherwise, delete the word "of", and delete the opening words such as "a kind", "applicable to", "used for", "with", "with". It should be noted that the first preset length of 10 here is only used as an example, and the specific value can be adjusted as needed.

[0061] Optionally, for the patent abstract text in each group of enterprise text data, delete the design patent abstract and the utility model patent abstract; use a preset word segmentation dictionary to perform word segmentation on the invention patent abstract text to obtain a fourth word segmentation set; delete stop words in each word segmentation in the fourth word segmentation set to obtain a fourth category target word segmentation set corresponding to the enterprise text data.

[0062] Specifically, considering that the abstract of a design patent often describes the design features, uses, views, colors, etc. of the product appearance, and also points out the product name in the title, the abstract of a design patent is not used. Most of the abstracts of utility model patents will involve the product name in the title, and some will point out the technical field involved, and then they will mostly describe the shape, structure or technical solution of the product, such as explaining the specific composition and function of the device, material, structure, etc. involved in the patent, so the abstract of a utility model patent is also not used. The abstract of an invention patent has no obvious characteristics. Considering that its text usually involves new technologies, new concepts and new statements, the embodiment of this application proposes a solution to construct a word segmentation dictionary to perform word segmentation, thereby avoiding the situation where the word segmentation granularity does not meet the requirements due to conventional word segmentation schemes, and can retain the effective information of the text to the greatest extent.

[0063] As an optional implementation, a word segmentation dictionary may be established through the following steps.

[0064] First, multiple historical invention patent abstract texts are obtained, stop words in the historical invention patent abstract texts are deleted, and the historical invention patent abstract texts with stop words deleted are segmented according to punctuation marks to obtain a fifth segmentation set; the segmentations in the fifth segmentation set whose length is less than the first preset length are determined to form a sixth segmentation set, and the segmentations in the sixth segmentation set are extracted based on a sliding window of a second preset length to construct a first word graph. Among them, the second preset length of the sliding window can be adjusted according to actual needs. The embodiment of the present application takes 2 as an example for illustration. Figure 2 A schematic diagram of a constructed word graph.

[0065] Considering the consistency dimension of the part of speech of the participles in the phrases, generally only the participles of certain fixed parts of speech can be combined into phrases. Therefore, the participles of non-target parts of speech in the first word graph can be deleted to obtain the second word graph, in which the target parts of speech include: nouns, verbs, gerunds, adjectives, English words, and related vocabulary of biology, chemistry, geography, mathematics, physics and other subjects.

[0066] Afterwards, for each pair of adjacent word segments in the second word graph, the degree of aggregation between the two adjacent word segments is calculated. If the degree of aggregation is greater than a preset degree of aggregation threshold, the two adjacent word segments are combined into a first phrase. The preset degree of aggregation threshold can be set based on experience and is not specifically limited here. The degree of aggregation between two adjacent word segments can be calculated using the following formula:

[0067]

[0068] In the formula, a and b represent two adjacent participles, |a| represents the length of participle a, and |b| represents the length of participle b. Generally, the longer the participle is, the less likely it is to be constructed into a new phrase; cohesion a→b represents the degree of aggregation between participle a and participle b; p(a→b) represents the probability of participle a and participle b appearing at the same time, which is used to measure the co-occurrence of participle a and participle b. count(a→b) indicates the frequency of word a pointing to word b, and count(all) indicates the total number of words in the second word graph; Indicates the degree of aggregation of participle a. The larger the value, the higher the probability that participle a and participle b are combined into a phrase. n represents the number of participles pointed to by participle a, count(a→w i ) indicates that participle a points to participle w i frequency; Indicates the indexing aggregation of word b. The larger the value, the higher the probability that word a and word b are combined into a phrase. m represents the number of participles pointing to participle b, count(w i →b) represents the participle wi The frequency of pointing to participle b.

[0069] By leveraging the idea of ​​key phrase extraction, the embodiment of the present application constructs a graph-based aggregation index to combine participles that have been screened by part of speech, fully considering the consistency and co-occurrence of the participle combination, and can effectively and quickly merge the participles to discover new phrases. At the same time, it can provide support for subsequent text classification through continuous accumulation and optimization.

[0070] After obtaining multiple first phrases, multiple second phrases can also be obtained from a preset source node, wherein the type of the source node includes at least one of the following: books, official websites, and term libraries, and the books here can be authoritative books in the relevant field. Finally, a word segmentation dictionary is constructed using the multiple first phrases and the second phrases. By constructing a dictionary for word segmentation, professional terms are retained and new words can be included, which can improve the accuracy of word segmentation to a certain extent.

[0071] Table 2 shows the four types of target word segmentation sets obtained after word segmentation of the enterprise text data in Table 1.

[0072] Table 2

[0073]

[0074]

[0075] When using two-dimensional tables to classify and predict enterprises in related schemes, each dimensional data is regarded as an isolated island, and the correlation between the enterprise and other enterprises is not taken into account, and the data cannot be fully utilized. Therefore, the embodiment of the present application proposes to construct an industrial map with an "enterprise-multi-dimensional data-enterprise" structure, which to a certain extent brings together enterprises with higher similarity, and makes full use of the information presented by the multi-dimensional data to improve the accuracy of enterprise classification.

[0076] As an optional implementation method, the target industry map can be constructed in the following ways:

[0077] Considering that there are certain similarities in the business scopes of enterprises, clustering similar business scopes into a set of business scopes can make the degree of aggregation between enterprises higher and improve the accuracy of subsequent classification using classification algorithms. Therefore, a density-based clustering algorithm can be used to cluster all the segmented words in multiple groups of first-class target segmented word sets corresponding to multiple groups of enterprise text data, and each segmented word set obtained is used as a business scope node. For example, the GloVe algorithm can be used to learn the segmented words in the first-class target segmented word set corresponding to each target enterprise, obtain the encoding of the business scope of each target enterprise, and then use the DBSCAN method to cluster the business scope to obtain multiple business scope nodes.

[0078] Software works are related to computer software and carry software work classification numbers. Therefore, the software work classification numbers can be used to cluster all the word segments in multiple sets of second-category target word segments corresponding to multiple sets of enterprise text data, and each set of word segments obtained can be used as a software work title node.

[0079] For patent titles, the IPC (International Patent Classification) number can be used to cluster all the word segments in multiple sets of third-category target word segments corresponding to multiple sets of enterprise text data, and each set of word segments obtained can be used as a patent title node. When building traditional industries, the main patent classification number can be directly referenced; if it involves emerging industries, the patent titles can be clustered in combination with the emerging industry classification and the reference relationship between the international patent classification to obtain multiple patent title nodes.

[0080] Similar to processing patent titles, IPC numbers can also be used to cluster all the word segments in multiple sets of fourth-category target word segments corresponding to multiple sets of corporate text data, and each obtained set of word segments is used as a patent abstract node.

[0081] Finally, each target enterprise is taken as an enterprise node, and the target industry map is constructed based on the relationship between each enterprise node and each business scope node, software copyright title node, patent title node, and patent abstract node. Figure 3 A structural diagram of an optional industry map is shown.

[0082] After obtaining the target industry map, the pre-trained enterprise classification model can be used to analyze the target industry map to obtain the first classification results of each target enterprise.

[0083] Since computers cannot process text data directly, the text needs to be converted into numerical form. Word embedding technology maps each word to a point in a real vector space through vectorization, so that the semantic relationship between words can be expressed and calculated in a numerical way. Considering that the business scope, software title and patent title have no context and the text is short, the GloVe algorithm can be used to encode each business scope node, software title node and patent title node; and the patent abstract itself has contextual expression and the text is generally long, so the Bi-LSTM (Bi-directional Long Short Term Memory) algorithm can be used to encode each patent abstract node to obtain a vectorized representation of the target industry map.

[0084] Afterwards, the GNN module in the enterprise classification model is used to analyze the target industry map represented by vectorization to obtain the enhanced word vectors corresponding to each target enterprise; then the classification module in the enterprise classification model is used to analyze each enhanced word vector to obtain the first classification result corresponding to each target enterprise, wherein the classification module is based on the probability conversion function for classification, and the types of probability conversion functions include at least one of the following: sigmoid, softmax, logistic. Taking the softmax function as an example, it can normalize the word embedding vector enhanced by the GNN model to realize the prediction of the corresponding industrial chain node of the enterprise. The specific formula is as follows:

[0085]

[0086] In the formula, vector x has k dimensions, and S(·) converts x i Mapped to the (0,1) interval to form a probability distribution.

[0087] Optionally, the above-mentioned enterprise classification model can be obtained through the following training process: constructing an initial model, wherein the initial model includes a GNN module and a classification module; obtaining multiple historical industry maps as training samples, and determining the classification results corresponding to the enterprises in each historical industry map as sample labels, wherein the historical industry maps here can be multiple industry maps pre-established according to the above-mentioned industry map construction process, and marking the standard classifications corresponding to the enterprises in the historical industry maps one by one as sample labels; and using the training samples to iteratively train the initial model.

[0088] Considering that a company may be classified into multiple standard classifications, this application specially designs enterprise-dimensional evaluation indicators to evaluate the prediction results. Specifically, when each training batch is completed, the F1 value of each company is calculated based on the classification results and sample labels of each company output by the model, and the average F1 value of all companies is calculated. The average F1 value is used as the final evaluation indicator, and the model parameters are adjusted based on the average F1 value. The specific evaluation indicator calculation formula is as follows:

[0089]

[0090] In the formula, represents the micro F1 of the i-th enterprise, which is calculated as follows: represents the micro-precision of the i-th enterprise, represents the micro recall rate of the i-th enterprise, l hi represents the number of companies that are predicted to be in category h and are actually in category h, m hi represents the number of companies i that are predicted to be in category h, n hiIndicates the number of items that actually belong to class h.

[0091] Optionally, in addition to the classification method using a model, the present application also proposes a rule-based classification method, which can analyze multiple groups of target word segmentation sets according to preset enterprise classification rules to obtain a second classification result for each target enterprise.

[0092] Specifically, for each type of target segmentation set corresponding to each target enterprise, the segmentations in the target segmentation set of this type are sorted according to the frequency of occurrence, and the segmentations of a first preset proportion ranked in front form a segmentation set to be tested; a standard segmentation set corresponding to the target segmentation set of this type is determined, wherein the standard segmentation set includes preset standard segmentations corresponding to industrial nodes in the target industrial chain; for each segmentation to be tested in the segmentation set to be tested, a standard segmentation matching the segmentation to be tested is determined from the standard segmentation set according to regular matching rules and / or similarity matching rules, and the standard segmentation is used as the sub-classification result corresponding to the segmentation to be tested; and multiple sub-classification results corresponding to multiple segmentations to be tested corresponding to the target enterprise are unified as the second classification result of the target enterprise.

[0093] Among them, when determining the standard participle that matches the participle to be detected based on the regular matching rule, for each standard participle in the standard participle set, the subset of the standard participle and the participle to be detected can be determined, and the ratio of the length of the subset to the length of the participle to be detected is calculated; if the largest ratio is greater than the preset ratio threshold, it is determined that the standard participle corresponding to the largest ratio matches the participle to be detected. The specific ratio calculation formula is as follows:

[0094]

[0095] In the formula, a represents the standard participle, b represents the participle to be tested, and a∈b, that is, the standard participle is a subset of the participle to be tested, count(a∩b) represents the word length of the standard participle and the participle to be tested, and count(b) represents the word length of the participle to be tested.

[0096] It should be noted that there are two rules in regular matching: identical and hit, both of which are applicable to the above calculation formula. When they are identical, the calculated ratio is 1, which must be the maximum ratio greater than the preset ratio threshold.

[0097] Optionally, when determining a standard participle that matches the participle to be detected based on a similarity matching rule, for each standard participle in the standard participle set, the similarity between the standard participle and the participle to be detected can be determined, and the type of similarity includes at least one of the following: Jaccard similarity, Dice similarity, overlap, cosine similarity, correlation coefficient, and Hamming distance; if the maximum similarity is greater than a preset similarity threshold, it is determined that the standard participle corresponding to the maximum similarity matches the participle to be detected.

[0098] In some scenarios, a to-be-detected word may match multiple standard word segments that meet the conditions at the same time. In this case, the standard word that comes later in the to-be-detected word segmentation is retained as the only matching sub-classification result. For example, when the to-be-detected word segmentation is car chassis, the standard word segmentation that is filtered out by calculation includes car and chassis, and the final sub-classification result is determined to be chassis.

[0099] After obtaining the first classification result based on model classification and the second classification result based on rule classification, the two classification results can be combined to improve the coverage of enterprise classification.

[0100] Optionally, since both the first classification result and the second classification result include multiple sub-classification results, when determining the final target classification result, the sub-classification results that exist in both the first classification result and the second classification result can be first determined as the target sub-classification result; for the sub-classification results that exist only in the first classification result, the sub-classification results can be sorted according to the probability values ​​corresponding to the sub-classification results predicted by the enterprise classification model, and the sub-classification results with the second preset ratio ranked first can be determined as the target sub-classification results, that is, only the top-n sub-classification results are retained; for the sub-classification results that exist only in the second classification result, the sub-classification results are sorted according to the frequency of occurrence of each sub-classification result, and the sub-classification results with the third preset ratio ranked first can be determined as the target sub-classification results, that is, only the top-k sub-classification results are retained; all target sub-classification results are unified as the target classification result. The values ​​of the second preset ratio n and the third preset ratio k can be adjusted according to the type of industrial chain, and are not specifically limited here.

[0101] Figure 4 A schematic diagram of a complete process of enterprise division is shown, which corresponds to the implementation process described above and will not be described in detail here.

[0102] In the embodiment of the present application, in terms of data usage and processing, different from the related technologies that model a single type of text, the embodiment of the present application integrates multi-dimensional text data to construct a heterogeneous graph, effectively breaking the data island; based on the knowledge graph, the data of each dimension of the enterprise is effectively associated to form a network structure, which can enhance the stability and accuracy of subsequent classification. In view of the fact that the business scope has short text and no context characteristics, the business scope collection node is constructed by word segmentation and clustering, which maximizes the aggregation of business scope word segments and reduces fragmentation. At the same time, multiple connections between enterprises are established to reduce the complexity of subsequent classification; according to the context characteristics of patent abstracts and patents usually involving new technologies, word segmentation is performed by constructing a word segmentation dictionary, and then the patent abstract collection node is constructed according to the patent IPC number; different data preprocessing mechanisms and means are used according to the different characteristics of the text, which helps to improve the accuracy of subsequent classification. On the basis of heterogeneous graphs, each node is encoded, and the GNN model is used to divide the industry category to which the enterprise belongs. Since the GNN model can deeply explore the associations between text data of various dimensions, the associations between enterprises, and the associations between text data of different dimensions and standard classifications, it can realize the learning of the global graph structure with high flexibility; while building the model, the enterprise is divided based on rules and the results are merged. Since the rule-based method is simple and easy, the matching logic can be personalized, especially the business or empirical processing can be customized, with high accuracy; the combination of model and rule-based method effectively integrates the advantages of the two methods, which can increase the coverage of classification and improve accuracy. This application scheme effectively solves the technical problem that it is difficult for relevant classification schemes to accurately divide enterprises into corresponding industrial chain nodes based on relevant data.

[0103] Example 2

[0104] According to an embodiment of the present application, there is also provided an enterprise division device for implementing the enterprise division method in embodiment 1, such as Figure 5 As shown, the enterprise classification device at least includes: an acquisition module 51, a word segmentation module 52, a first classification module 53, a second classification module 54 and a determination module 55, wherein:

[0105] An acquisition module 51 is used to acquire multiple groups of enterprise text data corresponding to multiple target enterprises associated with a target industrial chain, wherein each group of enterprise text data includes text data of different dimensions for describing the industry involved in the enterprise;

[0106] The word segmentation module 52 is used to perform word segmentation processing on each group of enterprise text data to obtain multiple groups of target word segmentation sets;

[0107] The first classification module 53 is used to construct a target industry map based on multiple sets of target word segments, and analyze the target industry map using a pre-trained enterprise classification model to obtain a first classification result for each target enterprise;

[0108] The second classification module 54 is used to analyze the multiple target word segmentation sets according to the preset enterprise classification rules to obtain the second classification results of each target enterprise, wherein the type of enterprise classification rule includes at least one of the following: regular matching rule and similarity matching rule;

[0109] The determination module 55 is used to determine the target classification result of each target enterprise according to the first classification result and the second classification result, wherein the target classification result is used to indicate the industrial node corresponding to the target enterprise in the target industrial chain.

[0110] The functions of each module of the enterprise division device are explained below in conjunction with the specific implementation process.

[0111] Optionally, when acquiring multiple groups of enterprise text data corresponding to multiple target enterprises associated with the target industrial chain, the acquisition module can determine the industry classification identifier associated with the target industrial chain in the preset economic industry classification standard; acquire multiple groups of enterprise text data corresponding to multiple target enterprises that match the industry classification identifier, wherein each target enterprise corresponds to a group of enterprise text data, and the text dimensions in each group of enterprise text data include at least one of the following: business scope, software copyright title, patent title, and patent abstract.

[0112] Optionally, when the word segmentation module performs word segmentation processing on each group of enterprise text data, the specific method is as follows: for the business scope text in each group of enterprise text data, the punctuation marks in the business scope text are adjusted according to the preset punctuation mark usage standard to obtain a first text; the first text is word segmented based on other punctuation marks except the first type of punctuation marks to obtain a first word segmentation set, and the first type of punctuation marks include: colon and semicolon; determine all the word segments in the first word segmentation set that match the preset first keyword to form a second word segmentation set, and the first keyword is a keyword associated with the target industrial chain; Based on the first type of punctuation marks, the words in the second word segmentation set are segmented to obtain a third word segmentation set; the verbs in each word in the third word segmentation set are deleted to obtain a first type of target word segmentation set corresponding to the enterprise text data; and / or, for each set of software title text in the enterprise text data, the enterprise name, region name and second keyword in the software title text are deleted to obtain a second type of target word segmentation set corresponding to the enterprise text data, wherein the second keyword includes: a, based on, of; and / or, for each set of patent title text in the enterprise text data, the design patent title is deleted. other fields except product name; for utility model patent titles whose title length is not greater than the first preset length, delete the third keyword in the utility model patent title, the third keyword includes: of, a kind of, applicable to, used for, with, with; for utility model patent titles whose title length is greater than the first preset length, use the TextRank algorithm to extract keywords in the utility model patent title; for invention patent titles whose title length is not greater than the first preset length and includes the fourth keyword, delete the third keyword in the invention patent title, the fourth keyword includes: method, comma; for invention patent titles whose title length is greater than the first preset length or includes the fourth keyword, use the TextRank algorithm to extract keywords in the invention patent title; use the processing result as the third type of target word segmentation set corresponding to the enterprise text data; and / or, for the patent abstract text in each group of enterprise text data, delete the design patent abstract and the utility model patent abstract; use the preset word segmentation dictionary to perform word segmentation processing on the invention patent abstract text to obtain a fourth word segmentation set; delete the stop words in each word segmentation in the fourth word segmentation set to obtain a fourth type of target word segmentation set corresponding to the enterprise text data.

[0113] The word segmentation dictionary can be established in the following manner: obtaining a plurality of historical invention patent abstract texts, deleting stop words in the historical invention patent abstract texts, and performing word segmentation processing on the historical invention patent abstract texts with the stop words deleted according to punctuation marks to obtain a fifth word segmentation set; determining that the word segmentations in the fifth word segmentation set whose length is less than a first preset length form a sixth word segmentation set, and extracting the word segmentations in the sixth word segmentation set based on a sliding window of a second preset length to construct a first word graph; deleting the word segmentations of non-target parts of speech in the first word graph to obtain a second word graph, wherein the target parts of speech include: nouns, verbs, gerunds, adjectives, English words, and subject-related vocabulary; for each pair of adjacent word segmentations in the second word graph, calculating the degree of aggregation between the two adjacent word segmentations, if the degree of aggregation is greater than a preset degree of aggregation threshold, the two adjacent word segmentations are combined into a first phrase; obtaining a plurality of second phrases from a preset source node, wherein the type of the source node includes at least one of the following: books, official websites, and term libraries; and constructing a word segmentation dictionary using the plurality of first phrases and second phrases.

[0114] Optionally, when calculating the degree of aggregation between two adjacent participles, the following formula may be used:

[0115]

[0116] In the formula, a and b represent two adjacent participles, |a| represents the length of participle a, |b| represents the length of participle b, and cohesion a→b represents the degree of aggregation between participle a and participle b, count(a→b) indicates the frequency of word a pointing to word b, and count(all) indicates the total number of words in the second word graph. represents the degree of aggregation of word a, n represents the number of words pointed to by word a, count(a→w i ) indicates that participle a points to participle w i The frequency of represents the indexing aggregation of word b, m represents the number of words pointing to word b, count(w i →b) represents the participle w i The frequency of pointing to participle b.

[0117] Optionally, when the first classification module constructs a target industry map based on multiple groups of target segmentation sets, it can use a density-based clustering algorithm to cluster all the segments in multiple groups of first-category target segmentation sets corresponding to multiple groups of enterprise text data, and use each obtained segmentation set as a business scope node; use software copyright classification numbers to cluster all the segments in multiple groups of second-category target segmentation sets corresponding to multiple groups of enterprise text data, and use each obtained segmentation set as a software title node; use international patent classification numbers to cluster all the segments in multiple groups of third-category target segmentation sets corresponding to multiple groups of enterprise text data, and use each obtained segmentation set as a patent title node; use international patent classification numbers to cluster all the segments in multiple groups of fourth-category target segmentation sets corresponding to multiple groups of enterprise text data, and use each obtained segmentation set as a patent abstract node; use each target enterprise as an enterprise node, and construct a target industry map based on the association relationship between each enterprise node and each business scope node, software title node, patent title node, and patent abstract node.

[0118] Optionally, when the first classification module uses the pre-trained enterprise classification model to analyze the target industry map, it can be performed in the following manner: use the GloVe algorithm to encode each business scope node, software title node, and patent title node, and use the bidirectional long short-term memory algorithm to encode each patent abstract node to obtain a vectorized representation of the target industry map; use the graph neural network module in the enterprise classification model to analyze the vectorized representation of the target industry map to obtain the enhanced word vector corresponding to each target enterprise; use the classification module in the enterprise classification model to analyze each enhanced word vector to obtain the first classification result corresponding to each target enterprise, wherein the classification module performs classification based on a probability conversion function, and the type of the probability conversion function includes at least one of the following: sigmoid, softmax, logistic.

[0119] The training process of the enterprise classification model includes: constructing an initial model, wherein the initial model includes a graph neural network module and a classification module; obtaining multiple historical industry maps as training samples, and determining the classification results corresponding to the enterprises in each historical industry map as sample labels; using the training samples to iteratively train the initial model, wherein, when each training batch is completed, the F1 value of each enterprise is calculated based on the classification results and sample labels of each enterprise output by the model, and the average F1 value of all enterprises is calculated, and the model parameters are adjusted based on the average F1 value.

[0120] Optionally, when the second classification module analyzes multiple groups of target word segmentation sets according to preset enterprise classification rules, for each type of target word segmentation set corresponding to each target enterprise, the word segmentations in this type of target word segmentation set can be sorted according to the frequency of occurrence, and the word segmentations with a first preset proportion of the sorting order can be formed into a word segmentation set to be tested; a standard word segmentation set corresponding to this type of target word segmentation set is determined, wherein the standard word segmentation set includes preset standard word segmentations corresponding to industrial nodes in the target industrial chain; for each word to be tested in the word segmentation set to be tested, a standard word that matches the word to be tested is determined from the standard word segmentation set according to regular matching rules and / or similarity matching rules, and the standard word is used as the sub-classification result corresponding to the word to be tested; and multiple sub-classification results corresponding to multiple word to be tested corresponding to the target enterprise are unified as the second classification result of the target enterprise.

[0121] Among them, when the second classification module determines the standard participle that matches the participle to be detected based on the regular matching rule, for each standard participle in the standard participle set, it determines the subset of the standard participle and the participle to be detected, and calculates the ratio of the length of the subset to the length of the participle to be detected; if the largest ratio is greater than the preset ratio threshold, it is determined that the standard participle corresponding to the largest ratio matches the participle to be detected.

[0122] Optionally, when the second classification module determines the standard participle that matches the participle to be detected based on the similarity matching rule, for each standard participle in the standard participle set, the similarity between the standard participle and the participle to be detected is determined, and the type of similarity includes at least one of the following: Jaccard similarity, Dice similarity, overlap, cosine similarity, correlation coefficient, Hamming distance; if the maximum similarity is greater than a preset similarity threshold, it is determined that the standard participle corresponding to the maximum similarity matches the participle to be detected.

[0123] Since the first classification result and the second classification result both include multiple sub-classification results, the determination module can determine the target classification results of each target enterprise in the following manner: determine the sub-classification results that exist in both the first classification result and the second classification result as the target sub-classification results; for the sub-classification results that only exist in the first classification result, sort the sub-classification results according to the probability values ​​corresponding to the sub-classification results predicted by the enterprise classification model, and determine the sub-classification results with a second preset proportion ranked first as the target sub-classification results; for the sub-classification results that only exist in the second classification result, sort the sub-classification results according to the frequency of occurrence of each sub-classification result, and determine the sub-classification results with a third preset proportion ranked first as the target sub-classification results; and unify all target sub-classification results as the target classification result.

[0124] It should be noted that each module in the enterprise division device in the embodiment of the present application corresponds one by one to each implementation step of the enterprise division method in Example 1. Since a detailed description has been given in Example 1, some details not reflected in this embodiment can be referred to Example 1 and will not be repeated here.

[0125] Example 3

[0126] According to an embodiment of the present application, a computer program product is also provided. The computer program product includes a computer program, wherein when the computer program is executed by a processor, the enterprise division method in Embodiment 1 is implemented.

[0127] According to an embodiment of the present application, a non-volatile storage medium is also provided, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the enterprise division method in Example 1 by running the computer program.

[0128] According to an embodiment of the present application, a processor is also provided, which is used to run a computer program, wherein the enterprise division method in Example 1 is executed when the computer program is running.

[0129] According to an embodiment of the present application, an electronic device is also provided, which includes: a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to execute the enterprise division method in Example 1 through the computer program.

[0130] Specifically, the computer program executes the following steps when it is running: obtaining multiple groups of enterprise text data corresponding to multiple target enterprises associated with the target industrial chain, wherein each group of enterprise text data includes text data of different dimensions for describing the industries involved in the enterprise; performing word segmentation processing on each group of enterprise text data to obtain multiple groups of target word segmentation sets; constructing a target industry map based on the multiple groups of target word segmentation sets, and analyzing the target industry map using a pre-trained enterprise classification model to obtain a first classification result for each target enterprise; analyzing the multiple groups of target word segmentation sets according to preset enterprise classification rules to obtain a second classification result for each target enterprise, wherein the type of enterprise classification rule includes at least one of the following: regular matching rule, similarity matching rule; determining the target classification result of each target enterprise based on the first classification result and the second classification result, wherein the target classification result is used to indicate the industry node corresponding to the target enterprise in the target industrial chain.

[0131] As an optional implementation, the electronic device may be in the form of a mobile terminal, a computer terminal or a similar computing device. Figure 6 FIG. 1 shows a hardware structure block diagram of an electronic device for implementing an enterprise division method. Figure 6As shown, the electronic device 60 may include one or more (602a, 602b, ..., 602n are used to illustrate) processors 602 (the processor 602 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 604 for storing data, and a transmission device 606 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Figure 6 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 6 More or fewer components as shown, or with Figure 6 Different configurations are shown.

[0132] It should be noted that the one or more processors 602 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the electronic device 60. As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0133] The memory 604 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the enterprise division method in the embodiment of the present application. The processor 602 executes various functional applications and data processing by running the software programs and modules stored in the memory 604, that is, realizing the vulnerability detection method of the above-mentioned application. The memory 604 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 604 may further include a memory remotely arranged relative to the processor 602, and these remote memories may be connected to the electronic device 60 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0134] The transmission device 606 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the electronic device 60. In one example, the transmission device 606 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 606 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0135] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the electronic device 60 .

[0136] The serial numbers of the above embodiments are only for description and do not represent the advantages or disadvantages of the embodiments.

[0137] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0138] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0139] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0140] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0141] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or all or part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of each embodiment method of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk, etc., which can store program code.

[0142] The above are only preferred implementations of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for dividing an enterprise, characterized in that: include: Acquire multiple groups of enterprise text data corresponding to multiple target enterprises associated with the target industrial chain, wherein each group of enterprise text data includes text data of different dimensions for describing the industry involved in the enterprise; Performing word segmentation processing on each group of the enterprise text data to obtain multiple groups of target word segmentation sets; Constructing a target industry map based on multiple groups of target word segmentation sets, and analyzing the target industry map using a pre-trained enterprise classification model to obtain a first classification result for each target enterprise; Analyze the multiple target word segmentation sets according to the preset enterprise classification rules to obtain the second classification results of each target enterprise, wherein the type of the enterprise classification rule includes at least one of the following: regular matching rule, similarity matching rule; The target classification result of each of the target enterprises is determined according to the first classification result and the second classification result, wherein the target classification result is used to indicate the industrial node corresponding to the target enterprise in the target industrial chain.

2. The method according to claim 1, characterized in that: Acquire multiple sets of enterprise text data corresponding to multiple target enterprises associated with the target industrial chain, including: Determine the industry classification identifier associated with the target industrial chain in the preset economic industry classification standard; Acquire multiple groups of enterprise text data corresponding to multiple target enterprises matching the industry classification identifier, wherein each of the target enterprises corresponds to a group of the enterprise text data, and the text dimensions in each group of the enterprise text data include at least one of the following: business scope, software copyright title, patent title, and patent abstract.

3. The method according to claim 2, characterized in that Perform word segmentation processing on each group of the enterprise text data to obtain multiple groups of target word segmentation sets, including: For each group of business scope text in the enterprise text data, the punctuation marks in the business scope text are adjusted according to a preset punctuation mark usage standard to obtain a first text; the first text is segmented based on other punctuation marks except the first type of punctuation marks to obtain a first segmentation set, where the first type of punctuation marks include: colon and semicolon; all segmentations in the first segmentation set that match a preset first keyword are determined to form a second segmentation set, where the first keyword is a keyword associated with the target industrial chain; the segmentations in the second segmentation set are segmented based on the first type of punctuation marks to obtain a third segmentation set; verbs in each segmentation in the third segmentation set are deleted to obtain a first type of target segmentation set corresponding to the enterprise text data; and / or, For each set of software title text in the enterprise text data, the enterprise name, region name and second keyword in the software title text are deleted to obtain a second type of target word segmentation set corresponding to the enterprise text data, wherein the second keyword includes: a, based on, of; and / or, For each set of patent title text in the enterprise text data, delete other fields in the design patent title except the product name; for utility model patent titles whose title length is not greater than the first preset length, delete the third keyword in the utility model patent title, and the third keyword includes: of, a kind of, applicable to, used for, with, with; for utility model patent titles whose title length is greater than the first preset length, use the TextRank algorithm to extract keywords in the utility model patent title; for invention patent titles whose title length is not greater than the first preset length and include the fourth keyword, delete the third keyword in the invention patent title, and the fourth keyword includes: method, comma; for invention patent titles whose title length is greater than the first preset length or include the fourth keyword, use the TextRank algorithm to extract keywords in the invention patent title; use the processing result as the third category target word segmentation set corresponding to the enterprise text data; and / or, For each set of patent abstract texts in the enterprise text data, the design patent abstracts and utility model patent abstracts are deleted; the invention patent abstract texts are segmented using a preset segmentation dictionary to obtain a fourth segmentation set; and stop words in each segmentation in the fourth segmentation set are deleted to obtain a fourth category of target segmentation set corresponding to the enterprise text data.

4. The method according to claim 3, characterized in that The process of establishing the word segmentation dictionary includes: Acquire multiple historical invention patent abstract texts, delete stop words in the historical invention patent abstract texts, and perform word segmentation processing on the historical invention patent abstract texts from which the stop words are deleted according to punctuation marks to obtain a fifth word segmentation set; Determine that the words in the fifth word segmentation set whose length is less than the first preset length form a sixth word segmentation set, and extract the words in the sixth word segmentation set based on a sliding window of a second preset length to construct a first word graph; Deleting the participles of non-target parts of speech in the first word graph to obtain a second word graph, wherein the target parts of speech include: nouns, verbs, gerunds, adjectives, English words, and subject-related vocabulary; For each pair of adjacent participles in the second word graph, calculating the degree of aggregation between the two adjacent participles, and if the degree of aggregation is greater than a preset degree of aggregation threshold, combining the two adjacent participles into a first phrase; Acquire a plurality of second phrases from a preset source node, wherein the type of the source node includes at least one of the following: a book, an official website, and a term library; The word segmentation dictionary is constructed using a plurality of the first phrases and the second phrases.

5. The method according to claim 4, characterized in that Calculate the degree of aggregation between two adjacent participles, including: The following formula is used to calculate the degree of aggregation between two adjacent participles: In the formula, a and b represent two adjacent participles, |a| represents the length of participle a, |b| represents the length of participle b, and cohesion a→b represents the degree of aggregation between participle a and participle b, represents the probability of word a and word b appearing at the same time, count(a→b) represents the frequency of word a pointing to word b, and count(all) represents the total number of word segments in the second word graph. represents the degree of aggregation of word a, n represents the number of words pointed to by word a, count(a→w i ) indicates that participle a points to participle w i The frequency of represents the indexing aggregation of word b, m represents the number of words pointing to word b, count(w i →b) represents the participle w i The frequency of pointing to participle b.

6. The method according to claim 3, characterized in that Constructing a target industry map based on multiple groups of target word segments includes: Using a density-based clustering algorithm to cluster all the segmented words in the multiple groups of the first target segmented word sets corresponding to the multiple groups of the enterprise text data, and taking each segmented word set obtained as a business scope node; Clustering all the word segments in the plurality of sets of the second target word segmentation sets corresponding to the plurality of sets of the enterprise text data using the software work classification number, and taking each obtained set of word segments as a software work title node; Clustering all the segmented words in the multiple sets of the third target segmented word sets corresponding to the multiple sets of the enterprise text data using international patent classification numbers, and taking each segmented word set obtained as a patent title node; Clustering all the segmented words in the plurality of sets of the fourth target segmented word sets corresponding to the plurality of sets of the enterprise text data using international patent classification numbers, and taking each obtained segmented word set as a patent abstract node; Each of the target enterprises is taken as an enterprise node, and the target industry map is constructed based on the association relationship between each of the enterprise nodes and each of the business scope nodes, the software copyright title nodes, the patent title nodes, and the patent abstract nodes.

7. The method according to claim 6, characterized in that The target industry map is analyzed using a pre-trained enterprise classification model to obtain a first classification result of each target enterprise, including: Using the GloVe algorithm to encode each of the business scope nodes, the software title nodes, and the patent title nodes, and using the bidirectional long short-term memory algorithm to encode each of the patent abstract nodes, to obtain a vectorized representation of the target industry map; Utilizing the graph neural network module in the enterprise classification model to analyze the target industry graph represented by the vectorization, and obtaining enhanced word vectors corresponding to each of the target enterprises; The classification module in the enterprise classification model is used to analyze each of the enhanced word vectors to obtain the first classification result corresponding to each of the target enterprises, wherein the classification module performs classification based on a probability conversion function, and the type of the probability conversion function includes at least one of the following: sigmoid, softmax, logistic.

8. The method according to claim 7, characterized in that The training process of the enterprise classification model includes: Constructing an initial model, wherein the initial model includes the graph neural network module and the classification module; Acquire multiple historical industry maps as training samples, and determine the classification results corresponding to the enterprises in each of the historical industry maps as sample labels; The initial model is iteratively trained using the training samples, wherein, when each training batch is completed, the F1 value of each enterprise is calculated based on the classification results and sample labels of each enterprise output by the model, and the average F1 value of all enterprises is calculated, and the model parameters are adjusted based on the average F1 value.

9. The method according to claim 3, characterized in that: Analyze the multiple target word sets according to the preset enterprise classification rules to obtain the second classification results of each target enterprise, including: For each target word set corresponding to each target enterprise, the words in the target word set are sorted according to the frequency of occurrence, and the words with a first preset proportion in the top order are formed into a word set to be detected; Determine a standard segmentation set corresponding to the target segmentation set, wherein the standard segmentation set includes preset standard segmentations corresponding to the industrial nodes in the target industrial chain; For each participle to be detected in the participle set to be detected, determine a standard participle matching the participle to be detected from the standard participle set according to a regular matching rule and / or a similarity matching rule, and use the standard participle as a sub-classification result corresponding to the participle to be detected; The multiple sub-classification results corresponding to the multiple to-be-detected segmented words corresponding to the target enterprise are unified as the second classification result of the target enterprise.

10. The method according to claim 9, characterized in that Determining a standard segmentation matching the segmentation to be detected from the standard segmentation set according to a regular matching rule and / or a similarity matching rule includes: For each standard participle in the standard participle set, determine a subset of the standard participle and the participle to be detected, and calculate the ratio of the length of the subset to the length of the participle to be detected; if the largest ratio is greater than a preset ratio threshold, determine that the standard participle corresponding to the largest ratio matches the participle to be detected; and / or, For each standard participle in the standard participle set, determine the similarity between the standard participle and the participle to be detected, where the type of similarity includes at least one of the following: Jaccard similarity, Dice similarity, overlap, cosine similarity, correlation coefficient, and Hamming distance; if the maximum similarity is greater than a preset similarity threshold, determine that the standard participle corresponding to the maximum similarity matches the participle to be detected.

11. The method according to claim 1, characterized in that: The first classification result and the second classification result both include a plurality of sub-classification results. The target classification results of each target enterprise are determined according to the first classification result and the second classification result, including: Determine a sub-category result that exists in both the first classification result and the second classification result as a target sub-category result; For sub-classification results that only exist in the first classification results, sort the sub-classification results according to the probability values ​​corresponding to the sub-classification results predicted by the enterprise classification model, and determine the sub-classification results with a second preset proportion in the top order as the target sub-classification results; For the sub-classification results that only exist in the second classification results, sort the sub-classification results according to the occurrence frequency of each sub-classification result, and determine that the sub-classification results with a third preset proportion that are ranked first are the target sub-classification results; All the target sub-classification results are unified as the target classification result.

12. An enterprise division device, characterized in that: include: An acquisition module, used to acquire multiple groups of enterprise text data corresponding to multiple target enterprises associated with a target industrial chain, wherein each group of enterprise text data includes text data of different dimensions for describing the industry involved in the enterprise; A word segmentation module is used to perform word segmentation processing on each group of the enterprise text data to obtain multiple groups of target word segmentation sets; A first classification module is used to construct a target industry map based on multiple groups of target word segmentation sets, and analyze the target industry map using a pre-trained enterprise classification model to obtain a first classification result for each target enterprise; A second classification module is used to analyze the multiple target word sets according to a preset enterprise classification rule to obtain a second classification result of each target enterprise, wherein the type of the enterprise classification rule includes at least one of the following: a regular matching rule and a similarity matching rule; A determination module is used to determine the target classification results of each of the target enterprises based on the first classification results and the second classification results, wherein the target classification results are used to indicate the industrial node corresponding to the target enterprise in the target industrial chain.

13. A computer program product, characterized in that include: A computer program, wherein when the computer program is executed by a processor, the enterprise division method described in any one of claims 1 to 11 is implemented.

14. An electronic device, characterized in that: include: A memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to execute the enterprise division method according to any one of claims 1 to 11 through the computer program.

Citation Information

Patent Citations

  • Method and device for automatically acquiring multi-level classification training data of enterprise

    CN112287075A

  • Enterprise classification method and system based on big data deep learning and electronic equipment

    CN112632980A

  • Industrial category determination method and device, storage medium and electronic equipment

    CN114297347A

  • Industrial chain construction method and system

    CN117633518A

  • Enterprise industry classification method and system based on knowledge graph enhancement

    CN118210925A

Cited By

  • Text information processing method and device and related equipment

    CN120633631A

  • Data cataloguing method, system and equipment based on artificial intelligence and storage medium

    CN120951074A

  • An artificial intelligence-based data cataloging method, system, device and storage medium

    CN120951074B