Inter-enterprise relationship matching method

Abstract

The invention relates to an inter-enterprise relationship matching method, electronic equipment and a computer readable storage medium. The inter-enterprise relationship matching method comprises the following steps: S1, based on a knowledge base, dividing an input enterprise name to obtain a plurality of segmented words; s2, vectorizing each segmented word to obtain a plurality of segmented word vectors; s3, inputting the plurality of word segmentation vectors into a similarity model; and S4, outputting a matching result, which is determined by the similarity model and is related to the relationship between the enterprises, of the enterprises.

Description

Technical Field

[0001] The present disclosure relates to the field of electronics, and particularly to an enterprise relationship matching method, an electronic device, and a computer-readable storage medium. Background Art

[0002] Currently, the accurate identification of the branch relationships between enterprises mainly relies on the direct matching of enterprise names, which has some significant defects. For example, the names are not standardized (using abbreviations, synonyms, etc.); the registration of branches in the public information is incomplete, resulting in incomplete data; the efficiency of manual verification is low and the coverage is insufficient, and so on. In addition, some existing matching algorithms cannot effectively handle industry term differences or cross-regional associations. Therefore, there is an urgent need to be able to automatically parse the semantics of enterprise names and identify implicit organizational relationships to improve the accuracy and comprehensiveness of the discovery of branches or parent institutions. Summary of the Invention

[0003] In view of the above problems, the present disclosure aims to provide an enterprise relationship matching method, an electronic device, and a computer-readable storage medium.

[0004] The enterprise relationship matching method according to the first aspect of the present disclosure includes: S1: Based on a knowledge base, divide the input name of an enterprise to obtain a plurality of word segments; S2: Vectorize each word segment to obtain a plurality of word segment vectors; S3: Input the plurality of word segment vectors into a similarity model; and S4: Output the matching result of the enterprise regarding the enterprise relationship determined by the similarity model.

[0005] According to the enterprise relationship matching method of some embodiments of the present disclosure, optionally, step S1 uses a natural language processing (NLP) algorithm.

[0006] According to the enterprise relationship matching method of some embodiments of the present disclosure, optionally, in step S1, the name of the enterprise is divided into a main body part and a branch part, the main body part indicates the parent institution of the enterprise, and the branch part indicates the branch institution of the enterprise.

[0007] According to the enterprise relationship matching method of some embodiments of the present disclosure, optionally, step S3 further includes inputting the word segment vectors of the main body part of the name of the enterprise into a main body part similarity model, and step S4 further includes outputting the matching result of the enterprise regarding the enterprise relationship determined by the main body part similarity model.

[0008] According to the enterprise relationship matching method of some embodiments of the present disclosure, optionally, the main body part includes a region word segment, a character word segment, an industry word segment, and an organizational form word segment, and the branch part includes a region word segment, a function word segment, and a branch word.

[0009] The method for matching inter-enterprise relationships according to some embodiments of the present disclosure. Optionally, the vectorization includes: performing word vectorization training on multiple word segmentations of the enterprise name using the Word2Vec algorithm to form multiple word vector sets for the multiple word segmentations.

[0010] The method for matching inter-enterprise relationships according to some embodiments of the present disclosure. Optionally, the similarity model is obtained through the following steps: inputting known enterprise names and their corresponding enterprise branch names as training samples; dividing the training samples to obtain multiple word segmentations; vectorizing the multiple word segmentations to obtain multiple word segmentation vectors; and training through the multiple word segmentation vectors to obtain the similarity model.

[0011] The method for matching inter-enterprise relationships according to some embodiments of the present disclosure. Optionally, the name of an enterprise is divided into a main part and a branch part. Training through the multiple word segmentation vectors to obtain the similarity model further includes: training through the multiple word segmentation vectors of the main part to obtain the main part similarity model.

[0012] The computer-readable storage medium of the second aspect of the present disclosure stores instructions that, when executed by a processor, implement the method for matching inter-enterprise relationships according to any one of the foregoing.

[0013] The electronic device of the third aspect of the present disclosure includes a processor and a memory. The memory stores instructions that, when executed by the processor, implement the method for matching inter-enterprise relationships according to any one of the foregoing. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 FIG. shows a schematic flowchart of an inter-enterprise relationship matching method 100 according to some embodiments.

[0015] Figure 2 FIG. shows a schematic flowchart of a model training process 200 according to some embodiments. DETAILED DESCRIPTION

[0016] Some of the multiple embodiments of the present disclosure are described below, aiming to provide a basic understanding of the present disclosure. It is not intended to identify the key or decisive elements of the present disclosure or to limit the scope to be protected.

[0017] For the sake of brevity and illustrative purposes, the principles of the present disclosure are mainly described with reference to its exemplary embodiments herein. However, those skilled in the art will readily recognize that the same principles can be equivalently applied to all types of inter-enterprise relationship matching methods, electronic devices, computer-readable storage media and can be implemented therein, and any such variations do not depart from the true spirit and scope of this patent application.

[0018] Moreover, in the following description, reference is made to the accompanying drawings, which illustrate specific exemplary embodiments. Electrical, mechanical, logical, and structural changes may be made to these embodiments without departing from the spirit and scope of the present disclosure. In addition, although a feature of the present disclosure is disclosed in conjunction with only one of several implementations / embodiments, this feature may be combined with one or more other features of other implementations / embodiments as may be desired and / or advantageous for any given or identifiable function. Therefore, the following description should not be taken in a limiting sense, and the scope of the present disclosure is defined by the appended claims and their equivalents.

[0019] Terms such as "comprising" and "including" mean that in addition to having the units (modules) and steps directly and explicitly stated in the specification and claims, the technical solutions of the present disclosure do not exclude the situation of having other units (modules) and steps not directly or explicitly stated.

[0020] Figure 1 A flowchart showing the enterprise - to - enterprise relationship matching method 100 according to some embodiments is presented. The method 100 includes the following steps:

[0021] In step S1, based on the knowledge base, the name of the input enterprise is segmented to obtain multiple segmented words. Specifically, the knowledge base may, for example, include a pre - constructed regional library, an industry term library, and an organizational form library to assist in achieving accurate word segmentation. The regional library, for example, includes administrative division names covering three levels of provinces, cities, and counties (such as "Beijing" and "Haidian District"). The industry term library can, for example, accurately identify industry keywords based on the national economic industry classification standard (such as "technology" corresponding to "science and technology promotion service industry", and "manufacturing" corresponding to "mechanical equipment manufacturing industry"). The organizational form library standardizes the expression of enterprise types, such as "limited company", "joint - stock company", and "branch company". The branch word library may also include branch identifier words such as "branch company", "branch factory", and "R & D center".

[0022] In some embodiments, in step S1 of the enterprise - to - enterprise relationship matching method 100, the name of an enterprise is divided into a main part and a branch part. The main part indicates the parent institution of the enterprise, and the branch part indicates the branch institution of the enterprise. For example, the named - entity recognition technology in natural language processing (NLP) algorithms can be used to decompose the enterprise name into a main part and a branch part. In some embodiments, the main part of the enterprise - to - enterprise relationship matching method 100 includes region word segmentation, character - of - the - firm word segmentation, industry word segmentation, and organizational - form word segmentation, and the branch part includes region word segmentation, function word segmentation, and branch - word segmentation. Among them, the main part can represent the core identifier of the enterprise's parent institution, including region word segmentation (such as "Beijing"), character - of - the - firm word segmentation (such as "XX"), industry word segmentation (such as "Technology"), and organizational - form word segmentation (such as "Co., Ltd."); the branch part describes the characteristics of the branch institution, including region word segmentation (such as "Shanghai"), function word segmentation (such as "R & D Center"), and branch - word segmentation (such as "Branch Company"). Through this structured division method, the relationship characteristics between the main body (parent) and the branch institution implicitly contained in the enterprise name can be effectively identified.

[0023] For example, for the input name "Beijing A Technology Information Technology Co., Ltd. Shanghai R & D Branch", the results of word segmentation can be, for example: in the main part, region "Beijing", character - of - the - firm "A Technology", industry "Information Technology", organizational - form "Co., Ltd."; in the branch part, region "Shanghai", function "R & D", branch - word "Branch". This division method can effectively distinguish the core identifiers of the parent and the branch institution, avoiding misjudgment caused by the mixed expression of names, such as misclassifying "Shenzhen" in "A Technology Group Shenzhen Factory" as the main - body region.

[0024] In step S2, each word segment is vectorized to obtain a plurality of word segment vectors. In some embodiments, the vectorization of the enterprise relationship matching method 100 includes: performing word vectorization training on a plurality of word segments of the enterprise name by using the Word2Vec algorithm to form a plurality of word vector sets of the plurality of word segments. By using, for example, the Word2Vec algorithm, word vector training is performed on the full-scale industrial and commercial enterprise name database (such as the enterprise registration database of the national industrial and commercial department), and word vector sets of each word segment can be formed. The Word2Vec maps each word in natural language to a short vector of a unified dimension (such as 300 dimensions), and automatically captures the semantic relationships of words from a large amount of text data through unsupervised learning without relying on manual annotation labels. Specifically, this algorithm uses the full-scale industrial and commercial enterprise name data to train a word vector model, so that words with similar semantics (such as "limited company" and "limited liability") are close in the vector space, while words with irrelevant semantics (such as "technology" and "catering") are far apart. This representation method can not only handle the abbreviations in the enterprise name (such as "group" corresponding to "head office") and synonyms (such as "branch company" and "branch factory"), but also quantify the semantic relevance between words through vector operations (such as the administrative division level relationship between "Beijing" and "Haidian District"). This method solves the problem of identifying branch institutions caused by name variants and non-standard expressions in traditional rule matching through the calculation of semantic similarity of word vectors, and at the same time significantly reduces the dependence on manual annotation and thesaurus maintenance. Moreover, word segmentation, vectorization or enterprise relationship matching can be performed based on a relatively large amount of data (even the full scale) of business entities, and the full-scale industrial and commercial enterprise name database is updated more timely, making the results of word segmentation, vectorization or enterprise relationship matching more comprehensive and accurate.

[0025] In step S3, the plurality of word segment vectors are input into a similarity model. Among them, in some examples, the similarity model is obtained through training, and the training process can refer to, for example, the process 200 described below. In some embodiments, the similarity model may refer to the main part similarity model, that is, step S3 of the enterprise relationship matching method 100 further includes inputting the word segment vectors of the main part of the enterprise name into the main part similarity model, and step S4 further includes outputting the matching result of the enterprise regarding the enterprise relationship determined by the main part similarity model.

[0026] In step S4, output the matching result of the enterprise regarding the enterprise relationship determined by the similarity model. By using the similarity model, the similarity of each word segment between the branch institution and the main institution related to the input enterprise can be evaluated (such as the similarity of each word segment of the main part), and one or more main body-branch institution relationships with the highest similarity can be determined.

[0027] In some examples, based on, for example, the compositional characteristics of enterprise names, after vector feature extraction, the transH knowledge graph completion algorithm can be used to obtain enterprise entity name relationship pairs. First, based on NLP technology, the enterprise name is structurally parsed and split into main components (region, brand name, industry, organizational form) and branch components (region, function, branch word). For example, "Shanghai R & D Branch of A Technology Information Technology Co., Ltd. in Beijing" is parsed into main components such as the region "Beijing", the brand name "A Technology", the industry "information technology", and the organizational form "Co., Ltd.", as well as branch components such as "Shanghai", "R & D", and "branch". Subsequently, the Word2Vec algorithm is used for unsupervised training of the full set of industrial and commercial data to generate 300-dimensional word vectors, so that words with similar semantics (such as "Co., Ltd." and "limited liability company") are close in the vector space, effectively capturing semantic relationships such as abbreviations and synonyms. On this basis, the TransH knowledge graph completion algorithm is introduced to construct an enterprise relationship network, and the implicit branch relationship is inferred through relationship hyperplane projection and vector translation constraints, even if the names are not directly related (such as "A Technology" and "A Cloud (Shanghai) Computing Co., Ltd.").

[0028] Then, the model can be trained in combination with the "publicity system - branch" relationship set (i.e., the national industrial and commercial enterprise registration publicity system), and expert rules can be integrated to parse the main - branch relationship set contained in the enterprise name. Specifically, the classifier can be trained with the verified branch relationship pairs of the national enterprise credit information publicity system, and weighted similarity calculation (for example, 60% for the brand name, 30% for the industry, and 10% for the region) and expert rules (such as partial matching of the brand name and industry relevance) can be adopted to achieve accurate identification combining automated determination and manual review. In this way, through the synergistic effect of semantic analysis, knowledge graph, and rule engine, the accuracy and coverage rate of branch identification are significantly improved, while the manual review cost is greatly reduced.

[0029] Figure 2 FIG. shows a schematic flowchart of a model training process 200 according to some embodiments. The similarity model or the main part similarity model described in this article can be trained through the following steps:

[0030] In step 210, the known enterprise name and its corresponding enterprise branch name are input as training samples. For example, the verified "main - branch" relationship pairs can be obtained from the national enterprise credit information publicity system. For example, a main enterprise "A Technology Group Co., Ltd." and its branch "A Technology Shanghai Branch" form a set of samples.

[0031] In step 220, the training sample is divided to obtain a plurality of word segments. In step 230, the plurality of word segments are vectorized to obtain a plurality of word segmentation vectors. The specific word segmentation and vectorization steps are described above and will not be repeated here.

[0032] In step 240, the similarity model is obtained by training through the multiple word segmentation vectors. Wherein, in step 240, the training process of the similarity model is carried out based on the structured data after word segmentation vectorization. First, the system uses the Word2Vec algorithm to perform unsupervised learning on the full amount of industrial and commercial enterprise names, and maps the various component words in the enterprise name (such as region, font size, industry, organizational form) into word vectors of unified dimensions. For example, synonyms such as "Limited Liability" and "Limited Liability" are close in distance in the vector space, while irrelevant words such as "technology" and "catering" are farther away. This process does not require manual labeling, and relies entirely on massive enterprise name data to automatically capture the semantic association between words, effectively solving the semantic alignment problem of abbreviations (such as "Group" corresponding to "Head Office") and synonyms (such as "Branch" and "Factory") in enterprise names, and providing basic semantic representation for subsequent similarity calculations.

[0033] Secondly, the core data for model training comes from the "branch information" publicly available on the National Enterprise Credit Information Publicity System. These verified "entity-branch" relationship pairs serve as positive examples for input into the system. After structurally decomposing the company name for each example, the model extracts the word vectors for the main components (region, font size, industry, and organizational form) to construct a weighted similarity calculation framework. For example, for the entity "Province A Technology Group Co., Ltd." and its branch "A Technology Shanghai Branch," the model calculates the font size vector similarity of the main components (e.g., the cosine similarity between "A Technology" and "A Technology" is 1.0) and the industry vector correlation (e.g., the cosine similarity between "Technology" and "Information Technology" is 0.85). Furthermore, the model combines regional consistency (e.g., a cross-regional variation coefficient of 0.1) to generate a comprehensive similarity score using preset weights (e.g., 60% for font size, 30% for industry, and 10% for region). Through iterative training with a large number of positive and negative examples (e.g., randomly generated pairs of unrelated company names), the model learns to distinguish valid associations from noisy data, ultimately forming a weighted similarity model for the main components.

[0034] Finally, in some other examples, during the training process, a knowledge graph completion algorithm and expert rules can be further integrated to enhance the robustness of the model. For example, the relationship between enterprise entities and their branches can be mapped to a low-dimensional vector space through the TransH algorithm to complete the branch relationships not explicitly expressed in the name. At the same time, the model output is verified in combination with the expert rule base. For example, it is forced that the font size part matches (e.g., "C Manufacturing" and "C Manufacturing Technology" are regarded as valid, while "C Manufacturing" and "D Technology" are invalid), or an artificial review process is triggered for cross-industry branches (such as an "A Food Company" associated with a "Biotechnology Research Institute"). In this way, high accuracy and high coverage of the model in complex scenarios can be ensured, while reducing the dependence on manual annotation and maintenance.

[0035] In some embodiments, the name of an enterprise is divided into a main part and a branch part. The similarity model obtained by training through the multiple token vectors described above may further include: training through the multiple token vectors of the main part to obtain a main part similarity model. That is to say, the main part similarity model can be obtained by training only on the tokens of the main part. That is, the training process of the similarity model can further focus on the main part of the enterprise name, and by optimizing the feature weights and training strategies, the recognition accuracy of the core components can be improved. Specifically, when the enterprise name is divided into a main part (region, font size, industry, organizational form) and a branch part (branch region, function, branch word), only the token vectors of the main part can be extracted for model training to construct a main part similarity model. For example, for a main enterprise "A Science and Technology Group Co., Ltd. in a certain province" and its branch "A Science and Technology Shanghai Branch", only the token vectors of the main components ("A in a certain province", "Science and Technology", "Group Co., Ltd.") are retained during training, and the token vectors such as "Shanghai" and "Branch" in the branch part are ignored, so as to focus on the core identification features of the parent institution. By focusing on the main components, the interference of redundant branch information is reduced, the matching accuracy of core identifications such as font size and industry is improved, and the robustness of the model in complex naming scenarios such as abbreviations and cross-regional branches is enhanced.

[0036] According to another aspect of the present disclosure, an electronic device is also provided. The electronic device includes a memory and a processor. The memory stores instructions that, when executed by the processor, implement the enterprise relationship matching method according to any one of the foregoing embodiments. According to still another aspect of the present disclosure, a computer-readable storage medium storing instructions is also provided, which executes the enterprise relationship matching method according to any one of the embodiments of the present disclosure when the instructions are run.

[0037] Among them, the computer-readable storage media, memories, storage units, storage modules, etc. referred to in this application include various types of computer-readable storage media and can be any available media accessible by a general or special-purpose computer. For example, the computer-readable medium can include RAM, ROM, EPROM, E 2 PROM, registers, hard disks, removable disks, CD-ROMs or other optical disk memories, magnetic disk memories or other magnetic storage devices, or any other transient or non-transient media that can be used to carry or store desired program code units in the form of instructions or data structures and can be accessed by a general or special-purpose computer or a general or special-purpose processor. The above combinations should also be included within the scope of protection of computer-readable media. An exemplary storage medium is coupled to the processor so that the processor can read from / write to the storage medium. In an alternative, the storage medium can be integrated into the processor. The processor and the storage medium can reside in an ASIC. The ASIC can reside in a computing terminal. In an alternative, the processor and the storage medium can reside in the computing terminal as discrete components.

[0038] The above mainly describes the enterprise relationship matching method, electronic device, and computer-readable storage medium of the present disclosure. Although only some specific embodiments of the present disclosure have been described, those of ordinary skill in the art should understand that the present disclosure can be implemented in many other forms without departing from its gist and scope. Therefore, the examples and embodiments shown are considered illustrative rather than restrictive, and the present disclosure may cover various modifications and substitutions without departing from the spirit and scope of the present disclosure as defined by the appended claims.

Claims

1. An enterprise - to - enterprise relationship matching method, characterized in that, The method includes the following steps: S1: Based on the knowledge base, divide the name of the input enterprise to obtain multiple word segments; S2: Vectorize each of the word segments to obtain multiple word segment vectors; S3: Input the multiple word segment vectors into a similarity model; and S4: Output the matching result of the enterprise regarding the inter-enterprise relationship determined by the similarity model.

2. The enterprise relationship matching method according to claim 1, wherein In step S1, a natural language processing (NLP) algorithm is used.

3. The enterprise relationship matching method according to claim 1, characterized in that, In step S1, the name of the enterprise is divided into a main part and a branch part. The main part indicates the parent institution of the enterprise, and the branch part indicates the branch institution of the enterprise.

4. The enterprise relationship matching method according to claim 3, wherein Step S3 further includes inputting the word segment vectors of the main part of the name of the enterprise into a main part similarity model. Step S4 further includes outputting the matching result of the enterprise regarding the inter-enterprise relationship determined by the main part similarity model.

5. The enterprise relationship matching method according to claim 3, characterized in that, The main part includes region word segments, character word segments, industry word segments, and organizational form word segments. The branch part includes region word segments, function word segments, and branch words.

6. The enterprise relationship matching method according to claim 1, wherein The vectorization includes: using the Word2Vec algorithm to perform word vectorization training on multiple word segments of the enterprise name to form multiple word vector sets of the multiple word segments.

7. The method for matching enterprise - to - enterprise relationships according to claim 1, wherein The similarity model is obtained through the following steps: Input known enterprise names and their corresponding enterprise branch names as training samples; Divide the training samples to obtain multiple word segments; Vectorize the multiple word segments to obtain multiple word segment vectors; Train through the multiple word segment vectors to obtain the similarity model.

8. The enterprise relationship matching method according to claim 7, wherein The name of the enterprise is divided into a main part and a branch part. Training through the multiple word segment vectors to obtain the similarity model further includes: training through the multiple word segment vectors of the main part to obtain a main part similarity model.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a processor, implement the method according to any one of claims 1-8.

10. An electronic device, characterized in that, The electronic device includes a processor and a memory. The memory stores instructions that, when executed by the processor, implement the method according to any one of claims 1-8.

Images