Enterprise name shape similarity intelligent checking method, system, device and medium

Abstract

The application discloses an enterprise name shape similarity intelligent checking method, system, device and medium, belongs to the technical field of enterprise name registration and compliance checking, and aims to solve the technical problem of how to overcome the defects of low calculation efficiency, poor recognition accuracy and weak rule expansibility in the prior art, and realize high accuracy, low delay and learnable evolution of the word number similarity judgment. The technical scheme is as follows: constructing a shape similar word rule library: the initial mapping dictionary is constructed by using and expanding the'shape similar word' file provided by the user; deploying an AI embedded model: deploying a lightweight deep learning model for mapping any word number into a fixed dimension semantic-word shape joint embedding vector, and performing semantic and word shape joint embedding representation on the word number; similarity calculation: calculating the similarity of the new word number input by the user and the existing word number in the shape similar word rule library through the rule channel and the AI channel respectively; fusing the score; judging whether the approximate word number is formed based on the dynamic threshold value, and outputting the checking result.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise name registration and compliance verification, and specifically to an intelligent verification method, system, device and medium for the similarity of enterprise names in similar shapes. Background Art

[0002] At present, the enterprise name registration system generally adopts a rule-based string matching method in the name verification link. For example, a "similar-shaped characters" file is pre-constructed, and easily confused Chinese characters such as "人" and "入", "未" and "末" are grouped together. By traversing and comparing whether the characters of the newly applied name and the names in the existing library belong to the same similar-shaped group, the similarity score is calculated. The following problems exist in the prior art: ① High data maintenance cost: The similar-shaped character table needs to be manually maintained and cannot cover rare characters, newly created characters, variant characters, etc.

[0003] ② Performance bottleneck: Comparing item by item in the entire library, there is serious lag when the amount of data is large.

[0004] ③ Unable to generalize: Poor recognition ability for newly registered names, combined words, homophonic words.

[0005] ④ Unable to self-learn: Unable to self-optimize from historical review results.

[0006] Therefore, how to overcome the defects of low calculation efficiency, poor recognition accuracy and weak rule extensibility in the prior art, and realize the judgment of name similarity with high accuracy, low latency and learnable evolution is a technical problem to be solved urgently at present. Summary of the Invention

[0007] The technical task of the present invention is to provide an intelligent verification method, system, device and medium for the similarity of enterprise names in similar shapes, so as to solve the problem of how to overcome the defects of low calculation efficiency, poor recognition accuracy and weak rule extensibility in the prior art, and realize the judgment of name similarity with high accuracy, low latency and learnable evolution.

[0008] The technical task of the present invention is realized in the following way. An intelligent verification method for the similarity of enterprise names in similar shapes is as follows: Construct a similar-shaped character rule library: Follow and expand the "similar-shaped characters" file provided by the user, construct an initial mapping dictionary similarMap, support the Unicode full character set, and realize the similarity mapping of glyphs, pinyin and radical structures; Deploy an AI embedding model: Deploy a lightweight deep learning model to map any name to a semantic-glyph joint embedding vector of a fixed dimension, and perform semantic and glyph joint embedding representation on the name; Similarity calculation: Calculate the similarity between the newly input font newWord by the user and the existing font oldWord in the near - shape character rule library through the rule channel and the AI channel respectively; among them, the rule channel obtains the rule similarity S_rule by calling PatternSimilarity(newWord, oldWord); the AI channel uses a lightweight dual - tower network, with the backbone being a trimmed Chinese - BERT (retaining 6 layers of Transformer), the input layer fuses glyph features, that is, extracts Chinese character image features or stroke sequence encodings through CNN, the output layer is connected to a similarity discriminator head, the overall number of parameters < 10M, and the inference latency < 10ms (CPU environment); Fusion score: Fusion the similarity scores of the rule channel and the AI channel to generate the final similarity value; Based on a dynamic threshold, determine whether it constitutes an approximate font and output the verification result.

[0009] Preferably, construct an initial mapping dictionary similarMap as follows: Load an authoritative near - shape character table (such as the "Comparison Table of Commonly Used Chinese Characters' Near - Shape Characters in Modern Chinese"); Supplement high - frequency pairs misrecognized by OCR (such as '日' and '曰'); Establish a reverse index for each character to ensure A ∈ similarMap[B] ⇔ B ∈ similarMap[A]; Support dynamic insertion of new character pairs and automatically maintain symmetry.

[0010] More preferably, the initial mapping dictionary similarMap is stored in JSON format, with the key being the Unicode code point (such as U+4EBA), and the value being a set of near - shape characters (such as [U+5165, U+4EC6]).

[0011] More preferably, the lightweight deep - learning model is a dual - tower neural network model dedicated to the similarity discrimination of enterprise font sizes, with a fine - tuned Chinese pre - trained language model (such as Chinese - BERT - wwm) as the backbone, and the input layer fuses Chinese character glyph features, including radical embedding and stroke sequence encoding; that is, use the pre - trained Chinese language model (such as Chinese - BERT - wwm) to perform vectorized comparison on positive - sample font sizes and negative - sample font sizes, and input the positive - sample font sizes and negative - sample font sizes into the pre - trained Chinese language model, and the pre - trained Chinese language model outputs the similarity score; Among them, the positive sample refers to the comparison of manually labeled near - shape font sizes, including near - shape, near - sound, and near - meaning; the negative sample refers to the comparison of randomly selected non - similar font sizes.

[0012] More preferably, based on a dynamic threshold, determine whether it constitutes an approximate font and output the verification result as follows: A regression model is built based on historical review results: using font size length, industry popularity, regional registration density, and historical conflict rate as features, the optimal threshold θ is predicted through XGBoost regression, and the regression model is incrementally trained every morning at midnight. Threshold tables for each industry-length combination are cached for real-time querying. Efficient retrieval and caching: LSH is used to index AI embedding vectors, and precise comparison is only performed on Top-K (e.g., K=100) candidates; high-frequency query results are cached in Redis to avoid duplicate calculations; an inverted index is built for the similar-looking character rule base, partitioned by the first character and the first letter of the pinyin to accelerate the initial screening; high frequency refers to character sizes that have been requested ≥5 times in the past 24 hours; access frequency is counted through a sliding window counter, and if the access frequency exceeds the threshold, the verification results (including Top-K candidates and final scores) are cached, with a cache validity period of 1 hour; Output results: If the score indicates a similar character size, a conflict warning is returned; the results of manual review are fed back to the training set, and the AI ​​model is trained incrementally and the rule library for similar-looking characters is updated regularly.

[0013] A smart verification system for the shape similarity of enterprise fonts, the system comprising: The rules engine module is used to load and parse files of similar-looking characters and build a bidirectional mapping dictionary; The AI ​​inference module is used to perform font size vectorization and similarity calculation using a lightweight Siamese network structure. The encoder of the lightweight Siamese network structure is fine-tuned based on Chinese-BERT-wwm and integrates glyph features (such as Chinese character stroke vectors and radical embeddings) at the input layer to form a 'semantic + glyph' bimodal embedding representation. In view of the short text characteristics of enterprise font sizes, the attention mechanism of standard BERT is pruned and optimized to reduce the number of parameters by 70% to meet the requirements of real-time response. The fusion scoring module is used to score the similarity between the rule-based channel and the AI ​​channel, generating the final similarity value. The rule-based channel obtains the rule similarity S_rule by calling PatternSimilarity(newWord, oldWord). The AI ​​channel adopts a lightweight dual-tower network with a backbone of a cropped version of Chinese-BERT (retaining 6 Transformer layers). The input layer fuses character shape features, that is, it extracts Chinese character image features or stroke order encoding through CNN. The output layer connects to the similarity discriminator. The overall parameter count is <10M and the inference latency is <10ms (CPU environment). The high-efficiency retrieval module integrates LSH indexing and caching mechanisms to achieve efficient retrieval; The dynamic threshold optimization module is used to automatically adjust the similarity threshold based on historical review data using an online learning mechanism. An incremental learning and rule self-update module, which is used to automatically learn new similar-looking Chinese character pairs from the manual review results and dynamically expand the rule base.

[0014] Preferably, the initial mapping dictionary similarMap is constructed as follows: Load an authoritative list of similar-looking Chinese characters (such as the "Comparison Table of Similar-looking Chinese Characters in Commonly Used Chinese Characters in Modern Chinese"); Supplement the high-frequency pairs misrecognized by OCR (such as '日' and '曰'); Establish a reverse index for each character to ensure that A∈similarMap[B] ⇔ B∈similarMap[A]; Support dynamic insertion of new character pairs and automatically maintain symmetry; Among them, the initial mapping dictionary similarMap is stored in JSON format, where the key is the Unicode code point (such as U+4EBA), and the value is the set of similar-looking Chinese characters (such as [U+5165, U+4EC6]).

[0015] More preferably, the efficient retrieval module uses LSH to index the AI embedding vectors, and only performs exact comparison on the Top-K (such as K=100) candidates; caches the results of high-frequency query character sizes to Redis to avoid repeated calculations; establishes an inverted index for the similar-looking Chinese character rule base, partitions it by the first character and the first letter of the pinyin to accelerate the initial screening; among them, high-frequency means that the character size has been requested ≥5 times in the past 24 hours; counts the access frequency through a sliding window counter, and if it exceeds the threshold, caches its verification results (including Top-K candidates and the final score), and the cache validity period is 1 hour; The dynamic threshold optimization module uses the character size length, industry popularity, regional registration density, and historical conflict rate as features, predicts the optimal threshold θ through XGBoost regression, incrementally trains the regression model every day at midnight, and caches the threshold table for each industry-length combination for real-time query.

[0016] An electronic device, comprising: a memory and at least one processor; Wherein, a computer program is stored on the memory; The at least one processor executes the computer program stored in the memory, so that the at least one processor executes the intelligent verification method for the similarity of enterprise character sizes as described above.

[0017] A computer-readable storage medium stores a computer program, and the computer program can be executed by a processor to implement the intelligent verification method for the similarity of enterprise character sizes as described above.

[0018] The intelligent verification method, system, device and medium for the similarity of enterprise character sizes of the present invention have the following advantages: (i) This invention solves the problems of low computational efficiency, poor recognition accuracy and weak rule extensibility in the prior art, and achieves high accuracy, low latency and learnable evolution of font similarity judgment. It is applicable to commercial service scenarios such as auxiliary verification of enterprise name registration, pre-screening of font compliance, and preliminary comparison of trademark similarity. (ii) This invention improves review efficiency, achieves millisecond response times for databases with millions of entries, and solves the problem of slow loading. (iii) This invention reduces labor costs and decreases manual review by more than 90%; (iv) This invention enhances compliance and avoids legal disputes arising from similar trade names; (v) The present invention forms a "rules + AI" dual-engine verification method, which can be extended to scenarios such as trademarks and domain names, and has the potential to become a platform. Attached Figure Description

[0019] The invention will be further described below with reference to the accompanying drawings.

[0020] Appendix Figure 1 This is a flowchart illustrating the working process of an intelligent verification system for the shape similarity of enterprise names. Detailed Implementation

[0021] The intelligent verification method, system, device, and medium for the shape similarity of enterprise names of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Example

[0022] This embodiment provides an intelligent method for verifying the similarity of enterprise font sizes, as detailed below: S1. Construct a rule base for similar-looking characters: Adopt and expand the "similar-looking characters" file provided by the user, construct an initial mapping dictionary similarMap, support the full Unicode character set, and realize the mapping of character shape, pinyin and radical structure similarity; S2. Deploy AI Embedding Model: Deploy a lightweight deep learning model to map any font size into a fixed-dimensional semantic-graph joint embedding vector, and perform semantic and grapheme joint embedding representation on the font size; S3. Similarity calculation: Calculate the similarity between the newly input font newWord by the user and the existing font oldWord in the similar font rule library through the rule channel and the AI channel respectively. Among them, the rule channel obtains the rule similarity S_rule by calling PatternSimilarity(newWord, oldWord); the AI channel uses a lightweight dual-tower network, with the backbone being a trimmed Chinese-BERT (retaining 6 layers of Transformer). The input layer integrates glyph features, that is, extracting Chinese character image features or stroke sequence encodings through CNN, and the output layer is connected to a similarity discriminator head. The overall number of parameters is <10M, and the inference latency is <10ms (CPU environment). S4. Fusion score: Fusion the similarity scores of the rule channel and the AI channel to generate the final similarity value. S5. Based on the dynamic threshold, determine whether it constitutes an approximate font and output the verification result.

[0023] The construction of the initial mapping dictionary similarMap in step S1 of this embodiment is specifically as follows: S101. Load the authoritative similar font table (such as the "Comparison Table of Commonly Used Chinese Characters with Similar Fonts in Modern Chinese"). S102. Supplement the high-frequency pairs misrecognized by OCR (such as '日' and '曰'). S103. Establish a reverse index for each character to ensure that A∈similarMap[B] ⇔ B∈similarMap[A]. S104. Support dynamic insertion of new character pairs and automatically maintain symmetry.

[0024] The initial mapping dictionary similarMap in step S1 of this embodiment is stored in JSON format, with the key being the Unicode code point (such as U+4EBA), and the value being the set of similar fonts (such as [U+5165, U+4EC6]).

[0025] The lightweight deep learning model in step S2 of this embodiment is a dual-tower neural network model dedicated to the similarity discrimination of enterprise font sizes, with a fine-tuned Chinese pre-trained language model (such as Chinese-BERT-wwm) as the backbone, and integrating Chinese character glyph features including radical embedding and stroke sequence encoding in the input layer. That is, using the pre-trained Chinese language model (such as Chinese-BERT-wwm) to perform vectorized comparison on the positive sample font sizes and negative sample font sizes, and inputting the positive sample font sizes and negative sample font sizes into the pre-trained Chinese language model, and the pre-trained Chinese language model outputs the similarity score. Among them, the positive sample refers to the comparison of manually labeled similar font sizes, including similar in form, pronunciation, and meaning; the negative sample refers to the comparison of randomly selected non-similar font sizes.

[0026] In this embodiment, step S5, which determines whether an approximate font size is constituted based on a dynamic threshold and outputs the verification result, is as follows: S501. Construct a regression model based on historical review results: Using font size length, industry popularity, regional registration density, and historical conflict rate as features, predict the optimal threshold θ through XGBoost regression, and incrementally train the regression model every morning at midnight, and cache the threshold table of each industry-length combination for real-time query. S502, Efficient Retrieval and Caching: LSH is used to index AI embedding vectors, performing precise comparisons only on Top-K (e.g., K=100) candidates; high-frequency query results are cached in Redis to avoid redundant calculations; an inverted index is built for the similar-looking character rule base, partitioned by the first character and the first letter of the pinyin to accelerate initial screening; high frequency refers to characters requested ≥5 times in the past 24 hours; access frequency is counted using a sliding window counter, and if the threshold is exceeded, the verification results (including Top-K candidates and final scores) are cached for 1 hour. S503. Output Results: Based on the score results, if the font size is determined to be similar, a conflict warning is returned; the results of manual review are fed back to the training set, and the AI ​​model is trained incrementally and the rule library for similar-looking characters is updated regularly. Example

[0027] This embodiment provides an intelligent verification system for the shape similarity of enterprise fonts, the system comprising: The rules engine module is used to load and parse files of similar-looking characters and build a bidirectional mapping dictionary; The AI ​​inference module is used to perform font size vectorization and similarity calculation using a lightweight Siamese network structure. The encoder of the lightweight Siamese network structure is fine-tuned based on Chinese-BERT-wwm and integrates glyph features (such as Chinese character stroke vectors and radical embeddings) at the input layer to form a 'semantic + glyph' bimodal embedding representation. In view of the short text characteristics of enterprise font sizes, the attention mechanism of standard BERT is pruned and optimized to reduce the number of parameters by 70% to meet the requirements of real-time response. The fusion scoring module is used to score the similarity between the rule-based channel and the AI ​​channel, generating the final similarity value. The rule-based channel obtains the rule similarity S_rule by calling PatternSimilarity(newWord, oldWord). The AI ​​channel adopts a lightweight dual-tower network with a backbone of a cropped version of Chinese-BERT (retaining 6 Transformer layers). The input layer fuses character shape features, that is, it extracts Chinese character image features or stroke order encoding through CNN. The output layer connects to the similarity discriminator. The overall parameter count is <10M and the inference latency is <10ms (CPU environment). An efficient retrieval module, which is used to integrate the LSH index and the cache mechanism to achieve efficient retrieval; A dynamic threshold optimization module, which is used to automatically adjust the similarity threshold according to historical review data by using an online learning mechanism; An incremental learning and rule self-update module, which is used to automatically learn new similar-shaped character pairs from the manual review results and dynamically expand the rule library.

[0028] In this embodiment, the construction of the initial mapping dictionary similarMap is as follows: ① Load an authoritative list of similar-shaped characters (such as the "Comparison Table of Commonly Used Similar-Shaped Characters in Modern Chinese Characters"); ② Supplement the high-frequency pairs misrecognized by OCR (such as '日' and '曰'); ③ Establish a reverse index for each character to ensure that A ∈ similarMap[B] ⇔ B ∈ similarMap[A]; ④ Support dynamic insertion of new character pairs and automatically maintain symmetry; Among them, the initial mapping dictionary similarMap is stored in JSON format, where the key is the Unicode code point (such as U+4EBA), and the value is a set of similar-shaped characters (such as [U+5165, U+4EC6]).

[0029] The efficient retrieval module in this embodiment uses LSH to index the AI embedding vectors, and only performs exact comparison on the Top-K (such as K=100) candidates; caches the results of high-frequency query character sizes to Redis to avoid repeated calculations; establishes an inverted index for the similar-shaped character rule library, partitions it by the first character and the first letter of the pinyin to accelerate the initial screening; where, high-frequency means that the character size has been requested ≥5 times in the past 24 hours; the access frequency is counted by a sliding window counter, and if it exceeds the threshold, its verification result (including Top-K candidates and the final score) is cached, and the cache validity period is 1 hour.

[0030] The dynamic threshold optimization module in this embodiment uses the character size length, industry popularity, regional registration density, and historical conflict rate as features, predicts the optimal threshold θ through XGBoost regression, incrementally trains the regression model every day at midnight, and caches the threshold table for each industry-length combination for real-time query.

[0031] As shown in the appendix Figure 1 The working process of the system is as follows: (1) Input a new enterprise character size: The user inputs a new enterprise character size as the text to be verified; (2) Use a pre-trained Chinese language model for vectorized comparison or perform nearest neighbor search via FAISS: Use a pre-trained Chinese language model (e.g., Chinese-BERT-wwm) to vectorize the input enterprise font size; or perform nearest neighbor search via FAISS (Facebook AI Similarity Search) to find similar font sizes; (3) Perform hybrid similarity calculation on Top-K candidates: Based on multiple dimensions such as text semantics and character similarity, perform hybrid similarity calculation on Top-K candidate font sizes obtained through vectorized comparison or nearest neighbor retrieval, and calculate the comprehensive similarity between candidate font sizes and input font sizes; (4) Determine if the similarity exceeds the threshold: Set a similarity threshold to determine if the similarity between the input font size and the candidate font size is too high: If the similarity exceeds the threshold, proceed to the "Output Verification Result - Conflict (Business Integration)" node; otherwise, proceed to the "Output Verification Result - Pass (Business Integration)" node. The output verification results are as follows: Conflict: If the similarity exceeds the threshold, the output verification result will be "conflict," meaning that the input company name is too similar to an existing name, potentially involving duplication or infringement issues. This result will be processed by the relevant business systems. Pass: If the similarity does not exceed the threshold, the output verification result is "Pass," meaning the input company name is unique and can be used for registration or other purposes. This result is also integrated into the relevant business systems. (5) Log recording for model optimization: Regardless of the verification result, log information will be recorded for subsequent model optimization and performance improvement. Example

[0032] This embodiment also provides an electronic device, including: a memory and at least one processor; The memory stores computer-executed instructions; The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to execute the intelligent verification method for the shape similarity of enterprise fonts in any embodiment of the present invention.

[0033] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0034] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices. Example

[0035] This embodiment also provides a computer-readable storage medium storing multiple instructions, which are loaded by a processor to cause the processor to execute the intelligent verification method for enterprise font shape similarity in any embodiment of the present invention. Specifically, a system or device equipped with a storage medium may be provided, on which software program code implementing the functions of any of the above embodiments is stored, and the computer (or CPU or MPU) of the system or device may read and execute the program code stored in the storage medium.

[0036] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0037] Storage media embodiments for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RYM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0038] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0039] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An enterprise font shape similarity degree intelligent checking method, characterized in that, The method is as follows: Build a similar character rule library: Adopt and expand the user-provided "similar character" file, build an initial mapping dictionary similarMap, support the full Unicode character set, and realize character shape, pinyin and radical structure similarity mapping; Deploy AI Embedding Model: Deploy a lightweight deep learning model to map any font size to a fixed-dimensional semantic-graph joint embedding vector, and perform semantic and grapheme joint embedding representation on the font size; Similarity Calculation: For the new character size `newWord` input by the user and the existing character size `oldWord` in the similar character rule library, the similarity is calculated through the rule channel and the AI ​​channel, respectively. The rule channel obtains the rule similarity `S_rule` by calling `PatternSimilarity(newWord, oldWord)`. The AI ​​channel adopts a lightweight dual-tower network with a pruned version of Chinese-BERT as the backbone. The input layer integrates character shape features, that is, it extracts Chinese character image features or stroke order encoding through CNN. The output layer connects to a similarity discrimination head. The overall number of parameters is <10M and the inference latency is <10ms. Fusion Score: The similarity scores of the rule channel and the AI ​​channel are merged to generate the final similarity value; The system uses a dynamic threshold to determine whether a font size is approximate and outputs the verification result.

2. The business font shape similarity intelligent verification method according to claim 1, characterized in that, The initial mapping dictionary, similarMap, is constructed as follows: Load the authoritative list of similar-looking characters; Supplementing high-frequency pairs of OCR misidentifications; Create a reverse index for each character to ensure that A∈similarMap[B] ⇔ B∈similarMap[A]; It supports dynamically inserting new character pairs and automatically maintains symmetry.

3. The business font shape near similarity intelligent checking method according to claim 1 or 2, characterized in that, The initial mapping dictionary `similarMap` is stored in JSON format, with keys being Unicode code points and values ​​being a set of similar-looking characters.

4. The business-size shape similarity intelligent checking method according to claim 3, characterized in that, The lightweight deep learning model is a dual-tower neural network model specifically designed for enterprise font size similarity discrimination. It uses a finely tuned pre-trained Chinese language model as the backbone and integrates Chinese character shape features, including radical embedding and stroke sequence encoding, into the input layer. Specifically, the pre-trained Chinese language model is used to perform vectorized comparison of positive and negative sample font sizes, and the positive and negative sample font sizes are input into the pre-trained Chinese language model, which then outputs a similarity score. Among them, positive samples refer to manually annotated comparisons of similar-looking characters, including those that are similar in form, sound, and meaning; negative samples refer to comparisons of randomly selected dissimilar characters.

5. The business-size shape similarity intelligent checking method according to claim 4, wherein, The system uses a dynamic threshold to determine whether an approximate font size exists, and outputs the verification results as follows: A regression model is built based on historical review results: using font size length, industry popularity, regional registration density, and historical conflict rate as features, the optimal threshold θ is predicted through XGBoost regression, and the regression model is incrementally trained every morning at midnight. Threshold tables for each industry-length combination are cached for real-time querying. Efficient retrieval and caching: LSH is used to index AI embedding vectors, and only Top-K candidates are precisely compared; the results of frequently queried character sizes are cached in Redis to avoid duplicate calculations; an inverted index is built for the rule base of similar-looking characters, partitioned by the first character and the first letter of the pinyin to accelerate the initial screening; where, high frequency refers to character sizes that have been requested ≥5 times in the past 24 hours; access frequency is counted by a sliding window counter, and if the access frequency exceeds the threshold, the verification result is cached, and the cache validity period is 1 hour; Output results: If the score indicates a similar character size, a conflict warning is returned; the results of manual review are fed back to the training set, and the AI ​​model is trained incrementally and the rule library for similar-looking characters is updated regularly.

6. A smart verification system for the shape similarity of enterprise names, characterized in that, The system includes: The rules engine module is used to load and parse files of similar-looking characters and build a bidirectional mapping dictionary; The AI ​​inference module is used to perform font size vectorization and similarity calculation using a lightweight Siamese network structure. The encoder of the lightweight Siamese network structure is fine-tuned based on Chinese-BERT-wwm and integrates glyph features at the input layer to form a 'semantic + glyph' dual-modal embedding representation. The fusion scoring module is used to score the similarity between the rule-based channel and the AI ​​channel, generating the final similarity value. The rule-based channel obtains the rule similarity S_rule by calling PatternSimilarity(newWord, oldWord). The AI ​​channel adopts a lightweight dual-tower network with a cropped version of Chinese-BERT as the backbone. The input layer fuses character shape features, that is, it extracts Chinese character image features or stroke order encoding through CNN. The output layer connects to the similarity discriminator. The overall number of parameters is <10M and the inference latency is <10ms. The high-efficiency retrieval module integrates LSH indexing and caching mechanisms to achieve efficient retrieval; The dynamic threshold optimization module is used to automatically adjust the similarity threshold based on historical review data using an online learning mechanism. The incremental learning and rule self-updating module is used to automatically learn new similar character pairs from the results of manual review and dynamically expand the rule base.

7. The intelligent verification system for the shape similarity of enterprise names according to claim 6, characterized in that, The initial mapping dictionary, similarMap, is constructed as follows: Load the authoritative list of similar-looking characters; Supplementing high-frequency pairs of OCR misidentifications; Create a reverse index for each character to ensure that A∈similarMap[B] ⇔ B∈similarMap[A]; Supports dynamic insertion of new character pairs and automatically maintains symmetry; The initial mapping dictionary, similarMap, is stored in JSON format, with keys being Unicode code points and values ​​being a set of similar-looking characters.

8. The intelligent verification system for the shape similarity of enterprise names according to claim 6 or 7, characterized in that, The high-efficiency retrieval module uses LSH to index AI embedding vectors and performs precise comparisons only on Top-K candidates; it caches high-frequency query results in Redis to avoid redundant calculations; it builds an inverted index for the similar-looking character rule base, partitioned by the first character and the first letter of the pinyin, to accelerate the initial screening; here, high frequency refers to character sizes that have been requested ≥5 times in the past 24 hours; it uses a sliding window counter to count access frequency, and if the access frequency exceeds the threshold, it caches the verification result, with a cache validity period of 1 hour; The dynamic threshold optimization module uses font size length, industry popularity, regional registration density, and historical conflict rate as features to predict the optimal threshold θ through XGBoost regression. It also incrementally trains the regression model every morning and caches the threshold tables for each industry-length combination for real-time querying.

9. An electronic device, characterized in that, include: Memory and at least one processor; The memory contains computer programs; The at least one processor executes the computer program stored in the memory, causing the at least one processor to perform the intelligent verification method for the shape similarity of enterprise fonts as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed by a processor to implement the intelligent verification method for the shape similarity of enterprise fonts as described in any one of claims 1 to 5.

Images