TWI851259B - A system of semantic analysis-based trademark class recommendation and the method thereof - Google Patents

Abstract

The present invention provides an online application system with a classification recommend module. The system includes an electronic device operated by a user, which connects to a server through internet. The electronic device comprises a processor, a memory, and a network interface controller. The server includes an application program. The user operates the electronic device to execute the application program's trademark online application module, conducting a trademark application online. The user inputs the required data for the trademark application through the electronic device. The system utilizes the classification recommend module and the risk assessment module to generate classification recommendation reports and risk assessment reports, providing solutions to the user's difficulties in selecting classifications.

Description

A semantic analysis trademark category recommendation system and method

The present invention relates to an online trademark application and management system, particularly a system with trademark category recommendation and risk assessment.

In the traditional trademark application process, whether domestic or foreign, paper documents need to be printed out and a large number of forms need to be filled out, which is not in line with environmental protection principles. Due to the formality of trademark application documents, if the applicant fills out incorrectly, the entire process may need to be repeated, which leads to a waste of time and human resources. Especially when applying for trademarks in other countries, in addition to traditional paperwork, if communication is carried out by personnel, due to their different professional backgrounds, different language usage, cultural differences and other unpredictable factors, inaccurate or misunderstood information may be conveyed, which leads to cognitive differences between applicants and foreign agents and government agencies. Therefore, applicants may not get the results they originally wanted.

On the other hand, in the past, the target users of trademark search were professional practitioners. For those without relevant professional background, it was very difficult to understand the search logic. Users could not intuitively understand the risk of the trademark they wanted to apply for being approved, and how to evaluate whether they could safely proceed with the trademark application or whether they needed to make design adjustments to the application text or images. For trademark searches, official search pages established by government agencies of various countries or private trademark search platforms often only list similar previous trademark cases during the trademark search process, and do not provide any result analysis, reference points or related evaluations. This makes it difficult for users who know nothing about intellectual property to judge the impact of these previous cases on their own trademark application cases, and users may also find it difficult to estimate the risk level.

In addition, in order to cope with the different definitions and interpretations of goods and services in accordance with the culture and customs of each country, the current way to select trademark classification is to search for similar results by precise or vague words based on the classification standards of each country and the internationally used Nice classification. However, for users who do not understand the classification standards published by official agencies of various countries or the international Nice classification description method, it will be very difficult to find the correct and corresponding results. Users may miss some important items that they really need. For example, the owner of a clothing store may not know that what he really needs is "wholesale and retail of clothing", but just find some items in the classification by himself, such as "women's clothing, men's clothing, shirts". Therefore, under this situation, users can only obtain suggestions on trademark classification and product items for their trademark applications through traditional manual consultation, and cannot directly apply for trademarks based on their own circumstances.

As a result, it is not easy for businesses to select trademark classifications and file trademark applications in various countries. Ultimately, users (businesses) can only seek help from local law firms to seek professional advice and entrust the execution of applications on the selection of appropriate trademark classifications and product items in a single country. Generally speaking, it takes one to several weeks from communication to receiving the final classification advice. If there are multiple countries or the advice is wrong, the multiple round trips often lead to missed opportunities. The traditional process not only makes the trademark application process more time-consuming, but also makes the application more time-consuming. The complexity of the process increases, and the time spent on consulting by a large number of trademark professionals also increases, which in turn leads to high costs for trademark applications. The pre-application processing and execution cycle are also longer, which in turn affects the general public. When they have new brand names, slogans or business (product) names, they are prone to give up trademark applications due to the difficulty and cost of trademark applications. This lays the groundwork for possible trademark disputes and irreversible risks.

After the epidemic, in the era of e-commerce and AI development revolution, cross-border e-commerce transactions and cross-border services are booming. More merchants need cross-border brands. When trademark applicants need to deploy trademarks across countries, it is usually because the contents of trademark product items in different countries are different, and it is impossible to directly convert the trademark application content of the original application country into the trademark application content of another country. At this time, it is necessary to entrust professionals again to enter the target country based on the product items of the original case. Manual comparison and selection are carried out in the national commodity item database; this requires a lot of man-hours and back-and-forth confirmation, which not only requires a higher execution cycle and cost, but also, through manual methods, trademark professionals are often not experts in the industry, and occasionally there will be misunderstandings or errors in recommended operations, which will lead to an increase in the risk of subsequent corrections and a slowdown in the speed of trademark applications, which leads to an increase in the execution cost when applying for trademarks abroad. On the other hand, trademark practitioners also suffer from reduced job satisfaction and increased burnout due to the repetitive execution of trademark recommendation work. Normally, a cross-border trademark recommendation takes several hours to several days to process from communication, understanding, judgment and preparation of recommendation opinions. Often, after the recommendation opinions are completed, they are provided to the trademark applicant, but the applicant does not apply in the end. The mental effort consumed in this process is only felt by trademark practitioners.

At present, some manufacturers or official systems have developed systems and functions for trademark category recommendation and trademark search, but they still have different shortcomings in practical use, which makes it difficult for users to obtain trademark application suggestions and evaluation content in a very intuitive user experience, so as to facilitate the subsequent (multi-national) trademark application process. The following lists and explains the unresolved problems one by one.

First of all, for creators using the LegalZoom platform (https://www.legalzoom.com/business/intellectual-property/trademark-registration-overview.html) provided by LegalZoom.com, Inc. in the United States, although the trademark application service process of LegalZoom has been implemented online, more specifically, it is a simple online order-taking process, that is, when issuing a trademark application demand, it is necessary to complete the trademark application form provided by the platform, as shown in Figures 1 to 3, including providing The trademark name to be applied for, the scenario of trademark name use, and the user's industry description; after completing the above information and payment, the system will transmit the information to the backend and the specialists assigned by LegalZoom will perform the subsequent manual offline and consulting services to complete the subsequent trademark application services and processes; in the above service process, the LegalZoom platform only realizes and replaces the traditional trademark affairs through online ordering, which requires face-to-face or offline contact and receiving customer application requirements by business personnel. However, in the above process, users are still unable to obtain trademark category proposals and execute trademark application processes by themselves, and still need to go through a lengthy interview process.

As disclosed in Chinese Patent Publication No. CN109800340A, please refer to Figures 4 and 5. The user needs to first enter the correct text or word in the trademark name field, and the user ultimately recommends at least one product/service item corresponding to each product category in at least one product category based on the product label information, including: the user ultimately determines the first similarity between the trademark and trademark information of each trademark category in at least one product label category in the database based on the product label information; and the user ultimately recommends at least one product/service item corresponding to each product category in the smallest trademark category based on the first similarity. Determine the second similarity between the trademark and trademark information of the similar group of each group in at least one group, wherein at least one commodity/service item corresponding to each commodity category in at least one commodity category is recommended according to the first similarity, including: recommending at least one commodity/service item corresponding to each commodity category in at least one commodity category according to the first similarity and the second similarity. For example, if the trademark category determined by the user is Class 7, if the first group in Class 7 and the second group in Class 12 are similar groups, then in addition to comparing the trademark information with the commodity label of the commodity label of Class 7 and determining the first similarity, it is also necessary to compare the trademark information with the trademark of the similar group (i.e., the second group) in the trademark category of Class 12 and determine the second similarity. Further, recommend the commodity/service item according to the first similarity and the second similarity. For example, the first group has no identical or similar trademarks, that is, the first similarity, and the registration success rate is high, but the second group has many similar trademarks and identical trademarks, that is, the second similarity is relatively high, and the registration success rate is relatively low. Therefore, considering the first and second similarities comprehensively, the registration success rate of the first group is relatively low, and it is not recommended to register the first group.

However, the above-mentioned Chinese patent (publication number CN109800340A) has several issues to be improved, for example: trademark recommendation comparison is mainly based on label information. If the label setting is not accurate enough or the same product item may meet multiple labels at the same time, the label comparison process may be inaccurate or a large number of non-compliant items may be matched, resulting in reduced recommendation efficiency. In addition, the above-mentioned Chinese patent uses the group relationship of category items in the Chinese trademark classification as one of the bases for judgment during comparison. The group relationship in the trademark category is certainly the basis for official search and retrieval, but when merchants or trademark practitioners need to know the recommended categories, it may also produce non-compliant items that are matched, resulting in too much data and reducing the effectiveness of the recommendation. Finally, when the above-mentioned Chinese patent receives information from users, the text entered by the user must be accurate and precise in the corresponding field, such as the trademark name, so that the label can be compared and the result information of recommending registration or not recommending registration can be generated.

As disclosed in Chinese Patent Publication No. CN109902196B, a trademark category recommendation method is applied to a trademark category recommendation device. Specifically, the method includes: constructing update data in real time; updating the trademark categories in a historical trademark database through the update data, wherein the trademark categories in the historical trademark database are historical recommended trademark categories; and determining the updated trademark category as a recommended trademark category. In this regard, the trademark category recommendation device constructs update data in real time, and, The trademark category recommendation device updates the trademark category in the historical trademark database according to the update data, wherein the trademark category in the historical trademark database is the historical recommended trademark category; then, the trademark category obtained after the update is determined as the recommended trademark category, wherein the recommended trademark category is the trademark category with the lowest registration frequency so far, thereby helping the user to determine the trademark category to be registered, and the user can also continue to select the trademark category with the lowest registration frequency from the recommended trademark category.

There are issues that need to be improved in the Chinese patent with announcement number CN109902196B. Its recommendation principle is based on historical recommendation records and statistical registration frequency records of various trademark categories. The data in the database is ensured to be up-to-date by constantly updating the data. However, for trademark categories with few recommendations or official registration frequencies in the recommendation system, such a recommendation model is obviously not necessarily the most suitable for users, because the least number of applicants means that users will not be interested. Furthermore, even though the Chinese patent with announcement number CN109902196B will also conduct a similarity comparison between the applicant's industry field and the trademark category, and find trademark categories with similar fields and low registration rates, it actually does not provide the applicant with the benefit of actually recommending categories or even product items. Finally, when the Chinese patent with announcement number CN109902196B receives user information data, the applicant must enter accurate and precise text in the corresponding fields, such as trademark name and industry field, so that subsequent industry similarity comparison can be carried out and trademark categories with few recommendations and low registration frequencies can be found.

As disclosed in Chinese Patent Publication No. CN111898022A, a trademark category recommendation method and apparatus, as well as a storage medium and an electronic device, receive an industry requested by a user; determine identification information of the industry requested by the user; based on the identification information of the industry requested by the user, obtain at least one trademark category corresponding to the industry requested by the user; and recommend the at least one trademark category to the user. Optionally, in an embodiment, based on the identification information of the industry requested by the user, obtaining at least one trademark category corresponding to the industry requested by the user includes: based on the identification information of the industry requested by the user, searching for trademark registration category statistics of the industry requested by the user from a database; wherein the trademark registration category statistics of the industry are used to represent the statistical information of the number of each trademark category registered for each industry. Optionally, the trademark registration category statistical information of the industry includes: determining the industry to which each company belongs based on the subject information of each company in the database; obtaining the trademark registration category of each company based on the trademark information of each company in the database; and performing statistics on the trademark registration categories of each industry based on the industry to which each company belongs and the trademark registration categories of each company to obtain the trademark registration category statistical information of each industry. Recommending the at least one trademark category to the user includes: determining at least one trademark registration category of each industry based on the trademark registration category statistics of each industry and the preset category threshold information of each industry; and recommending at least one trademark category of the industry requested by the user to the user based on the at least one trademark registration category of each industry.

As mentioned above, the Chinese patent (publication number CN111898022A) still has problems to be improved. Simply using statistical methods to collect statistics on the company's service (product) items registered with various Chinese public and commercial bureaus and the company's trademark application categories and store them in a database, this model is not applicable to other countries, let alone cross-border category recommendation analysis; when the user selects a specific industry, the corresponding trademark application category statistics for that industry will pop up. The system calculates the results and forms the recommendation results; it only recommends the trademark category, but does not recommend the actual product item content; the actual usability is not high. Even if the user can barely find and select the industry of his company from the industry classification provided by the system, he can only get very general "trademark category" information, but in practice, there is a high probability that he cannot apply for a single category of all product items. Therefore, the practical effect of recommending but assisting customers in trademark applications is limited.

For example, as disclosed in Chinese patent publication number CN107330109A, sample trademark images and contents are processed according to the preset trademark classification standard. The specific processing process includes: (1) establishing a trademark classification standard composed of multiple combinations of preset shape features, pronunciation features and meaning features, (2) determining whether the sample trademark is composed of Chinese characters, graphics, letters, numbers or symbols. (1) identify the content of the constituent elements, (2) identify the smallest unit of shape features, the smallest unit of pronunciation features, and the smallest unit of meaning features of each constituent element of the sample trademark; (3) extract the segmentation information of various texts and graphics generated or converted by each combination scheme according to the established trademark score card standard, use this segmentation information as the sample trademark score card information, and set the similarity evaluation score of each preset trademark score card standard. Using the input trademark subcard information set as the search keyword, the sample trademark subcard information stored in the trademark memory is searched to obtain the subcard information and subcard matching information of the relevant result trademarks; the calculation is performed according to the preset trademark shape similarity rate, trademark meaning similarity rate, trademark pronunciation similarity rate and search keyword matching score rate calculation formula; the trademark proximity comprehensive quantitative value is calculated and obtained, and then the result trademarks are sorted according to the size of the trademark proximity comprehensive quantitative value.

The above-mentioned Chinese patent (publication number CN107330109A) still has problems to be improved. This case is mainly aimed at the technical development of trademark search logic. It analyzes and compares the trademark logos input by users from different angles such as type, pronunciation, and meaning, calculates similarity and sorts them. Even if it helps users sort and cites many previous cases at the same time, users still don’t know whether to apply for it. In addition, the previous cases of trademarks are It is established under the concept of the same category or even the same "similar group". Even if your logo is similar to others, but you are selling food and the other person is selling mobile phones, in this case, the other person's logo should not affect your trademark application. This case only conducts a large number of searches and sorts out the similarities. In fact, the applicant still does not know which category to apply for, and still cannot solve the related issues such as cross-border trademark recommendation.

As disclosed in US Patent Publication No. US20140280104A1, for example, in one example of the solution described herein, a trademark search is performed, and the searched trademark is compared with a number of potentially relevant references, such as existing trademark registrations, common law references, domain names, etc. The search results are compiled into a data set, which can be stored in any appropriate data storage mechanism. The data set is analyzed to determine proximity scores for multiple categories associated with each reference. For example, these categories may include: appearance similarity, sound similarity, connotation similarity, business impression similarity, product/service similarity, transaction channel similarity, sales condition similarity, previous trademark popularity, etc. The associated proximity score is an objective measure of potential trademark risk and conflict. For example, in one example, the proximity score can be represented by a number from 0-5, where higher scores indicate a higher level of concern, in which case a higher proximity score for appearance similarity indicates a higher appearance similarity, and a higher score for a prior trademark indicates that the trademark is more famous. In other words, the higher the proximity score, the more likely that factor is to indicate potential infringement or conflict.

Although the above-mentioned US patent (publication number US20140280104A1) shows the degree of similarity by scoring after comparison, and displays the categories of the similar previous case through the interface, there are still problems to be improved. For example, although users can know which categories to avoid applying in, they do not know which categories are suitable for application. If they are conducting cross-border trademark recommendations, they cannot assist in completing them. In addition, users need to enter precise and accurate text in the corresponding fields, such as trademark names and product fields, in order to effectively conduct subsequent comparisons and scoring. The benefits in actual use are also quite limited.

For example, as disclosed in US Patent Publication No. US20170322983A1, a special UI is used to indicate the similarity between the previous case and the searched trademark. The "display" displays the similarity (similar threat) of the previous case to the current case through the position of the previous case displayed on different displays. Users can observe the content classification of various degrees of "similar previous cases" through different block positions of the display.

The trademark search logic and function of the above-mentioned US patent (publication number US20170322983A1) are limited to the similarity comparison and score conversion of "text trademarks". The difference between this case and this case lies in the starting point. This case uses the trademark application classification recommendation as a starting point to conduct previous case search and risk assessment, and make category recommendations. The purpose of the previous case is to be used for highly professional trademark search, and through the screen presentation and operation of the "display", users can classify and view trademarks with different degrees of similarity.

Another example is the trademark category search system on the official website of the Intellectual Property Office of Singapore (IPOS), see Figures 6 and 7. It allows users to enter the name of a product or service, such as car repair. The system matches the entered text in the database to find all product items containing car or repair, and displays all product items containing car or repair in the search results, and indicates which category they belong to.

The trademark category search system of IPOS mentioned above is just a general simple full text comparison. If one or two words are different, the search cannot be complete. In the process, all the included product items will be listed. For users, it does not achieve the effect of recommendation or even risk assessment. Moreover, the words entered by the user must be clear product words or service words, otherwise the system cannot perform the comparison. For example, if you enter "I run a car repair store", the search result is 0. It is also difficult for general applicants to easily know the trademark category item of the application in terms of input, because general applicants do not necessarily know the name of the trademark product item or which words are in line with the search target.

From the above description, it can be seen that it is necessary to improve or adjust the known technology to enhance its convenience in use. In view of this, the inventor of this invention has made great efforts to research and create, and finally developed and completed the system of this invention.

The purpose of this invention is to propose an online application system with trademark category recommendation, which solves the problems existing in the above-mentioned existing technologies.

Therefore, in order to achieve the above-mentioned purpose of the present invention, the present invention provides a semantic analysis trademark category recommendation system, which includes: an electronic device operated by a user, the electronic device is connected to a server through network information, wherein the electronic device includes a processor, a memory and a network interface controller, and the server includes an application.

The processor connects to the server through a network interface controller and executes the application program, thereby configuring and activating the trademark online application module and the category recommendation module. The processor further configures and activates the application data collection unit, the category item selection unit, the data transmission unit, the text parsing unit, the category item comparison unit, and the report generation unit.

In addition, the processor is also configured to enable a data transceiver module to load data from multiple databases into the memory.

The processor is further connected to multiple submission servers via a network interface controller, allowing the data transmission unit to upload the user's trademark application documents to the submission server in the selected country.

The user operates the processor of the electronic device to execute the trademark online application module of the application to apply for the trademark online. The user enters the data required for the trademark application through the electronic device, such as: trademark name, trademark image, applicant, contact person, product items to be applied for, or description of the product and service, etc. The display screen of the electronic device will display the corresponding fields for the user to fill in the corresponding data. The processor then activates the application data collection unit, and the data entered by the user in the corresponding fields is classified and stored in the memory. Since the trademark name, trademark image, product items to be applied for, and description of the product and service in the trademark application data will be used in particular in the future, the information in these fields is first classified by the application data collection unit.

Preferably, the processor is configured to enable the category item selection unit, and provide the trademark category items pre-stored in the memory for the user to select through the electronic device.

Preferably, the processor is configured to enable the data transmission unit to connect to a filing server in at least one country through the network interface controller to upload the trademark application documents completed by the user.

Preferably, the processor is configured to enable the text parsing unit to perform semantic analysis on the trademark application data input by the user to extract keywords.

Preferably, the processor is configured to enable the category item matching unit, compare the keywords extracted by the text parsing unit with the trademark category items pre-stored in the memory, and store the comparison results back in the memory.

Preferably, the processor is configured to enable the report generation unit to sort the matching results in the memory according to the degree of word similarity to generate the category recommendation report.

Preferably, the processor connects to the server through the network interface controller and executes the application, thereby further configuring and activating a risk assessment module to compare the trademark name in the trademark application data input by the user with previous cases, and the processor executes the risk assessment module in the application to further configure and activate the text parsing unit, a search comparison unit and the report generation unit.

Preferably, the processor is configured to activate the text parsing unit, to perform text arrangement and combination on the trademark name of the trademark application data input by the user and store it in the memory, and the processor is further configured to activate the search and comparison unit, to compare all the text arrangement combinations in the memory with the trademark previous cases pre-stored in the memory, and to generate the comparison results and store them back in the memory.

Preferably, the processor is configured to enable the report generation unit to sort the matching results in the memory according to the degree of word similarity to generate the risk assessment report.

The following is a detailed description using only specific implementation examples and accompanying drawings.

100: Electronic devices

110: Processor

120: Memory

130: Network interface controller

200: Internet

300: Server

310: Applications

311: Trademark online application module

3111: Application for data collection unit

3112:Category item selection unit

3113:Data transmission unit

312: Category recommendation module

3121: Text parsing unit

3122: Category item comparison unit

3123: Report generation unit

3124: Knowledge graph comparison unit

313: Data transceiver module

314: Risk Assessment Module

3142: Search and match unit

400: Database

500: Delivery server

600: Input module

601: Information acquisition unit

700:Multilingual translation model

U: User

S101, S102, S103, S104, S105, S106, S107, S108: Steps

S201, S202, S203, S204, S205, S206, S207, S208: Steps

S301, S302, S303, S304, S305, S306, S307, S308, S309: Steps

S401, S402, S403, S404, S405, S406, S407, S408: Steps

S501, S502, S503, S504, S505, S506, S507, S508, S509: Steps

S601, S602, S603, S604, S605, S606, S607, S608, S609, S610, S611: Step

S701, S702, S703, S704, S705, S706, S707, S708, S709: Steps

S801, S802, S803, S804, S805, S806, S807, S808, S809, S810, S811, S812, S813: Steps

S901, S902, S903, S904, S905, S906, S907, S908, S909, S910, S911, S912, S913; S914: Step

FIG. 1 is a schematic diagram of the prior art; FIG. 2 is another schematic diagram of the prior art; FIG. 3 is another schematic diagram of the prior art; FIG. 4 is another schematic diagram of the prior art; FIG. 5 is another schematic diagram of the prior art; FIG. 6 is another schematic diagram of the prior art; FIG. 7 is another schematic diagram of the prior art; FIG. 8 is a schematic diagram showing the system of the present invention; FIG. 9 is a schematic diagram showing the electronic device of the system of the present invention; FIG. 10 is a schematic diagram showing the server of the system of the present invention; FIG. 11 is a schematic diagram showing the application of the system of the present invention; FIG. 12 is a flow chart showing the method of the present invention; FIG. 13 is a flow chart showing another embodiment of the method of the present invention; FIG. 14 is a flow chart showing another embodiment of the method of the present invention; FIG. 15 is a flow chart showing the method of the present invention. Another embodiment flow chart; FIG. 16 is a flow chart showing another embodiment of the method of the present invention; FIG. 17 is a flow chart showing another embodiment of the method of the present invention; FIG. 18 is a flow chart showing another embodiment of the method of the present invention; FIG. 19 is a flow chart showing another embodiment of the method of the present invention; FIG. 20 is a flow chart showing another embodiment of the method of the present invention; FIG. 21A and FIG. 21B are schematic diagrams showing application examples of the present invention; FIG. 22 is a flow chart showing another embodiment of the method of the present invention; FIG. 23 is a flow chart showing another embodiment of the method of the present invention; FIG. 24 is a schematic diagram showing the use scenario of the method of the present invention; FIG. 25 is a schematic diagram showing the use scenario of the method of the present invention; and FIG. 26 is a schematic diagram showing the use scenario of the method of the present invention.

The inventive concept will now be more fully described below with reference to the accompanying drawings in which exemplary embodiments of the inventive concept are shown. The advantages and features of the inventive concept and the method of achieving the same will become apparent from the exemplary embodiments described in more detail below with reference to the accompanying drawings.

The terms used herein are used only to describe specific embodiments and are not intended to limit the present invention. Unless the context clearly indicates otherwise, the singular forms of the terms "a", "an" and "the" used herein are intended to include the plural forms as well. The term "and/or" used herein includes any and all combinations of one or more of the relevant listed items. It should be understood that when an element is said to be "connected" or "coupled" to another element, the element may be directly connected or coupled to the other element or there may be intermediate elements.

Exemplary embodiments are described herein with reference to figures, which are idealized exemplary illustrations. Therefore, deviations from the illustrated shapes due to, for example, manufacturing techniques and/or tolerances are expected. Therefore, the regions shown in the figures are schematic, and their shapes are not intended to illustrate the actual shape of the device nor to limit the scope of the exemplary embodiments.

Please refer to FIG. 8 and FIG. 9 . FIG. 8 is a schematic diagram showing the structure of the system of the present invention; FIG. 9 is another schematic diagram showing the structure of the system of the present invention. The system of the present invention is implemented by providing an electronic device 100 operated by a user U. The electronic device 100 is connected to a server 300 through a network 200. The electronic device 100 includes a processor 110, a memory 120 and a network interface controller 130. The server 300 includes an application 310. The processor 110 connects to the server 300 through the network interface controller 130 and executes the application 310, thereby configuring and activating the system. The online trademark application module 311 and the category recommendation module 312 are configured. Furthermore, the processor 110 is configured to enable the application data collection unit 3111, the category item selection unit 3112, the data transmission unit 3113, the text analysis unit 3121, the category item comparison unit 3122 and the report generation unit 3123. In addition, the processor 110 is also configured to enable a data transceiver module 313 to load the data in the multiple databases 400 into the memory 120.

The server can be a cloud server or a local server.

In this embodiment, the creator uses a server with the following specifications to perform the trademark product item conversion model training and inference execution of this technology; the processor (CPU) uses a high-performance multi-core processor, especially for processing large amounts of data and performing complex calculations, at least one CPU with 16 cores or more (such as AMD Ryzen Threadripper or Intel Xeon series) is used. And the size of the memory (RAM) can be but is not limited to 64GB or higher RAM specifications so that the size of the corpus and the size of the word vector model can be processed. The network interface controller is a high-speed and stable network connection hardware, especially for use in cloud computing resources or downloading/uploading large amounts of data. Among them, the most important graphics processing unit (GPU) uses a high-performance GPU (such as NVIDIA's RTX 30 series or Tesla series) to reduce the time of model training; in the server training architecture of this case, a higher-specification training server as above is used, and when the model training is completed and the inference processing is performed, a lower-specification server host can be used, and it can be a cloud or ground server host deployed in this system. In the architecture of this case, the use of server host specifications does not affect the technical features emphasized in this case, so any server host specifications should still fall within the technical scope of this case. The processor 110 is further connected to multiple delivery servers 500 via the network interface controller 130, so that the data transmission unit 3113 can upload the trademark application documents of user U to the delivery server 500 of the selected country.

Specifically, the user U operates the processor 110 of the electronic device 100 to execute the trademark online application module 311 of the application 310 to apply for the trademark online. The user U uses the electronic device 100 to input the data required for the trademark application, such as the trademark name, trademark image, applicant, contact person, product items to be applied for or description of the product or service, etc., and the display screen of the electronic device 100 will display the corresponding The fields are for user U to fill in the corresponding data. The processor 110 then activates the application data collection unit 3111, and classifies the data entered by user U in the corresponding fields and stores them in the memory 120. Since the trademark name, trademark image, product items to be applied for, and description of the product and service in the trademark application data will be particularly used later, the information in these fields is first classified by the application data collection unit 3111.

Specifically, the processor 110 executes the application 310 to activate the category item selection unit 3112 of the trademark online application module 311. The user U can directly select the product item for which the trademark is to be applied through the electronic device 100. The processor 110 accesses the trademark category and item list pre-stored in the memory 120 and displays them on the display screen of the electronic device 100, so that experienced or professional users can directly find the category items corresponding to the product or service, and continue to complete the trademark online application after selection.

If the user U is not an experienced or professional person, he can select the category recommendation option through the display screen of the electronic device 100, and the processor 110 will activate the category recommendation module 312 to recommend trademark product items, so that the user U can refer to the recommended items and quickly select the appropriate category items.

The processor 110 further activates the text parsing unit 3121 to read the trademark application data classified by the application data collecting unit 3111 in the memory 120, especially the product items and descriptions of the products and services, and performs semantic analysis on the text of the product and service description input by the user U to extract keywords, wherein the product and service description can be, for example: I want to open a cafe selling hand-brewed coffee, selling beverages, cakes, hand-made cookies and accessories, etc.

Specifically, keyword extraction is a natural language processing technology that aims to automatically extract important keywords or phrases from text. The method can be but not limited to statistical methods, frequency methods, statistical methods, text vectorization methods or machine learning methods.

Frequency methods judge the importance of words based on their frequency in the text. Common methods include TF-IDF (Term Frequency-Inverse Document Frequency) and Term Frequency. TF-IDF considers the frequency of a word in the text and its importance in the entire collection, while term frequency only considers the frequency of a word in the text.

Text statistics is based on statistical models to analyze the distribution and relevance of words in text. Common methods include mutual information, pointwise mutual information, and chi-squared test. These methods usually require the establishment of a statistical model between words and texts, and the calculation of word importance based on the model.

The text vectorization method converts the text into a vector representation and then uses the Vector Space Model to calculate the importance of words. Common methods include the Bag-of-Words Model, Word Embeddings, and text vectorization methods (such as TF-IDF vectorization).

Machine learning methods use machine learning algorithms to train models and learn the importance of words from text. Common methods include text classification, text clustering, and keyword extraction models. These methods require the use of labeled text data for model training.

The above can also be done by writing a program using a scripting language (such as Python).

After the text parsing unit 3121 extracts the keywords, the processor 110 then activates the category item matching unit 3122 to compare the product item input by the user U in the memory 120, the extracted keywords and the trademark category item. The trademark category item has been connected to multiple databases 400 through the data transceiver module 313 to store trademark category items of different countries in the memory 120 in advance. The trademark category item of the country to which the user U wants to apply is selected for comparison.

Specifically, the comparison method may be, but is not limited to, using a scripting language (such as Python) to write a program to read each category name one by one, and then use the program to read the trademark category item text in the memory 120 to search for matching or similar items, or using SQL language for comparison, using a SELECT statement to select category items that meet specific conditions, so as to find out whether they exist in the trademark category items.

If you use a scripting language (such as Python) to write a program, you can read each trademark category item one by one. Then, through the file processing function in Python, read the text in the trademark category item to search for matching or similar text. In order to perform text matching, you can use the string processing methods in Python, such as string matching, regular expressions, etc. Through these methods, the program can determine whether the text in the trademark category item contains words that match the product item or keyword to determine whether there is matching text.

In addition, the comparison method can be but is not limited to string matching algorithms (such as Levenshtein Distance, Jaro-Winkler Distance, etc.). Levenshtein Distance (also known as Edit Distance) is an algorithm that measures the similarity of two strings. The principle is to calculate the minimum number of operations required to convert one string to another. These operations can be to insert a character, delete a character, or replace a character.

For example, suppose there are two strings "kitten" and "sitting". To calculate the Levenshtein Distance between them, we need to convert "kitten" to "sitting" through the following steps: replace "k" with "s" to get "sitten"; replace "e" with "i" to get "sittin"; insert a "g" to get "sitting". Therefore, the Levenshtein Distance between the two strings is 3, which means that 3 operations are required to convert one string into another. Through this algorithm, we can calculate the distance between any two strings and use it to compare their similarities.

In addition, Jaro-Winkler Distance is an algorithm for measuring the similarity between two strings. It is an improved version of Jaro Distance and is usually used to compare shorter strings, such as names, addresses, etc. Jaro-Winkler Distance will calculate a similarity value between 0 and 1 based on the distance between the same characters and character positions of the two strings, where 0 means completely dissimilar and 1 means exactly the same. The main steps of the Jaro-Winkler Distance algorithm are as follows: Calculate the Jaro distance, which is the similarity between two strings. It mainly considers the order of the characters in the strings and the number of characters in common in the two strings; calculate the Winkler correction factor, which is mainly used to deal with similar strings. It adjusts the value of the Jaro distance by calculating the length of the common prefix and the value of a constant p; calculate the Jaro-Winkler Distance, Jaro-Winkler Distance = Jaro distance + Winkler correction factor.

After the category item comparison unit 3122 generates the comparison result, it is stored in the memory 120. At the same time, the processor 110 activates the report generation unit 3123 to sort the comparison results according to the word similarity. The completely identical or highly similar category items are listed as the priority recommended category items. Specifically, the similarity can be calculated in different ways during the comparison process, such as edit distance, cosine similarity or Jaccard similarity. After the similarity is sorted, a recommendation report is generated. The user U can clearly know how to choose the trademark category item through the recommendation report.

Furthermore, the processor 110 connects to the server 300 through the network interface controller 130 and executes the application 310, thereby configuring and activating a risk assessment module 314 to perform a previous case comparison on the trademark name in the trademark application data input by the user U, and the processor 110 executes the risk assessment module 314 in the application 310 to further configure and activate the text parsing unit 3121, the search comparison unit 3142 and the report generation unit 3123.

The text parsing unit 3121 performs text arrangement and combination on the trademark name in the trademark application data stored in the memory 120. Specifically, if a word is to be arranged and combined as a single word, the following steps can be used: first, the first character of the word is taken out as the starting character; then, the remaining characters are arranged and combined. This can be done by recursion or iteration. Recursion is a self-calling method. The process is to decompose the problem into smaller sub-problems and process them in a loop until the termination condition is reached. Iteration is to use loops to repeatedly perform the same operation until the termination condition is reached; insert the first character into a different position of each permutation combination to form a new permutation combination. The insertion position can be tried one by one from the beginning to the end, so that all possible permutations and combinations can be generated; repeat the above steps until all characters are processed.

It should be noted that the above arrangement and combination method is only for demonstration and not for limitation. Its main function is to help find similar names as much as possible in the subsequent comparison process.

After the text parsing unit 3121 completes all permutations and combinations of the trademark name input by the user U, it is stored in the memory 120, and the processor 110 simultaneously activates the search and comparison unit 3142. Prior to this, the data transceiver module 313 has been connected to the database 400 to store all the applied trademark information in the memory 120. These applied trademark information are the trademark cases, which serve as the basis for subsequent comparison.

Specifically, the search and comparison unit 3142 compares all permutations and combinations in the memory 120 with the previous trademarks in the memory 120. The comparison method is similar to that of the above-mentioned category item comparison unit 3122, so it will not be described in detail.

The comparison results generated by the search and comparison unit 3142 will also be stored back in the memory 120. At the same time, the processor 110 activates the report generation unit 3123 to sort the comparison results by similarity. The similarity calculation method is the same as the above-mentioned category recommendation report. The higher the similarity, the higher the ranking order, and the risk assessment report is generated. The report can be presented in the form of an online image, a downloadable document file, a file format that can be forwarded to a communication software, a file format that can be shared to a social software, an electronic bulletin board, or a combination thereof.

In another embodiment, the processor 110 may simultaneously enable the category recommendation module 312 and the risk assessment module 314, and the report generation unit 3123 may generate a comprehensive report combining trademark category recommendation and risk assessment.

Specifically, the text parsing unit 3121 simultaneously accesses the trademark name, product item, and product service description input by the user U in the memory 120, and performs text arrangement and combination on the trademark name, and keyword extraction on the product service description, and stores them back in the memory 120. Then the processor 110 activates the category item matching unit 3122 and the search matching unit 3142, respectively, to perform corresponding comparisons on the keywords and all the text arrangement combinations, and stores the generated comparison results back in the memory 120.

Next, the processor 110 activates the report generation unit 3123 to read the comparison results in the memory 120, and arranges the trademark category items and the similarity of the trademark names. Specifically, the similarity of the trademark names is first arranged, and then the similarity between the category items to which the trademark belongs and the recommended trademark category items is analyzed and calculated. If the trademark names are highly similar and the category items to which the trademark belongs meet the highly recommended trademark category items, it means that the trademark predecessor is a highly similar predecessor, which will be presented in a prominent color in the comprehensive report.

The colors displayed in the comprehensive report are used to clearly distinguish the trademark category items that User U tries to avoid. Such trademark category items have many highly similar trademark precedents, providing User U with a clear way to choose appropriate and lower-risk category items.

Please refer to FIG. 10 for another embodiment of the present invention. The system is used to receive a string provided by a user terminal through an electronic device and at least one target country to be queried. A processor of the electronic device is connected to a server through a network interface controller and executes an application program to calculate and generate a trademark category recommendation report. The system at least includes: an input module 600, a multilingual translation model 700, a category recommendation module 312, a risk assessment module 314 and a database 400, wherein the database 400 can be the Nice trademark category database, trademark category databases of multiple countries, and a multinational trademark database.

Among them, the Nice Trademark Classification Database and the Trademark Classification Database of Multiple Countries mainly store the classification and product item data of the Nice International Trademark Classification, as well as the trademark class classification and product item data of each corresponding country, while the Multinational Trademark Database stores the trademark application data of each corresponding country.

The user can choose any combination of brand owners, trademark applicants, trademark practitioners, and legal practitioners.

The input module 600 is used to receive the string input by the user, tokenize the string, and send a string information. Specifically, string tokenization is the process of splitting a sentence or document into individual words (tokens). The following is an example of a simple string tokenization method: Remove punctuation: Use regular expressions or predefined punctuation lists to remove punctuation in the string, such as periods, commas, question marks, etc. Split words: Split the string according to spaces or other specific separators, and each split part is a word. Handle special cases: Handle special cases, such as abbreviations, hyphens, numbers, etc. Regular expressions or specific rules can be used to handle these cases and split them into appropriate words. Convert to lowercase (if the language is case-sensitive): Convert all words to lowercase to unify the way words are represented.

The multilingual translation model 700 is a computational model trained through large-scale language translation. When receiving the string information, it determines whether the input string language is consistent with the official language of the target country selected by the user. If not, the string input by the user is translated into the official language of the target country and the translated string information is sent.

The category recommendation module 312 is an operational model trained through a natural language model and a trademark classification table and detailed items, and is used to parse the string information of ambiguous semantics (or inaccurate description) into trademark category recommendation information, which further includes: a text parsing unit 3121, a category item matching unit 3122, and a report generation unit 3123.

A natural language model can predict the next word or generate a sentence that conforms to grammar and semantics based on known text data. It can be, but is not limited to: N-gram model: The N-gram model is a probability-based language model that assumes that the probability of a word appearing is only related to the previous N-1 words. For example, in a bigram model, for a given previous word, the probability of predicting the next word.

Recurrent Neural Network (RNN) Model: RNN is a neural network suitable for processing sequence data, which can capture the temporal dependencies between words. In natural language processing, RNN is often used to construct language models, where each word is regarded as a time step.

Pre-trained language models (e.g. BERT): Pre-trained language models are models obtained through large-scale unsupervised training that can understand and generate natural language. The BERT model uses the Transformer network architecture to be pre-trained on a large amount of text data, and then fine-tuned to adapt to specific tasks such as text classification, named entity recognition, etc.

The trademark classification table and details are the classifications extracted from the trademark databases of various countries, as well as all the product items in the classifications. They are extracted and listed to provide natural language models for training and calculation.

The text analysis unit 3121 is used to analyze the translated string information and extract keywords and descriptive texts related to business activities, products, services, etc., collectively referred to as industry information.

The category item comparison unit 3122 is used to receive the relevant description texts of the keywords and industry information, and compare them with the trademark category database of the target country, and calculate and generate the closest at least one category information and one product item information and one service item information of the category, that is, the recommended category and recommended product item. One recommended category is matched with at least one recommended product item.

The report generation unit 3123 is used to receive the recommended category, the recommended product item and the target country, and generate a trademark category recommendation report in a template format using the input language. Specifically, if the language of the string initially input by the user is different from the official language of the target country, the generated trademark category recommendation report will be translated back to the language used when the string was input through the multilingual translation model 700.

Specifically, the input string may contain trademark words, trademark descriptions, business practices, company names, stock symbols, product names, service names, or a combination thereof.

The user can input strings through text input, voice input or video input.

The input module 600 further includes an information acquisition unit 601, which is used to actively send the labeled string information to the network space to acquire relevant information, and can transmit the information to the multilingual translation model 700 for translation.

Specifically, the information acquisition unit 601 crawls information in the cyberspace based on the industry-related information generated after the text analysis unit 3121 analyzes the string input by the user. The crawled information may be the user's official website, news or other online articles containing the extracted keywords.

The risk assessment module 314 is used to compare the keywords extracted from the input string with the trademark database of the target country and generate risk information. The risk information can be integrated into the trademark category recommendation report through the report generation unit 3123 to generate a risk assessment report.

Specifically, the recommended trademark categories and recommended product items are generated after calculation by the category recommendation module 312, and then the risk assessment module searches and compares the recommended categories. First, it compares whether there are similar names in the target country's trademark database with the keywords. If there are similar previous cases in the recommended categories, the similarity of the product items is compared. Finally, a risk assessment report is generated after cross-analysis.

Please continue to refer to Figure 11. The category recommendation module 312 of this system further includes a knowledge graph comparison unit 3124. Specifically, the information acquisition unit 601 can crawl the user's official website or related network information, and can also crawl company information of similar industries or similar backgrounds based on the parsed industry information, and then compare these information through the knowledge graph comparison unit 3124, and search out the previous trademark cases applied for by companies of similar industries or similar backgrounds, and extract the categories applied for in these cases and the product items therein and merge them into the recommended categories and recommended product items compared by the category item comparison unit 3122.

The Knowledge Graph, in which each node represents the trademark application case record data of different companies, and the edge represents the industry relevance or similarity between the companies. This method can intuitively show the relationship between the categories/product items and industries in the trademarks applied for by different companies. The construction of this knowledge graph is collected over time, including but not limited to the following information: the announcement information of registered trademarks (trademark families), the previous case comparison table constructed through word meaning expansion and translation comparison, and some trademark application cases disclosed by the National Intellectual Property Administration.

Moreover, its detailed construction process is as follows: Data collection: Collect existing trademark data, including trademarks applied by the company, trademark names, categories and other related information. These data can be obtained from sources such as trademark databases and patent offices.

Knowledge graph construction: Combine trademark data with other relevant information, such as company information, industry categories, etc., to construct a knowledge graph. Knowledge graphs can use graphical databases or graphical representations to represent the relationships between different entities (such as companies, trademarks, industries).

Similarity calculation: Calculate the similarity between different trademarks based on the relationships and attributes in the knowledge graph. This can be based on different features, such as similarity of trademark names, similarity of categories, etc.

Please refer to Figure 12 for another embodiment of the present invention, which shows a flow chart of the method of the present invention, which includes steps S101~S108.

In step S101, the user inputs the introduction text, mainly by operating the electronic device 100 to input any text, and it is not limited to the trademark name or standardized words; in step S102, the text parsing unit 3121 performs semantic analysis on the introduction text input by the user, and determines whether the input text contains the user's brand name, company name, company introduction, product introduction and other information; if so, step S103 is performed, and the text parsing unit 3121 extracts the brand name and company name from the introduction text input by the user; and then step S104 is performed to determine the identity of the extracted text, mainly because one of the approval conditions for the trademark is that it must have identity. In step S105, the text parsing unit 3121 performs semantic analysis on the introduction text input by the user, and determines whether the input text contains the user's brand name, company name, company introduction, product introduction and other information; if so, step S103 is performed, and the text parsing unit 3121 extracts the brand name and company name from the introduction text input by the user; and then step S104 is performed to determine the identity of the extracted text, mainly because one of the approval conditions for the trademark is that it must have identity. 1. First, exclude non-identifying words and extract identifying words. For example, "株式会社" in ABC股份有限公司 is a non-identifying word. If the above judgment is negative, after step S105, proceed to step S106, and the introduction words or identifying words entered by the user are again analyzed by the text analysis unit 3121 for industry, and also for product and In step S107, the service content analysis will also activate the category item comparison unit 3122 to compare the identification text and the introduction text after analysis in step S106 with the trademark category items in the database 400, and in step S108, the comparison result will be generated into a category recommendation report through the report generation unit 3123, and the recommendation reason will be attached.

Another embodiment of the present invention is different from the embodiment of FIG. 12 in that after step S106, the introductory text or identifying text input by the user is again analyzed by the text parsing unit 3121 in step S106, and the product and service content is also analyzed. Furthermore, the category item comparison unit 3122 first compares the analyzed industry description to perform a comparison. The recommended trademark category is generated, and then the category item comparison unit 3122 compares the parsed product and service content descriptions in the generated recommended trademark category, thereby generating recommended product items of the trademark category. Finally, the process proceeds to step S108, and the comparison results are formed into a category recommendation report through the report generation unit 3123, and the recommendation reason is further attached.

Please continue to refer to Figure 13, which shows a flowchart of another embodiment of the method of the present invention, which includes steps S201~S208.

In this embodiment, in step S201, the user copies the text description of the company introduction information by operating the electronic device 100 and then pastes it. In particular, the user inputs a text description in step S201 instead of simple words or nouns; in step S202, the text analysis unit 3121 performs semantic analysis on the text input by the user and extracts keywords; and determines whether the input text description contains information such as the user's brand name and company name, and the subsequent steps S203~S208 are the same as the aforementioned steps S103~S108, so they are not repeated here.

Please continue to refer to Figure 14, which shows a flowchart of another embodiment of the method of the present invention, which includes steps S301~S309.

In step S301, the user inputs the introduction text by operating the electronic device 100, and the text analysis unit 3121 determines whether the input language is the default received language and whether translation is required; if translation is required, step S302 is performed first to perform language translation; if translation is not required or after completing step S302, step S303 is performed, and the text analysis unit 3121 performs semantic analysis on the introduction text input by the user and extracts keywords; and the subsequent steps S304~S309 are the same as the aforementioned steps S103~S108, so they are not repeated.

Please continue to refer to Figure 15, which shows another method flow chart of the present invention, which includes steps S401~S408.

In step S401, the user uses voice to input the explanatory text by operating the electronic device 100, and the text parsing unit 3121 converts the voice into text and stores it in the memory 120. The voice recognition technology can compare the audio data with the voice model based on the model and algorithm, identify the corresponding text content, and perform post-processing of the text, such as grammar correction, spelling correction and semantic analysis; then judge the text to determine whether it contains the user's brand name, company name and other information; and the subsequent steps S403~S408 are the same as the aforementioned steps S103~S108, so they are not repeated.

Please continue to refer to Figure 16, which shows a flowchart of another embodiment of the method of the present invention, which includes steps S501~S509.

Steps S501 to S507 in this embodiment are the same as the aforementioned steps S101 to S107, and therefore will not be described in detail. In step S508, the search comparison unit 3142 is activated to compare the extracted identifying text with the applied trademark previous cases in the database 400, and the comparison results are stored in the memory 120; in step 509, the comparison results are arranged by the report generation unit 3123 to generate a category recommendation report and a list of similar previous cases, and the reasons for the category recommendation are also attached.

Please continue to refer to Figure 17, which shows a flowchart of another embodiment of the method of the present invention, which includes steps S601~S611.

In this embodiment, the main purpose is to provide users with trademark category recommendations of different countries (multinational). Steps S601 to S608 are the same as the aforementioned steps S101 to S108, and therefore will not be described in detail. After step S608, the display screen of the electronic device 100 displays whether to perform a transnational product item conversion for the user to choose. If not, the process ends. If yes, step S609 is performed to provide multiple countries in the database 400 for the user to choose from. Different from the country of the original application, at least one second application country is selected. After the user selects a country in step S610, the processor 110 reads the trademark category items of the country in the memory 120 and activates the category item comparison unit 3122 for comparison. In step S611, the comparison result is displayed on the display screen of the electronic device 100, so as to achieve the effect of transnational trademark category recommendation.

Please continue to refer to Figure 18, which shows a flowchart of another embodiment of the method of the present invention, which includes steps S701~S709.

In this embodiment, after the user inputs the introduction text by operating the electronic device 100 in step S701, the application country is selected first. After the trademark application country is determined, step S702 is performed. The text parsing unit 3121 performs semantic analysis on the introduction text input by the user and determines whether the introduction text contains information such as brand name, company name, etc. The subsequent steps S703~S707 are the same as the aforementioned steps S103~S107, so they are not repeated here. In step S708, the comparison result generated in step S707 is converted across countries according to the country selected by the user in step S701, and the original recommended category items are converted into commodity items of the selected country; in step S709, the report generation unit 3123 outputs the conversion result to generate a cross-country category recommendation report.

Please continue to refer to Figure 19, which shows a flowchart of another embodiment of the method of the present invention, which includes steps S801~S813.

In this embodiment, category recommendations and risk assessments that can be converted across countries are also provided, and steps S801 to S806 are the same as the aforementioned steps S101 to S106, so they are not repeated here. In step S807, the extracted identifiable text and the parsed text are converted to English by the text parsing unit 3121. In this embodiment, all text is converted to English by default. In step S808, the category item comparison unit 3122 compares the text converted to English with the information data that has been pre-converted to English in the database 400. In steps S809 and S810, the product items are compared and the applied cases are compared. The case name is compared with the case name that has been applied for through the search and comparison unit 3142; in step S811, the comparison results of steps S809 and S810 are comprehensively analyzed by the report generation unit 3123; in step S812, the report generation unit 3123 distinguishes the priority of the category recommendation by color in the comparison result; in step S813, the report generation unit 3123 comprehensively analyzes the category recommendation and the search result and integrates them to generate a cross-border risk assessment report.

Please continue to refer to Figure 20, which shows another method flow chart of the present invention, including steps S901~S914.

This embodiment also provides a cross-national category recommendation and risk assessment method, and steps S901 to S906 are the same as the aforementioned steps S101 to S106, and step S907 is the same as the aforementioned step S807, so they are not repeated. In step S908, the text parsing unit 3121 is activated to generate synonyms or similar words for the extracted identification text and the parsed text. The method of generating similar words or synonyms can be, but is not limited to, the text parsing unit 3121 using an existing vocabulary, including synonyms or similar words of the vocabulary into the vocabulary, and generating similar words or synonyms by matching the words in the vocabulary. For example, WordNet is a synonym generation system based on vocabulary expansion, or using a large corpus to learn the relationship between words, including word co-occurrence and context similarity, to generate similarity scores between words, and then generate similar words or synonyms. For example, LSI (Latent Semantic Indexing) and LDA (Latent Dirichlet Allocation).

WordNet is a computerized vocabulary of English words, containing a large number of English words, and organizing and managing words based on the semantic and grammatical relationships of the words. The purpose of WordNet is to provide a reliable language resource for natural language processing and semantic analysis.

LSI (Latent Semantic Indexing) is a natural language processing technology used to identify synonyms and similar words. Its basic principle is to identify the semantic relationship between words by converting the text into a vector space model and then performing singular value decomposition (SVD). Specifically, LSI uses the word frequency statistics method in the text to process the words in the text and convert them into a vector space model. Before performing SVD, LSI usually uses TF-IDF (Term Frequency-Inverse Document Frequency) weights to weight the words to eliminate the influence of some common words, so as to better capture the correlation between words. Through such processing, LSI can generate a word-text matrix in which each word corresponds to a vector. Through singular value decomposition, LSI can decompose the singular values and singular vectors of this word-text matrix to capture the implicit semantic information in the text. Based on this semantic information, LSI can calculate the similarity between two words to find synonyms or similar words.

In addition, LDA (Latent Dirichlet Allocation) is a commonly used topic model that can distribute text in a collection to multiple topics and find the vocabulary distribution represented by each topic. This vocabulary distribution can be used to find synonyms or similar words. First, you need to prepare a text set, which can be any text set containing the target vocabulary. Then, use LDA to distribute this text set to multiple topics and find the vocabulary distribution represented by each topic. For a target vocabulary, you can use its corresponding vocabulary distribution in the LDA model to find synonyms or similar words. Specifically, you can calculate the similarity between the vocabulary distribution of the target vocabulary and the vocabulary distribution of other words, and find words similar to the target vocabulary as synonyms or similar words.

The text parsing unit 3121 can also use a scripting language (such as Python) to write a program to execute WordNet, LSI (Latent Semantic Indexing), LDA (Latent Dirichlet Allocation) to perform text parsing and generate synonyms or similar words.

It is worth pointing out that the machine translation in the present invention adopts a direct method and a deep learning-based machine translation model, such as Google's Transformer model or Facebook's FAIR SEQ model, to directly translate the trademark product item language of one country into the language of another country. Then, the product item database of another country is entered for direct text comparison.

In addition, a knowledge graph of multinational trademarked product items is established in the present invention. The knowledge graph is used to establish node matrix associations, where each node represents a product item from a different country, and the edge represents the correlation or similarity between product items. Based on the dynamic learning model of time evolution based on the equivalence relationship between product items and trademark items of each country, the knowledge graph unit conversion includes at least the following steps: collecting data on the relationship between multinational trademarked product items and trademark projects and learning dynamic models, and calculating the best match, shortest path, and maximum flow through graph query and natural and abrasive reasoning, and optimizing the equivalence relationship of product items based on the final decision result. This method can intuitively show the conversion relationship between commodity items in different countries. The construction of this knowledge map is collected over time, including but not limited to the following information: information on the announcement of registered transnational trademarks (trademark families), the results of conversion completed by previous users, a commodity item comparison table constructed through word meaning expansion and translation comparison, and a partial commodity item translation comparison table published by the National Intellectual Property Administration.

The detailed construction process is as follows: In the cross-border application transfer technology of trademark goods or services, the knowledge graph may include the following elements: Entity: trademark goods items in various countries.

Relationship: Similarity or equivalence between product items. For example, a certain trademark product item in the United States may have the same or similar meaning as a certain product item in Germany. The following steps are followed to use the knowledge graph for cross-border trademark application conversion: Data collection: First, collect data on trademark product items in various countries and the relationships between these product items. This part requires a lot of manual effort, and partly requires automated data crawling and processing technology. And use the collected data to construct a knowledge graph, in which each product item is a node and the relationship between each node is an edge.

Graph query and reasoning: Using query and reasoning techniques, the best match between the trademark product project of the original country and the product project of the target country is found based on the knowledge graph, which involves problems such as the shortest path and maximum flow in graph theory. Finally, the result evaluation and feedback are performed: the query and reasoning results are evaluated. If the results are inaccurate or unsatisfactory, the knowledge graph can be updated based on the feedback.

On the other hand, in the technology of the present invention, the word vector analysis is used in the cross-border application conversion technology of trademark product items to capture and understand the semantic relevance of product items in the trademark classification systems of different countries and realize the conversion process.

The following provides one of the application methods of the present invention and the database model and training method: Establish a multilingual word vector model: First, we convert the descriptions of trademark product items in different countries into word vectors. This is achieved by training a multilingual word vector model. For example, we can use a large-scale public multilingual text dataset to train this model, or use a pre-trained multilingual word vector model, such as Facebook's FastText but not limited to the above language models.

Transform product item descriptions: Transform the product item descriptions of the trademark in the original country into word vectors, and then transform these word vectors into the language of the target country. This step can be achieved through the similarity between word vectors. For example, we can find the nearest neighbors of the word vectors in the original language and the word vectors in the target language, and use these nearest neighbor words to form new product item descriptions.

Manual review and adjustment: Although the word vector model is powerful, it still cannot fully understand the subtle differences in language and cultural differences. Therefore, after the conversion is completed, the training effect can be strengthened through manual review and correction. For example, a team of experts can review and correct the conversion results, and use these corrected data to train the model again to continuously improve the conversion accuracy of the model. And continue to optimize the model: Over time, the model can be continuously optimized according to actual needs and effects, for example, adding support for more languages, or improving the ability to handle specific types of product items.

The above method is just one possible way, and the actual implementation may be adjusted according to specific needs and conditions. For the actual conversion case in this case, it is completed through the following "word vector" conversion model, which includes: Word vector training: Use a large-scale corpus to train the word vector model, and map each word to a high-dimensional space based on the context, where similar words will be close to each other in this space. The conversion formula is the following target function:

where T is the number of words in the corpus, c is a chosen window size, θ is the model parameter, and p ( w_ { t + j }| w_t ) is the conditional probability of the context word w_ { t + j } given the center word w_t.

Conversion function: Using the trained word vector model, we can generate corresponding word vectors for the product items of country A and country B, and we establish a conversion function to map the word vector of the product item of country A to the word vector of the product item of country B. And minimize the distance between the word vector of country A and its corresponding word vector of country B. Suppose we have a set of product item pairs of country A and country B (x_i, y_i), where i=1,...,N, N is the number of pairs, then the formula for minimizing the objective is: min _ W Σ _{ i =1}^{ N }∥ Wx _ i - y _ i ∥^2

Where W is the transformation matrix we want to learn. Assuming that the U.S. trademark product item we want to transform is "computer software", we first convert this phrase into a word vector, and using the trained word vector model, we can get the word vector representation v_US of "computer software". Then, we apply the transformation matrix W to v_US, that is, W * v_US, and the result is the transformed word vector v_TW. Next, we need to find the Taiwanese trademark product item that is closest to v_TW. This can be achieved by calculating the similarity between v_TW and the Taiwanese trademark product item word vector. The similarity is generally measured using cosine similarity. The calculation formula is as follows: sim ( a , b ) = a . b /(∥ a ∥＊∥ b ∥) where a and b are two word vectors, a.b represents the dot product of vectors, and ∥a∥ and ∥b∥ represent the norms of vectors a and b respectively.

Finally, the system searches all Taiwan trademark product item word vectors and finds the item with the highest cosine similarity to v_TW. This item is the result of the conversion. Among them, the system can determine that the most similar Taiwan trademark product item found is "computer software" and set it as the result of the conversion.

Among them, in the framework of this system and method, ensuring the update of the commodity item database of each country is the core focus. Since the commodity item database of each country may be updated once or twice a year, old commodity item names may be eliminated and new commodity items may be added. Among them, if the official has informed that the newly added commodity item corresponds to a certain commodity item of the original old version, such information must be updated in the knowledge atlas of the system; on the contrary, if the official has not informed the corresponding relationship, the corresponding relationship can be maintained through word image comparison or manual methods.

In step S909, the text in steps S907 and S908 is compared with the language-converted trademark category items in the database 400 through the category item comparison unit 3122; in steps S910 and S911, the text is analyzed to generate category recommendations, and the previous cases are searched through the search comparison unit 3142 based on the identification text; in step S912, the report generation unit 3123 comprehensively analyzes the search results and the comparison results, and distinguishes the recommended categories by color in step S913, and then in step S914, the report generation unit 3123 compares the searched previous cases in the recommended category to generate a risk assessment report.

Next, please refer to Figures 21A and 21B, which are schematic diagrams showing application examples of the present invention. As shown in the figure, the user can enter any descriptive text or introduction text in the webpage of this system (https://inta.aiplux.com/). The text content can include but is not limited to brand name, company name, product description, service description, company philosophy, company introduction, etc. After entering the descriptive text, you can press the analysis button.

The results are generated immediately after the category recommendation module of this system calculates, and the results include the brand name, company name, trademark name extracted from the description text, and the industry information parsed from the description text. Most importantly, the recommended trademark category is generated after comparison based on the parsed industry information, and the recommended product items in the trademark category.

Next, this case will explain and demonstrate the specific process, and further explain that when the user enters any description into the system of the present invention, the background operation process of the system is specifically divided into the following five main actions and implemented using the corresponding technologies: entity recognition (NER), topic modeling or text classification, discriminative text judgment, trademark category recommendation, and trademark goods and services recommendation. For example, when the user enters the system message "AIPLUX Technology Co., Ltd. published Taiwan's first large language model, which was established by supercomputers with up to 176 billion parameters, combining semantic understanding and text generation capabilities to launch an enterprise-level generative AI solution. 'For example, in the entity recognition (NER) stage: the system uses the NER model to further identify the organization name in the text (such as "AIPLUX Technology Co., Ltd.").

Furthermore, in the topic modeling or text classification stage, it helps the system understand the main topics in the text and classify them into corresponding industries. It then uses topic modeling to identify keywords such as "artificial intelligence", "natural language processing" and "big data" and classify them into corresponding industry categories. In addition, in the discriminative text judgment stage, the system needs to find keywords in the text that can represent the main name of the company. In this example, the system will consider "AIPLUX" as a discriminative term.

In the trademark category recommendation stage, the system recommends the international category of the trademark based on the above information. Among them, if a company is mainly engaged in artificial intelligence, natural language processing and big data related businesses, the system will recommend relevant trademark categories such as 9 categories (scientific instruments), 42 categories (scientific and technical services) and 41 categories (education and entertainment services). Finally, in the trademark goods and services recommendation stage, the system recommends related goods and services items based on the company's industry category and main business. Among them, if a company is mainly engaged in the research and development of artificial intelligence, we can recommend goods and services related to artificial intelligence, such as "artificial intelligence consulting services", "artificial intelligence technology research", etc.

To explain in more detail, entity recognition (NER) is a natural language processing (NLP) technology and is used to identify words with specific meanings in text, such as names of people, organizations, places, etc. It is not limited but can be completed using the following models: hidden Markov model (HMM), maximum entropy Markov model (MEMM), conditional random field (CRF), or neural network-based models such as BiLSTM-CRF. Among them, in this case demonstration, it is implemented through the BiLSTM-CRF model. The specific steps are to embed each word and convert the word into a dense vector representation. The word embedding can be obtained using pre-trained models such as word2vec, GloVe (Global Vectors for Word Representation), CKIP (Chinese Knowledge and Information Processing) or BERT (Bidirectional Encoder Representations from Transformers). The word embedding is processed using a bidirectional long short-term memory (BiLSTM) network. At this stage, the representation of each word in its context can be obtained. Then, the output of the BiLSTM is sent to the CRF layer: the CRF layer can consider the dependency between labels, making the sequence labeling result more accurate.

In this demonstration, the user inputs "AIPLUX Technology Co., Ltd. has released Taiwan's first large language model for enterprise use. The model has up to 176 billion parameters built by supercomputers, combining semantic understanding and text generation capabilities to launch an enterprise-level generative AI solution. 』 Take the sentence as an example, split the sentence into word sequences: ["AIPLUX Technology Co., Ltd.", "published", "Taiwan", "first", "enterprise use", "big", "language model", "the model", "through", "super", "computer", "established", "out", "of", "up to", "1,760", "billion", "parameters", "combined", "traditional Chinese", "of", "semantic understanding", "with", "text generation capability", "launched", "enterprise-level", "generative", "AI", "solution"], and further use pre-trained word embedding models (such as word2vec, GloVe or BERT) to convert each word into a vector.

For example, we assume that the word vector of "AIPLUX Technology Co., Ltd." is [0.1, 0.2, ..., 0.5] (in fact, a word vector usually has hundreds to thousands of dimensions), the word vector of "publish" is [0.2, 0.3, ..., 0.6], and so on. Feed the word vector into the BiLSTM network. BiLSTM can obtain contextual information for each word. For example, given the word vector of "AIPLUX Technology Co., Ltd.", BiLSTM can generate a new vector based on the words before and after it ("publish" and "Taiwan"), which contains the contextual information of "AIPLUX Technology Co., Ltd." in the sentence. Feed the output of BiLSTM into the CRF layer. The CRF layer will give the label probability of each word. For example, CRF may determine that the label of "AIPLUX Technology Co., Ltd." is "ORG" (representing the organization name), the label of "publish" is "O" (representing a non-entity word), and so on.

Finally, we select the label sequence with the highest probability as the result of NER. For example, for the above sentence, the NER result may be: ["ORG","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O","O"]

Therefore, according to this result, "AIPLUX Technology Co., Ltd." is identified as an organization name, and the rest of the words are identified as non-entity words.

Next, in the topic modeling or text classification stage, topic modeling usually uses unsupervised learning algorithms, such as Latent Dirichlet Allocation (LDA), to identify the topic of the document by analyzing the word frequency of the document. Text classification uses supervised learning algorithms, which require pre-annotated data. The algorithm technology used in this case includes but is not limited to the following algorithms: Naive Bayes, support vector machine (SVM), or deep learning methods such as convolutional neural network (CNN), long short-term memory network (LSTM), etc. In this demonstration, the main technology of the topic modeling LDA algorithm used is to regard each document as a mixture of topics, and each topic can be regarded as the probability distribution of words. In practice, LDA uses Dirichlet distribution to model the distribution of documents to topics and the distribution of topics to words, and then through repeated calculations and optimization, we get the distribution of topics for each document and the distribution of words for each topic. After word segmentation, we get the term representation of the document, and in the calculation process of LDA, the system can get the term distribution of each topic and the topic distribution of each document, and then find the keywords of each topic by checking the term distribution of each topic. Among them, "artificial intelligence", "natural language processing" and "big data" are determined as keywords of the same topic, and then excerpted.

In the identification word determination stage, keyword extraction can be done through methods such as TF-IDF or TextRank, where TF-IDF considers term frequency (TF) and inverse document frequency (IDF) to evaluate the importance of a word. The main task of identification word determination is to find those parts that are unique to the trademark, that is, those words that can explain consumers to identify and distinguish the source of the product. In the example you provided, "AIPLUX" may be determined as an identification word because it may be a specific brand name or company name, while "科技股份有限公司" may be determined as a non-identifying word because it is a general term that cannot help consumers distinguish the source of the product. This task can be achieved through the following steps: first, build a dictionary containing all registered trademarks, which can be obtained from the trademark databases of various intellectual property offices; second, build a dictionary containing all common words, which can be obtained from large-scale text data, such as Wikipedia or other large corpora. Finally, use the above two dictionaries to identify the identifiable and non-identifiable words in the text.

Among them, simply using the above two dictionaries for evaluation still has deficiencies and the risk of judgment distortion. Identification words usually have certain uniqueness and are therefore unlikely to appear in common vocabulary dictionaries. However, this assumption is not always true, and some identification words may also be common words. Therefore, in practical applications, we may need to combine other information to improve the accuracy of identification word judgment, such as using machine learning models to predict whether a word is identification, or using knowledge graphs to understand the meaning of a word in a specific context.

Based on the above, in the embodiment of the present invention, the system further uses the TF-IDF method to extract keywords. It is then used to evaluate the importance of a word in a document set. The calculation formula of TF-IDF is: TF - IDF ( t , d , D ) = TF ( t , d ) * IDF ( t , D )

Where: t is a word, d is a document, D is the set of all documents, TF(t,d) is the term frequency of word t in document d, which can usually be calculated by counting the number of times word t appears in document d and dividing it by the total number of words in document d; IDF(t,D) is the inverse document frequency of word t, which can be calculated by calculating the total number of documents divided by the logarithm of the number of documents containing word t.

For "AIPLUX Technology Co., Ltd.", the system determines that it is a document containing two words, in which "AIPLUX" appears once and "Technology Co., Ltd." also appears once, so their TF values are both 1/2=0.5. Then we need to calculate the IDF value. Assuming that our document library contains 10,000 documents, "AIPLUX" only appears in 10 documents, and "Technology Co., Ltd." appears in 1,000 documents, so their IDF values are log(10000/10)=3 and log(10000/1000)=1 respectively. Finally, we multiply the TF value and the IDF value, and get the TF-IDF value of "AIPLUX" is 0.5 * 3=1.5, and the TF-IDF value of "Technology Co., Ltd." is 0.5 * 1=0.5. From this we can see that "AIPLUX" is more important than "Technology Co., Ltd." in this document.

Furthermore, the present invention adopts the knowledge graph technology to further optimize the identification text judgment effect in the present technology, wherein the knowledge graph can store a large amount of information about entities and their attributes and relationships. In this problem, we can take "AIPLUX" and "科技股份有限公司" as entities, their attributes can include the number of times they appear in the document, their TF-IDF values, etc., and their relationship can be that they appear together in the same document, etc. In addition, the knowledge graph in this case integrates and stores data from the trademark review process of the Intellectual Property Office. For example, during the review process, "科技股份有限公司" or "股份有限公司" was mentioned many times and was considered as non-identifiable text. Then we can add this information to the knowledge graph, and add a "Identification: Low" label to the attribute of "科技股份有限公司"; in this way, in the subsequent keyword extraction process, we can use this information to assign a lower weight to the term "科技股份有限公司", making the model more inclined to identify "AIPLUX" as a keyword. This will help us better understand the entities in the text, their attributes, and relationships, so as to more accurately extract keywords and make identification text judgments.

In the stage of trademark category recommendation, the system uses the keywords and topics previously extracted from the text, as well as their corresponding industry classifications, to recommend corresponding trademark categories for users. For example, for a company that mainly engages in artificial intelligence, natural language processing, and big data-related businesses, the system will recommend relevant trademark categories such as Category 9 (scientific instruments), Category 42 (scientific and technical services), and Category 41 (education and entertainment services). The process is achieved through the use of a lookup table and a recommendation system algorithm. The system creates a lookup table in the background to map keywords or topics to their corresponding trademark categories, as shown in the following table:

Then, the system backend can find the corresponding trademark categories based on the keywords or topics in the text entered by the user, and recommend these categories to the user. At the same time, the system backend merges the industry information results of the user, enters the business information query platform, finds other companies in similar industries, and then integrates and summarizes the previous trademarks applied for by these other companies. Finally, the word vector comparison analysis recommendation results are combined with the industry information comparison analysis recommendation results to integrate the recommended trademark categories and product items (goods and services) selection suggestions.

The system backend can not only select goods and services through the general Nice classification content, but also select goods and services from various countries based on the product information provided by the user. Specifically, the system backend provides word vector analysis of the goods and services added by each country, and then performs word vector analysis and comparison based on the user's introduction information, and finally selects the corresponding recommended "goods and services" content.

Among them, it mainly includes two stages: one is to recommend products and services by comparing word vectors, and the other is to further optimize the recommendation results by comparing industry information. When the system background already has word vectors for products and services of various countries, the system can first convert the product name in the user introduction information into a word vector. For example, assuming that the product introduced by the user is "AI intelligent robot", in this demonstration process, the Word2Vec model is used to convert this word into a word vector v.

Then, the system background can traverse the word vectors of goods and services in various countries and calculate their similarity with v, and use cosine similarity to calculate the similarity between word vectors. The formula is as follows: similarity = cos (θ) = v ＊ w /(∥ v ∥＊∥ w ∥)

Among them, v and w are two word vectors, and "∥" represents the length of the vector. Finally, the system can select the products and services with the highest similarity as the recommendation results.

On the other hand, in the process of recommending products and services through industry information comparison, the system background uses the user's industry information to find other companies in the same industry, and then finds the most common products and services from the trademark applications of these companies. Specifically, first search for companies in the same industry on the business information query platform, and then collect the trademark applications of these companies. We can regard the products and services in each application as a word and calculate the frequency of each word. Then, the system background can select the words with the highest frequency as recommended products and services.

In this way, the system background has currently obtained two recommendation results: one is based on word vector comparison, and the other is based on industry information comparison. Then, the system background can combine these two results to form the final recommendation result. Specifically, a score is assigned to each product and service item, which is a weighted average of the item's ranking in the word vector comparison result and the ranking in the industry information comparison result. Then, we can choose the item with the highest score as the final recommendation result.

The following is the formula for this step: score ( item ) = α * rank _ vector ( item ) + β * rank _ industry ( item )

Among them, rank_vector(item) is the ranking of the item in the word vector comparison results, rank_industry(item) is the ranking of the item in the industry information comparison results, α and β are the weights of the two methods, and their sum is 1. In practical applications, the values of α and β can be determined by cross-validation to optimize the quality of the recommendation results. Therefore, α and β are adjustable parameters. In this case introduction, the weight adjustment method and process will not be elaborated. Adjustable parameters are customized according to the user end. The adjustable parameters can be arbitrarily selected from the options of user end preferences, user end habits, and user end risk tolerance.

If the company name, brand name or trademark name is not found in the description text entered by the user after being parsed by this system, it will not affect the subsequent category recommendation. This system can still parse the description text to generate industry information, and then generate recommended trademark categories and recommended product items in the trademark category based on the comparison of industry information.

Please continue to refer to Figures 22 and 23, which are flow charts of the embodiments of the present invention.

The user enters a description text through the input module of the electronic device, and further selects the target country for which the trademark recommendation is to be generated. The system first determines whether the description text language is the same as the official language of the target country. If they are different, the description text will be translated first. If they are the same, the subsequent process will be directly carried out. The description text will then be analyzed and extracted to generate keywords and industry-related texts. Industry-related texts are industry information.

The keyword part is judged and extracted for identifying characters, mainly by comparing with the non-identifying character database to confirm whether the keyword contains non-identifying characters. If so, the non-identifying characters are removed and only the identifying characters are retained. Specifically, the non-identifying character database can be established by, but not limited to, using a large language model database for training, and common characters or interjections, prepositions, etc. are judged as non-identifying.

On the other hand, the industry information (industry-related text) generated after parsing is used to generate industry categories corresponding to the descriptive text input by the user through a trained language generation model.

The category recommendation module compares the trademark category and product items according to the above-generated industry categories. Specifically, the comparison database will be different depending on the target country selected by the user. For example, if the user selects the United States as the target country, it will be compared with the official US trademark category database. If the user selects Taiwan, it will be compared with the official Taiwan trademark category database. The purpose is that the product items in the trademark categories of different countries may be different, and sometimes there may be differences. There are some product items in Taiwan that cannot be found in the same category in the United States. Therefore, it is very important to compare with the trademark category database of the target country so as not to generate recommended product items that do not actually exist in that country. This system will use adjustable parameters and, for example, the similarity comparison of word vectors to generate a comparison result of at least one recommended category and at least one recommended product item, including at least one trademark category recommendation information , to find similar trademark categories and product items, and achieve the effect of accurate recommendation. The adjustable parameters are customized according to the user end, and the adjustable parameters can be arbitrarily selected from the options of user end preferences, user end habits, and user end risk tolerance.

The comparison results are then arranged and filled into a modular format by the report generation unit to generate a trademark category recommendation report. The risk assessment module is used to compare the keywords extracted from the string with the trademark database of the target country and generate risk information. The category recommendation module also includes a knowledge graph unit, which is used to establish node matrix associations and dynamically learn the equivalent relationship between product items and trademark items in different countries according to the time evolution. The knowledge graph unit conversion includes at least the following steps: collect the relationship data between multi-national trademark product items and trademark projects and learn the dynamic model, calculate the best match, shortest path, and maximum flow through graph query and natural and abrasive reasoning, and optimize the equivalent relationship between product items according to the final decision result.

The report can be presented in the form of online images, downloadable documents, files that can be forwarded to communication software, files that can be shared to social networking software, electronic bulletin boards, or a combination thereof.

In FIG. 23, the difference from FIG. 22 is that after generating the target country's trademark category recommendation information, the risk assessment module further compares the trademark category recommendation information with previous cases through the search and comparison unit. The target country's trademark database searches for previous trademarks with similar identifying words. The search and comparison unit searches in the recommended category to see if there are any identifying words extracted from the keywords above. The search results are combined with the recommended categories and recommended product items for analysis. For example, if the recommended category is Category 9, the search comparison finds that there is a similar previous case in Category 9. At this time, the search comparison unit will further compare and analyze the recommended product items and the previous product items in detail. If the product items are completely non-overlapping, the similarity will be lower than the similarity of some product items overlapping. The search results and the identifying text are similarly calculated and arranged, and the similarity between the previous application category in the search results and the recommended information of at least one trademark category is calculated to generate an evaluation result.

Finally, the integrated analysis results are put into the modular format through the report generation unit to generate a risk assessment report.

Furthermore, the system has also established a Chiming trademark database. In the above comparison, even if the user and the previous case found by the search are in different trademark categories, if the previous case belongs to one of the Chiming trademark database, the search result will still show the previous case, and it will be prioritized and displayed as high risk.

Furthermore, after the category recommendation module parses the description text and extracts keywords and industry-related text to form industry information, the information extraction unit can proactively search and match companies with similar industry-related text descriptions in external network information based on the extracted industry-related text, and transmit the searched and matched company information to a knowledge graph comparison unit of the category recommendation module, and then perform trademark category comparison through the knowledge graph comparison unit, and generate at least one trademark category recommendation information, which at least includes: a recommended category and at least one recommended product item.

Please continue to refer to Figures 24 to 26, which are schematic diagrams showing the use scenarios of the system of the present invention.

As shown in Figure 24, in this usage scenario, the user is a general merchant, brand owner or trademark applicant. Through a vehicle such as a mobile phone, computer or other electronic device, the server is activated after connecting to the Internet and any descriptive text string is entered through the field of the web page. These strings are semantically analyzed through the server of the cloud part. Specifically, it is first determined whether the language of the string is the same as the official language of the target country. If not, it is translated through a multilingual translation model, and then the string is parsed, keywords are extracted and industry information is generated. Then, it is compared with the trademark classification database of the target country through the category item comparison unit, and a trademark category recommendation report is generated and fed back to the user's vehicle.

Furthermore, the previous trademark cases are searched in the recommended trademark categories to find out whether there are previous trademark cases in the recommended categories and the similarity between the trademarks and the product items. After the search results are combined with the recommended categories and the recommended product items, risk assessment information is generated, and finally a risk assessment report is generated and fed back to the user's vehicle.

Furthermore, the trademark category recommendation report may be in the form of an online image, a downloadable document file, a file format that can be forwarded to a communication software, a file format that can be shared to a social networking software, an electronic bulletin board, or a combination thereof.

It can also crawl information from local servers such as official trademark application databases of various countries or any online information based on the input string, search for trademark records applied for by other companies in similar industries, and obtain trademark categories and product items as recommended options.

The end user can provide this report to a professional firm or directly submit the application documents to the official agency for trademark application.

As shown in Figure 25, in this usage scenario, the user's role is a professional trademark practitioner or a law firm. The user uses a carrier, usually an electronic device such as a computer, to connect to the Internet, activate the server, and enter any descriptive text string through the webpage field to obtain a trademark category recommendation report or even a risk assessment report through the system of the present invention. The user can directly or again organize the report in a modular fixed format to generate a trademark proposal. After automatically applying the trademark practitioner or law firm's own format, it will be provided to the user's client, which is the brand owner, merchant or trademark applicant. After the client confirms, the application documents can be submitted to the official agency to complete the case processing.

The schematic diagram of the risk assessment report can be found in Figure 26. It should be noted that Figure 26 is only an example in the usage scenario and does not limit the display method and format of the report. As shown in Figure 26, in the risk assessment report, similar trademarks in the recommended category are listed, as well as their basic information such as trademark name/image, country, application date, applicant, category and product item, official number, risk level, etc. Through the display interface, users can clearly know the risk level of the trademark to be applied for in the corresponding trademark category. Furthermore, they can also switch The target country then displays risk assessment reports for different target countries. The recommended categories for different target countries are likely to be the same, but the product items will be slightly different, and the similar trademark cases will also be different. Through the semantic analysis, automatic translation and judgment functions of the system of the present invention, multiple reports for different target countries can be quickly generated, greatly reducing the time cost and the threshold of previous category analysis and retrieval.

As can be seen from Figures 24 and 25, whether the user is an end customer or an intermediate practitioner, the time cost of searching and recommending trademarks can be greatly shortened. Moreover, through the characteristics of the system of the present invention: inputting any descriptive text can be used to perform semantic analysis and text generation through a multilingual translation model and a text parsing unit, understand the descriptive text in a short time, and perform category comparison and similar arrangement to generate recommended categories and product items, which is even more effective in cross-border applications.

Finally, the technical features of the present invention and the technical effects it can achieve are summarized as follows:

First, the semantic analysis trademark category recommendation system of the present invention can solve the problem that when general users apply for trademarks online, they do not know which trademark category the goods they sell or the services they provide belong to, and do not know how to choose the appropriate category for their trademarks. The trademark category recommendation report of the present invention can provide users with a quick choice.

Secondly, the semantic analysis trademark category recommendation system of the present invention can solve the problem that users do not know whether their trademarks have similar cases before applying for trademarks. The risk assessment report of the present invention provides users with an immediate and easy-to-understand arrangement of search results.

Third, through the comprehensive report generated by the semantic analysis trademark category recommendation system of the present invention, combined with trademark category recommendation and previous case search, it is clearly marked which category items have more or fewer trademark previous cases, which can be clearly distinguished by color or presented in any other way that can be marked, providing users with clearer trademark category item selection suggestions.

It must be emphasized that the above detailed description is a specific description of the feasible embodiments of the present invention, but the embodiments are not intended to limit the patent scope of the present invention. Any equivalent implementation or modification that does not deviate from the technical spirit of the present invention should be included in the patent scope of this case.

100: electronic device 200: network 300: server 400: database 500: delivery server U: user

Claims (19)

A semantic analysis trademark category recommendation system is used to receive a string provided by a user terminal by operating an electronic device and at least one target country to be queried, and a processor of the electronic device is connected to a server through a network interface controller and executes an application program to calculate and generate a trademark category recommendation report. The system at least includes: an input module, which is used to receive the string input by the user, label the string, and send a string information; a category recommendation module, which is an operation model trained by a natural language model and a trademark classification table and detailed headings, and is used to convert the string into a The information is parsed into trademark category recommendation information, which further includes: a text parsing unit, which is used to parse the string information, extract keywords and industry information, and send them; a category item comparison unit, which is used to receive the keywords and industry information, compare them with the trademark category database of the target country, and calculate and generate at least one recommended category and at least one recommended product item; and a report generation unit, which is used to receive the recommended category, the recommended product item and the target country, and generate a trademark category recommendation report in a template format using the input language. The semantic analysis trademark category recommendation system as described in claim 1, wherein the system also includes a multilingual translation model, which is an operation model trained through large-scale language translation, for receiving the string information, determining whether the string input by the user is the same as the official language of the target country, and if different, translating the string into the official language of the target country, and sending all translated string information. A semantic analysis trademark category recommendation system as described in claim 1, wherein the string may further include trademark text, trademark description, business behavior, company name, stock code, product name, service name or a combination thereof. The semantic analysis trademark category recommendation system as described in claim 1, wherein the user terminal inputs the string by text input, voice input or video input. The semantic analysis trademark category recommendation system as described in claim 2, wherein the input module further includes an information acquisition unit, which is used to actively send the labeled string information to the network space to acquire relevant information, and can transmit the information to the multilingual translation model for translation. The semantic analysis trademark category recommendation system as described in claim 1, wherein the trademark category database described in the category item comparison unit further includes the trademark category database of the target country and the International Nice Trademark Category Database. The semantic analysis trademark category recommendation system as described in claim 1, wherein the system further comprises a risk assessment module for comparing the keywords extracted from the string with the trademark database of the target country and generating risk information. The semantic analysis trademark category recommendation system as described in claim 7, wherein the risk information can be further integrated into the trademark category recommendation report through the report generation unit. The semantic analysis trademark category recommendation system as described in claim 1, wherein the report can be presented in the form of an online image, a downloadable document file, a file format that can be forwarded to a communication software, a file format that can be shared to a social software, an electronic bulletin board, or a combination thereof. The semantic analysis trademark category recommendation system as described in claim 1, wherein the user terminal can be selected from any combination of brand owners, trademark applicants, trademark practitioners, and legal practitioners. As described in claim 1, the semantic analysis trademark category recommendation system, wherein the category recommendation module further includes a knowledge graph unit, which is used to establish node matrix associations, and according to the dynamic learning model of time evolution based on the equivalence relationship between product items and trademark items of various countries, the knowledge graph unit conversion at least includes the following steps: collecting multi-national trademark product items and trademark project relationship data and dynamic model learning, through graph query and natural and abrasive reasoning, inferring the best match, shortest path, maximum flow, and optimizing the equivalence relationship of product items according to the final decision result. A semantic analysis trademark category recommendation method is provided, wherein a user operates an electronic device to query at least one target country, and a processor of the electronic device is connected to a server through a network interface controller and executes an application program. The method of calculating the output includes at least the following steps: (1) the user enters a description text through an input module of the electronic device; (2) a category recommendation module parses the description text and extracts keywords and product information. (3) using a category recommendation module to compare the parsed industry-related text with the target country's trademark category database, outputting a comparison result according to an adjustable parameter, and generating at least one trademark category recommendation information including at least one recommended category and at least one recommended product item; and (4) using a report generation unit to generate a trademark category recommendation report using the trademark category recommendation information of step (3) in a templated format. The semantic analysis trademark category recommendation method as described in claim 12, wherein the step (2) may further include extracting identifying words from the extracted keywords through a text analysis unit, which can be divided into: step (21) the text analysis unit generates extracted identifying words by comparing with a non-identifying word database; and step (22) the text analysis unit links the language model of the category item comparison unit to convert industry-related words to generate industry categories. The semantic analysis trademark category recommendation method as described in claim 12, wherein step (1) further comprises: step (11) the user selects at least one target country for which the trademark category recommendation is to be generated; and step (3) further comprises: step (31) the category recommendation module compares the category in the trademark category database of the target country based on the country selected by the user. The semantic analysis trademark category recommendation method as described in claim 12, wherein after step (3), further comprising: Step (301) performing risk assessment on the identifying text through a risk assessment module, and using a search comparison unit to compare the identifying text in at least one trademark category recommendation information generated in step (3) with the trademark database of the target country to generate a search result; Step (302) the report generation unit The at least one trademark category recommendation information is integrated and analyzed with the search results, the search results and the identifying text are similarly calculated and arranged, and the similarity between the previous application category in the search results and the at least one trademark category recommendation information is calculated to generate an evaluation result; and in step (303), the report generation unit combines the at least one category recommendation category with the evaluation result to generate a risk assessment report in the template format. The semantic analysis trademark category recommendation method as described in claim 12, wherein after step (2), further comprises: step (21) an information extraction unit actively searches for companies with similar industry-related text descriptions in external network information based on the industry-related text extracted in step (2), and transmits the searched and matched company information to a knowledge graph comparison unit of the category recommendation module; and step (22) performs trademark category comparison through the knowledge graph comparison unit and generates at least one trademark category recommendation information. The semantic analysis trademark category recommendation method as described in claim 14, wherein after step (11), it further comprises: step (12) the input module determines whether the input string language is the same as the official language of the target country. If not, it will first be translated through a multilingual translation model to translate the string language into the official language of the target country. The semantic analysis trademark category recommendation method as described in claim 12, wherein the adjustable parameter is customized according to the user terminal, and the adjustable parameter can be arbitrarily selected from the options of user terminal preferences, user terminal habits, and user terminal's ability to bear risks. The semantic analysis trademark category recommendation method as described in claim 14, wherein the trademark category recommendation information at least includes: a recommended category and at least one recommended product item.