A trademark infringement risk multi-factor intelligent scoring method integrating official registration state activity ranking

Abstract

The present application belongs to the technical field of computer-aided analysis, and particularly relates to a trademark infringement risk multi-factor intelligent scoring method fusing official registration state activity ranking, comprising the following steps: step one, vector retrieval is performed on a target trademark to obtain a candidate hit; step two, hierarchical rearrangement is performed according to a state code-activity level mapping; step three, the same level is further sorted according to a similarity score; step four, NICE categories / state / goods descriptions are supplemented in real time from an official library; step five, five-factor large model scoring is performed; and step six, a risk score with official state and five factors is outputted. According to the present application, the hierarchical rearrangement of activity levels enables effective trademarks to participate in judgment preferentially, avoids interference of invalid trademarks, and makes the conclusion more suitable for legal practice; through hierarchical rearrangement and same-level sorting, effective trademarks are placed in front and invalid trademarks are placed in back, so that a large language model preferentially evaluates candidate trademarks with actual legal effectiveness, which is in line with the core principle of the Trademark Law that "effective registered trademarks are the basis for rights".

Description

Technical Field

[0001] This invention belongs to the field of computer-aided analysis technology, specifically relating to a multi-factor intelligent scoring method for trademark infringement risk that integrates the activity ranking of official registration status. Background Technology

[0002] With the rapid development of cross-border e-commerce, products need to be quickly assessed for trademark infringement risks in the target market's legal jurisdiction before being listed for sale. Traditional trademark infringement risk analysis mainly relies on manual searches and judgments by trademark agents or lawyers, which is inefficient, costly, and lacks standardized judgment criteria when dealing with massive amounts of products. In recent years, trademark similarity search technologies based on vector retrieval and artificial intelligence have gradually emerged, providing a new technological path for the automation of trademark infringement risk analysis.

[0003] Existing trademark infringement risk analysis technologies mainly have the following problems: First, current trademark similarity searches often return candidate trademarks based on vector or text similarity ranking, with infringement risk then assessed manually based on experience. This approach treats expired trademarks and valid trademarks equally, resulting in a large number of expired trademarks being mixed in with valid ones in the ranking, interfering with the accuracy of subsequent risk assessments. For example, an expired trademark may rank high in the similarity ranking, but it no longer possesses legal exclusivity and should not be considered a primary reference for infringement risk; conversely, a trademark with slightly lower similarity but still valid may constitute a genuine source of infringement risk. This equal treatment significantly increases the workload of manual review and is prone to misjudgment.

[0004] Second, some solutions attempt to directly use large language models to assess the risks of goods and candidate trademarks, but fail to differentiate the official registration status of candidate trademarks at the input stage, nor do they incorporate legal factors such as the overlap of NICE goods / service categories. If the input of a large language model assessment does not include official category and status information, it is prone to creating "illusions" and providing risk conclusions that are inconsistent with the facts. Furthermore, due to the lack of official legal elements, the model's output conclusions are difficult to trace and verify, and cannot serve as a defensible basis in legal practice.

[0005] Third, the existing scheme lacks a quantitative stratification mechanism for the activity level of official trademark registrations, failing to achieve interpretability and verifiability of risk scoring at the legal element level. The lack of a structured legal intermediary between search results and risk assessment results in a clear end-to-end black-box characteristic, making it difficult to meet the requirements of verifiable and probable risk conclusions in trademark law practice.

[0006] In view of this, the present invention provides a multi-factor intelligent scoring method for trademark infringement risk that integrates the ranking of official registration status activity to solve the above problems. Summary of the Invention

[0007] To achieve the above objectives, the present invention provides the following technical solution: a multi-factor intelligent scoring method for trademark infringement risk that integrates official registration status activity ranking, comprising the following steps: Step 1: Perform vector retrieval on the target trademark to obtain candidate matches (including similarity); that is, use the embedding model to map the target trademark into a vector, search for approximate matches in the trademark vector library, and obtain candidate similar trademarks and their similarity scores; Step 2: Perform hierarchical rearrangement based on status code → activity level mapping; that is, according to the official registration status code of the candidate trademarks, classify them into different activity levels such as valid, uncertain, and invalid according to the preset mapping relationship; Step 3: Sort by similarity score within the same level (without re-querying the vector library); that is, within the same activity level, sort by the original similarity score from high to low. This operation is a post-processing of the already obtained matches and does not require querying the vector library again. Step 4: Real-time updates of NICE category / status / product description from official databases; that is, real-time acquisition of the latest NICE classification information, registration status, and product / service descriptions of candidate trademarks from official trademark databases such as the USPTO. Step 5: Five-factor large model scoring: confusion / salience / category overlap / visual / matching; that is, submit the target trademark and the completed candidate trademark to the large language model, evaluate them from five dimensions: likelihood of confusion, trademark salience strength, NICE category overlap, visual similarity, and matching details, and output a structured risk score and reasoning. Step Six: Output a risk score with official status and five-factor basis; that is, output a risk score conclusion that includes official registration status, NICE category information, and five-factor assessment scores and basis, for subsequent review and evidence application.

[0008] As a preferred embodiment of the present invention, which integrates the activity ranking of official registration status for trademark infringement risk multi-factor intelligent scoring method, in step one, the input form of the target trademark includes trademark text, trademark image, or a combination thereof; for trademarks in text form, they are mapped to text vectors through a text embedding model and approximately searched in a trademark text vector library; for trademarks in image form, they are mapped to image vectors through a multimodal embedding model and approximately searched in a multimodal trademark vector library; the text vector library and the image vector library adopt a dual-index architecture.

[0009] As a multi-factor intelligent scoring method for trademark infringement risk that integrates the activity ranking of official registration status according to the present invention, preferably, in step one, after retrieving the set of candidate similar trademark hits, pre-filtering is performed based on the NICE category number, retaining only the candidate hits that are the same as or similar to the NICE category to which the target trademark belongs, and then proceeding to step two; the similarity refers to the NICE category number belonging to the same major category group or having a similar relationship of goods / services.

[0010] As a preferred embodiment of the present invention, which integrates the activity ranking of official registration status, the activity level in the "Official Registration Status Code → Activity Level" mapping table includes at least three levels: valid, uncertain, and invalid. The valid level takes precedence over the uncertain level, and the uncertain level takes precedence over the invalid level. The valid level includes status codes with official registration statuses of "registered valid" and "renewed valid". The invalid level includes status codes with official registration statuses of "expired and not renewed", "cancelled", and "invalid declaration". The uncertain level includes status codes with official registration statuses of "under review" and "objection".

[0011] As a multi-factor intelligent scoring method for trademark infringement risk that integrates official registration status activity ranking according to the present invention, preferably, the specific steps of the hierarchical reordering in step two are as follows: traverse each hit record in the candidate similar trademark hit set, read its official registration status code, and map it to the corresponding activity level according to the mapping table; establish a hierarchical queue according to the activity level from high to low, that is, the effective layer queue takes priority, the uncertain layer queue takes second place, and the ineffective layer queue takes last place.

[0012] As a preferred method of trademark infringement risk multi-factor intelligent scoring that integrates the activity ranking of official registration status in this invention, in step three, the similarity scores are ranked from high to low within the same activity level. If the similarity scores are the same, they are further ranked a second time according to the timeliness of the official registration status code, and candidates with more recent registration dates or renewal dates are ranked higher.

[0013] As a multi-factor intelligent scoring method for trademark infringement risk that integrates official registration status activity ranking of the present invention, preferably, the hierarchical reordering in step two and the ranking in step three are post-processing operations on the already obtained hits, without re-querying the vector library; the hierarchical reordering and ranking of the candidate similar trademark hit set retains the original vector similarity score, which serves as a reference input for the large language model evaluation in step five, so that the large language model can simultaneously refer to vector similarity and activity ranking information for comprehensive judgment.

[0014] As a multi-factor intelligent scoring method for trademark infringement risk that integrates the activity ranking of official registration status according to the present invention, preferably, in step four, the official trademark database is the USPTO trademark database, and the completion operation performs a precise query through the unique identifier of the candidate trademark; the elements to be completed include at least the NICE category number, NICE category title, detailed description of goods / services, and official registration status code; if the official registration status code of the candidate trademark cannot be found in the mapping table, it is classified into the uncertain level by default.

[0015] As a preferred embodiment of the present invention, which integrates the activity ranking of official registration status for trademark infringement risk multi-factor intelligent scoring method, in step five, the confusion probability factor assesses the overall probability of confusion between the target trademark and the candidate trademark in terms of pronunciation, shape, and meaning; the trademark distinctiveness strength factor determines the distinctiveness level based on the type of the candidate trademark, which includes coined trademarks, arbitrary trademarks, suggestive trademarks, and descriptive trademarks; the NICE category overlap factor calculates the degree of overlap between the NICE category to which the target product belongs and the NICE category registered by the candidate trademark; the visual similarity factor calculates the vector distance between the target trademark image and the candidate trademark image based on multimodal embedding vectors; and the matching detail factor assesses the degree of association between the target product and the goods approved for use by the candidate registered trademark in terms of sales channels, target consumers, and intended use.

[0016] As a multi-factor intelligent scoring method for trademark infringement risk that integrates official registration status activity ranking according to the present invention, preferably, in step five, the large language model returns the evaluation results according to a preset structured output format, including the comprehensive risk score, the scores of each factor, the reasons for the risk judgment, and the corresponding official legal basis citations; in step six, the risk scoring conclusion is verifiable structured data, the data structure of which includes candidate trademark identifiers, candidate trademark official registration status, candidate trademark NICE category information, comprehensive risk score, scores and comments of each of the five factors, and legal basis citations; the structured data is used to support subsequent manual review and legal evidence presentation.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention employs a hierarchical reordering of activity levels to prioritize valid trademarks in the assessment, avoiding interference from expired trademarks and resulting in conclusions that are more aligned with legal practice. In existing technologies, expired and valid trademarks are treated equally, leading to a large number of expired trademarks interfering with the accuracy of risk assessment. This invention, through hierarchical reordering and equal-level ranking, places valid trademarks first and expired trademarks last, enabling the large language model to prioritize the evaluation of candidate trademarks with actual legal effect, thus conforming to the core principle of trademark law that "rights are based on valid registered trademarks."

[0018] This invention employs hierarchical rearrangement and sorting as post-processing of existing matches, eliminating the need for secondary vector database queries and resulting in low computational overhead. Vector retrieval involves calculating similarity between high-dimensional vectors, which is computationally intensive and time-consuming. This invention performs vector retrieval only once in step one; subsequent hierarchical rearrangement and sorting are completed in memory based on the obtained matches, achieving a computational complexity of only O(n·logn), far lower than the overhead of vector retrieval. This makes it suitable for real-time risk analysis scenarios involving large volumes of goods.

[0019] This invention employs a five-factor framework and official category / status injection to make scoring interpretable and defensible, significantly reducing the risk of illusion in large language model assessments. This invention supplements elements such as the NICE category and registration status from the official trademark database in real time, ensuring that the input to the large language model has an official legal basis, reducing illusions from the outset. The five-factor assessment framework decomposes risk judgment into five independently verifiable legal dimensions, and the output conclusions are traceable to specific official status codes and category entries, overcoming the unverifiable shortcomings of black-box models and meeting the requirements of traceable and verifiable conclusions in legal practice. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart of a multi-factor intelligent scoring system for trademark infringement risk provided in this application embodiment. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 The present invention provides the following technical solution: a multi-factor intelligent scoring method for trademark infringement risk that integrates the ranking of official registration status activity, comprising the following steps.

[0023] Step 1: Perform vector retrieval on the target trademark to obtain candidate matches (including similarity).

[0024] Specifically, the system receives input information for the target trademark. The input form of the target trademark can be trademark text, trademark image, or a combination thereof. For the text-based trademark "OLA", the system maps it to a 768-dimensional text vector using a text embedding model, performs an approximate search in the Pinecone text vector library, uses cosine similarity measurement, sets a similarity threshold of 0.75, and retrieves a candidate hit set of 20 records with similarity scores greater than this threshold. Each hit record contains the candidate trademark identifier and its corresponding similarity score, with similarity scores ranging from 0.76 to 0.93. For image-based trademarks, a multimodal embedding model maps them to a 1408-dimensional image vector, and an approximate search is performed in a multimodal trademark vector library. The text vector library and the image vector library employ a dual-index architecture.

[0025] Furthermore, after obtaining the set of candidate similar trademark matches through vector retrieval, pre-filtering is performed based on the NICE category number. For example, if the target trademark "OLA" is intended for use in Class 25 (clothing), the pre-filtering only retains candidate matches with the NICE category number in Class 25 or those similar to Class 25 (such as Class 24 textiles), excluding candidates from other irrelevant categories, thereby reducing the number of invalid candidates and improving processing efficiency. "Similar" refers to NICE category numbers belonging to the same major category group or having a similar relationship in goods / services.

[0026] Step 2: Perform hierarchical rearrangement based on the status code → activity level mapping.

[0027] Registered (Registered is valid) efficient Renewed (Valid for renewal) efficient Published for Opposition (Publication under objection) uncertain Pending (Under review) uncertain Expired Failure Cancelled Failure Abandoned Failure The specific steps of the hierarchical reordering are as follows: traverse each hit record in the candidate similar trademark hit set, read its official registration status code, and map it to the corresponding activity level according to the mapping table; establish a hierarchical queue according to the activity level from high to low, that is, the effective layer queue takes priority, the uncertain layer queue takes second place, and the ineffective layer queue takes last place.

[0028] Step 3: Within the same level, sort by similarity score (without re-checking the vector library).

[0029] Specifically, within the same activity level, the candidates are sorted in descending order of their original similarity scores. The three hierarchical queues are then concatenated sequentially to obtain the rearranged set of candidate similar trademark matches. This hierarchical rearrangement and sorting is a post-processing operation on the matches already obtained; it does not require re-querying the vector database, resulting in extremely low computational overhead.

[0030] Taking the target trademark "OLA" as an example, its 20 candidate trademarks include several similar trademarks that have been cancelled. For instance, the "OLAY" trademark has expired in some categories, and its similarity score is 0.89, which is higher than the 0.82 of a certain valid trademark. According to the traditional similarity ranking, the expired trademark would be ranked higher; however, according to the method of this invention, after the activity level is hierarchically reordered, the valid trademark is prioritized at the top of the list, and the invalid trademark is ranked at the bottom of the list.

[0031] Furthermore, for candidate matches with the same similarity score within the same activity level, a secondary sort is performed based on the validity period of the official registration status code, placing candidates with more recent registration or renewal dates higher. The hierarchical rearrangement and sorted set of candidate similar trademark matches retains the original vector similarity scores, serving as reference input for the subsequent evaluation of the five language models.

[0032] Step 4: Complete the NICE category / status / product description in real time from the official database.

[0033] Specifically, the system continuously completes the NICE category number and title, goods / service description, and registration status code elements of the re-ranked candidate similar trademarks from the official trademark database. The official trademark database is the USPTO trademark database, and the completion operation uses the unique identifier of the candidate trademark (such as a serial number or registration number) for precise searching. For valid trademarks ranking high after the re-ranking, the system continuously completes their NICE category number ("025"), category title ("Clothing, footwear, headgear"), goods / service description ("Clothing, namedly shirts, pants, dresses; footwear; headgear, namedly hats, caps"), and registration status code ("live / registered") from the USPTO trademark database.

[0034] The elements to be completed must include at least: NICE category number, NICE category title, detailed description of the product / service, and official registration status code. This completion is performed in real-time to ensure that the obtained element information is the latest status in the official database. If the official registration status code of a candidate trademark cannot be found in the mapping table, it will be classified as an uncertain level by default.

[0035] Step 5: Five-factor large model scoring: confusion / significance / category overlap / visual / matching.

[0036] Specifically, the system submits the target trademark and the completed candidate similar trademarks together to the large language model, and evaluates them according to five factors: likelihood of confusion, trademark distinctiveness, overlap of NICE categories, visual similarity, and matching details, and outputs a structured risk score and reasons.

[0037] The system assembles the target trademark "OLA" and its product category "Class 25 Clothing" with the completed candidate trademark information (such as the "OLAY" trademark, Class 25, and validity status) into an evaluation request according to the preset prompt word template, and submits it to the large language model.

[0038] The evaluation dimensions of the five factors are as follows: Confusion Probability Factor: Assess the overall probability of confusion between the target trademark and candidate trademarks in terms of pronunciation, shape, and meaning. A 100-point scale is used for scoring, with higher scores indicating a higher probability of confusion.

[0039] Trademark distinctiveness strength factor: assesses the distinctiveness of candidate trademarks in their registered goods / services categories, and determines the distinctiveness level based on the type of candidate trademark, which includes coined trademarks, arbitrary trademarks, suggestive trademarks, and descriptive trademarks.

[0040] NICE Category Overlap Factor: Calculates the degree of overlap between the NICE category to which the target product belongs and the NICE category of the candidate trademark registration.

[0041] Visual similarity factor: The vector distance between the target trademark image and the candidate trademark image is calculated based on multimodal embedding vectors and converted into a similarity percentage.

[0042] Matching detail factor: Assess the degree of relevance between the target goods and the goods for which the candidate trademark is approved in terms of sales channels, target consumers, and intended use.

[0043] The large language model returns the evaluation results according to a pre-defined structured output format, including a comprehensive risk score, scores for each factor, reasons for the risk assessment, and corresponding official legal citations. An example prompt template is as follows: text Please conduct an infringement risk assessment for the following target trademarks and candidate trademarks, scoring each of the five factors (0-100 points each), and provide a comprehensive risk score (0-100 points) and the rationale for the assessment: Target trademark: [Target trademark name] Target product category: [NICE category number and title] Candidate Trademark: [Candidate Trademark Name] Candidate trademark registration category: NICE: [NICE category number and title] Official registration status of the candidate trademark: [Official status code] Evaluation factors: 1. Probability of confusion (overall probability of confusion regarding pronunciation, character form, and meaning); 2. Trademark distinctiveness (figurative / arbitrary / suggestive / descriptive); 3. Category overlap (complete / partial / no overlap); 4. Visual similarity (based on image features); 5. Matching details (degree of relevance in sales channels, target consumers, etc.); Please output the evaluation results in JSON format.

[0044] In practical use, after submitting the target trademark "OLA" and the completed candidate trademark "OLAY" (Class 25, valid status) together, the big language model outputs: a comprehensive risk score of 87 (high risk); a confusion probability factor of 92, because "OLA" and "OLAY" are both disyllabic in pronunciation, have the same first and last letters, and differ in shape by only one letter, resulting in a high overall possibility of confusion; a trademark distinctiveness strength factor of 85, because "OLAY" is an arbitrary trademark and has strong distinctiveness in Class 25; a NICE category overlap factor of 95, because both are in Class 25 and completely overlap; a visual similarity factor of 78, because the trademark images differ in font style but have similar letter composition; a matching detail factor of 80, because both products are mass-market clothing, and their sales channels and target consumers highly overlap; and a risk assessment reason of "high possibility of confusion with a valid registered trademark in Class 25 in appearance and pronunciation, the candidate trademark is valid, and the NICE category completely overlaps."

[0045] Step Six: Output a risk score with official status and five-factor basis.

[0046] Specifically, the system outputs a risk score conclusion with official status, category, and five-factor criteria. The risk score conclusion is verifiable structured data, and its data structure includes the candidate trademark identifier, the candidate trademark's official registration status, the candidate trademark's NICE category information, a comprehensive risk score, individual scores and comments for each of the five factors, and legal citations. An example data structure is as follows: json { "candidate_mark":"OLAY", "candidate_serial_number":"12345678", "official_status":"live / registered", "nice_classes":["025"], "nice_description":"Clothing,footwear,headgear", "comprehensive_risk_score":87, "risk_level":"high", "factor_scores":{ "confusion_likelihood":{"score":92,"comment":"The pronunciation and character shape are highly similar"}, "distinctiveness":{"score":85,"comment":"A trademark with strong distinctiveness"}, "class_overlap":{"score":95,"comment":"Class 25: Complete overlap"}, "visual_similarity":{"score":78,"comment":"The letter composition is similar, and the font differences are minor"}, "matching_details":{"score":80,"comment":"Sales channels and target consumers highly overlap"} }, "reasoning": "There is a high possibility of confusion with the valid registered trademark OLAY in Class 25 in terms of appearance and pronunciation." "legal_basis":["LanhamAct§32(15U.SC§1114)"] }

[0047] The structured data is used to support subsequent manual review and legal evidence submission.

[0048] The working principle of this invention in specific use is as follows: This invention, after obtaining candidate similar trademarks through vector retrieval, no longer relies solely on similarity ranking. Instead, it introduces a hierarchical re-ranking mechanism based on official registration status activity. Through a mapping table of "official registration status code → activity level," valid trademarks are prioritized over invalid trademarks. This allows the large language model to evaluate candidate trademarks with actual legal effect first, avoiding interference from invalid trademarks in risk assessment. Simultaneously, before evaluation, the NICE category and status information are supplemented in real-time from the official database, ensuring that the input to the large language model contains sufficient legal elements. The final output is an interpretable and defensible structured risk scoring conclusion. Activity re-ranking, as a post-processing step for identified similarities, eliminates the need for secondary vector library queries, ensuring computational efficiency. Vector similarity scores are retained as reference input for the large language model, achieving a dual-track fusion of mathematical similarity judgment and legal semantic judgment.

[0049] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-factor intelligent scoring method for trademark infringement risk that integrates official registration status activity ranking, characterized in that, Includes the following steps: Step 1: Perform vector retrieval on the target trademark to obtain candidate matches (including similarity); that is, use the embedding model to map the target trademark into a vector, search for approximate matches in the trademark vector library, and obtain candidate similar trademarks and their similarity scores; Step 2: Perform hierarchical rearrangement based on status code → activity level mapping; that is, according to the official registration status code of the candidate trademarks, classify them into different activity levels such as valid, uncertain, and invalid according to the preset mapping relationship; Step 3: Sort by similarity score within the same level (without re-querying the vector library); that is, within the same activity level, sort by the original similarity score from high to low. This operation is a post-processing of the already obtained matches and does not require querying the vector library again. Step 4: Real-time updates of NICE category / status / product description from official databases; that is, real-time acquisition of the latest NICE classification information, registration status, and product / service descriptions of candidate trademarks from official trademark databases such as the USPTO. Step 5: Five-factor large model scoring: confusion / salience / category overlap / visual / matching; that is, submit the target trademark and the completed candidate trademark to the large language model, evaluate them from five dimensions: likelihood of confusion, trademark salience strength, NICE category overlap, visual similarity, and matching details, and output a structured risk score and reasoning. Step Six: Output a risk score with official status and five-factor basis; that is, output a risk score conclusion that includes official registration status, NICE category information, and five-factor assessment scores and basis, for subsequent review and evidence application.

2. The multi-factor intelligent scoring method for trademark infringement risk based on official registration status activity ranking as described in claim 1, characterized in that: In step one, the input form of the target trademark includes trademark text, trademark image, or a combination thereof; for trademarks in text form, they are mapped to text vectors through a text embedding model and approximately searched in a trademark text vector library; for trademarks in image form, they are mapped to image vectors through a multimodal embedding model and approximately searched in a multimodal trademark vector library; the text vector library and the image vector library adopt a dual-index architecture.

3. The multi-factor intelligent scoring method for trademark infringement risk based on official registration status activity ranking as described in claim 1, characterized in that: In step one, after obtaining the set of candidate similar trademarks, pre-filtering is performed based on the NICE category number, retaining only candidate matches or similar to the NICE category to which the target trademark belongs, and then proceeding to step two; the similarity refers to the NICE category number belonging to the same major category group or having a similar relationship between goods / services.

4. The multi-factor intelligent scoring method for trademark infringement risk based on official registration status activity ranking as described in claim 1, characterized in that: In step two, the "Official Registration Status Code → Activity Level" mapping table includes at least three levels: valid, uncertain, and invalid. Valid levels take precedence over uncertain levels, and uncertain levels take precedence over invalid levels. Valid levels include official registration status codes of "Registered Valid" or "Renewed Valid". Invalid levels include official registration status codes of "Expired and Not Renewed", "Cancelled", or "Declared Invalid". Uncertain levels include official registration status codes of "Under Review" or "Under Objection".

5. The multi-factor intelligent scoring method for trademark infringement risk based on official registration status activity ranking as described in claim 1, characterized in that: In step two, the specific steps of hierarchical rearrangement are as follows: traverse each hit record in the candidate similar trademark hit set, read its official registration status code, and map it to the corresponding activity level according to the mapping table; establish a hierarchical queue according to the activity level from high to low, that is, the effective layer queue takes priority, the uncertain layer queue takes second place, and the ineffective layer queue takes last place.

6. The multi-factor intelligent scoring method for trademark infringement risk based on official registration status activity ranking as described in claim 1, characterized in that: In step three, candidates are sorted from high to low based on similarity scores within the same activity level. If the similarity scores are the same, they are further sorted based on the validity of the official registration status code, with candidates whose registration date or renewal date is more recent being ranked higher.

7. The multi-factor intelligent scoring method for trademark infringement risk based on official registration status activity ranking as described in claim 1, characterized in that: The hierarchical rearrangement in step two and the sorting in step three are post-processing operations on the already obtained hits, without re-querying the vector library; the hierarchical rearrangement and sorted candidate similar trademark hit sets retain the original vector similarity scores, which are used as reference inputs for the large language model evaluation in step five, so that the large language model can make a comprehensive judgment by simultaneously referring to vector similarity and activity ranking information.

8. The multi-factor intelligent scoring method for trademark infringement risk based on official registration status activity ranking as described in claim 1, characterized in that: In step four, the official trademark database is the USPTO trademark database. The completion operation uses the unique identifier of the candidate trademark for precise query. The elements to be completed include at least the NICE category number, NICE category title, detailed description of goods / services, and official registration status code. If the official registration status code of the candidate trademark cannot be found in the mapping table, it will be classified into the uncertain level by default.

9. The multi-factor intelligent scoring method for trademark infringement risk based on official registration status activity ranking as described in claim 1, characterized in that: In step five, the confusion probability factor assesses the overall probability of confusion between the target trademark and the candidate trademark in terms of pronunciation, shape, and meaning; the trademark distinctiveness strength factor determines the distinctiveness level based on the type of the candidate trademark, which includes coined trademarks, arbitrary trademarks, suggestive trademarks, and descriptive trademarks; and the NICE category overlap factor calculates the degree of overlap between the NICE category to which the target product belongs and the NICE category registered by the candidate trademark. The visual similarity factor is calculated based on multimodal embedding vectors to determine the vector distance between the target trademark image and the candidate trademark image; the matching detail factor assesses the degree of correlation between the target product and the product for which the candidate trademark is registered in terms of sales channels, target consumers, and intended use.

10. The multi-factor intelligent scoring method for trademark infringement risk based on official registration status activity ranking as described in claim 5, characterized in that: In step five, the large language model returns the evaluation results according to a pre-set structured output format, including a comprehensive risk score, scores for each factor, reasons for risk assessment, and corresponding official legal basis citations. In step six, the risk score conclusion is verifiable structured data, the data structure of which includes candidate trademark identifiers, official registration status of candidate trademarks, NICE category information of candidate trademarks, comprehensive risk score, scores and comments for each of the five factors, and legal basis citations. The structured data is used to support subsequent manual review and legal evidence presentation.

Images