IPO: Check if you could register your trade mark tool
A tool that uses AI to improve the quality of trade mark applications by providing a pre-application search function.
1. Summary
1 - Name
Check if you could register your trade mark tool
2 - Description
A trade mark is a sign which can distinguish the trade origin of goods and/or services from those of competitors. They are considered acceptable if distinctive or, in other words, can be recognized as a sign that differentiates the origin of goods or services from those of other sources. The ‘Check if you could register your trade mark’ tool improves the quality of trade mark applications by providing a pre-application search function to aid evaluation of similar existing trademarks. This tool enables a simple means for users to select goods and services from a list of pre-approved terms using AI-powered automated search of the register to identify similar marks.
3 - Website URL
https://trademarks.ipo.gov.uk/ipo-improve
4 - Contact email
Tier 2 - Owner and Responsibility
1.1 - Organisation or department
Intellectual Property Office
1.2 - Team
Argo
1.3 - Senior responsible owner
Divisional Director: Secure Trade Marks and Designs
1.4 - Third party involvement
Yes
1.4.1 - Third party
Deloitte
1.4.2 - Companies House Number
OC303675
1.4.3 - Third party role
A collaborative and multi-skilled team from the IPO alongside external partners Deloitte was formed to co-develop our tailored solution. This allowed us to build our own cutting-edge tool to improve customer experience.
1.4.4 - Procurement procedure type
G-Cloud framework
1.4.5 - Third party data access terms
N/A
Tier 2 - Description and Rationale
2.1 - Detailed description
The ‘Check if you could register your trade mark’ tool enables users to evaluate if their application for a trade mark will be successful prior to applying. The tool provides a user with relevant search results regarding their trade mark application without the requirement of understanding Vienna/Nice systems (which are international systems for classifying the figurative elements of trademarks). This tool utilises machine-learning to ensure that the tool performance can be continuously improved and ensures that it requires minimal time/financial investment from the user. This allows the user to focus their limited activity on uploading marks for a selection of goods/services (This tool does not require personal details and no fee is charged for the checking service) and delivers relevant information quickly inside a short user journey.
2.2 - Benefits
The key benefits of this tool are the improved quality of trade mark applications filed and an improved user journey for those filing their applications.
2.3 - Previous process
Historically, trade mark applications were submitted without meeting essential criteria, resulting in automatic rejection.
2.4 - Alternatives considered
N/A
Tier 2 - Deployment Context
3.1 - Integration into broader operational process
Users and organisations will first use this tool to check to see if a trademark is already in place prior to making a full application to the Intellectual Property Office. To undertake this task users and organisations must submit a smaller subset of information that will help identify the goods and services that they want to protect their trade mark for, check if there are any trade marks that are similar that could cause conflict and identify whether aspects of the trade mark might not be appropriate such as, offensive words or protected symbols e.g. a crown or crest. Depending on the outcome of this checking tool, the user or organisation may then decide to make a full application to the IPO or instead reconsider their signs, words, logos, slogans, or designs and make further changes. When a full application is made to the IPO this application is reviewed and a decision is made by a human.
3.2 - Human review
Monthly performance stats are reviewed by AI engineers to ensure system remains performant. In the event of a downturn in expected effectiveness of the tool, analysis would be conducted and the model retrained if appropriate.
3.3 - Frequency and scale of usage
Check if you could register your trade mark receives an average 3000 visits per month. With a monthly average of 14,400 applications this service currently supports approximately 20% of filings.
3.4 - Required training
No training is required by those using the tool, users are guided through key requirements before submission. This includes automated checks and tailored advice to help applicants assess the viability of their trade mark.
3.5 - Appeals and review
Some users may challenge or misunderstand the outcome provided by the tool. It’s important to note that the tool offers indicative guidance only and does not prevent users from submitting a formal trade mark application. The tool is designed to help users assess the likelihood of success based on basic criteria, but it does not provide a definitive decision. All applications must be submitted in full to undergo assessment and examination by the IPO. Users seeking further clarification or advice are directed to official resources at: https://www.gov.uk/defend-your-intellectual-property/get-help-and-advice
Tier 2 - Tool Specification
4.1.1 - System architecture
The tool searches trade marks for any that may match the user’s supplied candidate mark. The tool uses: - IPO’s Database to check for similar text in trade marks - Word embeddings with a vector database to search for most relevant “approved goods and service” terms; also used for most similar “Vienna image descriptors” (as required) - If a user describes their image, an additional set of image matching results are presented that are a filtered subset of the primary image results (via Vienna image descriptors) - Image embeddings with a vector database to search for similar image trade marks - Text also scanned for any potentially offensive terms - Data is updated nightly from the record of published trademarks
4.1.2 - System-level input
Current “live” trademarks are ingested nightly; the user enters their proposed text, image and goods and services for the trade mark.
4.1.3 - System-level output
Tool reports similar trademarks, so a user is aware of trade marks that could potentially raise an objection to their application. Potentially offensive terms are also flagged.
4.1.4 - Maintenance
Models are monitored monthly for performance degradation.
4.1.5 - Models
Word embeddings use FastText; image embeddings use Google Big Vision model. Database text similarity uses manually defined rules for approximate matching. Additional model provided by Deloitte to determine if an image is abstract or describable, used to determine if asking a “user to describe their mark” is worthwhile.
Tier 2 - Model Specification: FastText; self-hosted (1)
4.2.1. - Model name
FastText; self-hosted
4.2.2 - Model version
Pre-trained: v0.9.2
4.2.3 - Model task
Generate word embeddings from text, where embeddings reflect semantic similarity between words
4.2.4 - Model input
Text string
4.2.5 - Model output
300 dimensional embedding
4.2.6 - Model architecture
FastText is a library for learning word representations and performing text classification. It is designed to be fast and efficient, working on standard hardware, and can be used for tasks like data compression, sentence classification, and as a feature for other models. https://fasttext.cc/
4.2.7 - Model performance
A dataset of 409 goods and service terms with corresponding approved goods and service terms. A ~50-50 split between terms that are a correct match and terms that are not.
Both terms (entered and related approved) are embedded, the cosine similarity is used to determine semantic similarity. This score is compared with the label (correct match Y/N) and the Spearman Rank Coefficient is calculated. This value was used to validate the suitability of the embedder.
4.2.8 - Datasets and their purposes
Dataset of 409 goods and service terms with potential approved term
Tier 2 - Model Specification: Google Big Vision; self-hosted (2)
4.2.1. - Model name
Google Big Vision; self-hosted
4.2.2 - Model version
ViT-B-16-SigLIP
4.2.3 - Model task
Generate embedding given an image
4.2.4 - Model input
RGB bitmap
4.2.5 - Model output
Embedding representing the image
4.2.6 - Model architecture
Ref: https://github.com/google-research/big_vision Ref: https://huggingface.co/timm/ViT-B-16-SigLIP
Vision Transformer.
Weights released in the public domain by Google, ported to PyTorch by timm library (now used by Hugging Face).
4.2.7 - Model performance
Trade marks have “earlier marks” attached to a trade mark application, which indicate what marks (if any) a trade mark examiner is concerned could raise an objection.
The model is evaluated by assigning an embedding to each trade mark image; embedding similarity then used to determine image similarity. The model is scored on how many “earlier marks” appear in the top N results, indicating how close the model can get to an examiner’s opinion. We use Hit Rate @ N, as this represents how often the model would have shown an “earlier mark” to the end customer. Such as HR@5 reflects how often an earlier mark appears in the first results page, and HR@50 how often for the second page.
4.2.8 - Datasets and their purposes
~3 million trade marks were used. An 80-20 test/train split was created, with 20% used to evaluate pre-trained models before using in production.
Of the entire trade mark dataset, ~1.1 million are image marks of which 20% (as per split) were used to evaluate the image embeddings.
Tier 2 - Model Specification: Deloitte’s abstract classifier; self-hosted (3)
4.2.1. - Model name
Deloitte’s abstract classifier; self-hosted
4.2.2 - Model version
Supplied pre-trained by Deloitte
4.2.3 - Model task
Given an image, is it describable or abstract? It is used to avoid asking a user to describe an indescribable image, as this wouldn’t result in a useful Vienna code filter for the optional 2nd set of image results presented to the user.
4.2.4 - Model input
RGB bitmap
4.2.5 - Model output
Boolean
4.2.6 - Model architecture
Convolutional Neural Network (CNN)
4.2.7 - Model performance
Set of abstract and describable images; accuracy performance used to confirm model suitability.
4.2.8 - Datasets and their purposes
Pre-trained; validated using a set of abstract and describable images.
2.4.3. Development Data
4.3.1 - Development data description
Historical published trade marks were used, where “earlier marks” cited by examiners were used to validate the responses from the models.
4.3.2 - Data modality
Text, image
4.3.3 - Data quantities
~3 million trade marks were used. An 80-20 test/train split was created, with 20% used to evaluate pre-trained models before using in production.
4.3.4 - Sensitive attributes
N/A; IPO do not use customer personal data to train models.
4.3.5 - Data completeness and representativeness
N/A; We use all available published trademarks, no missing datasets.
4.3.6 - Data cleaning
N/A
4.3.7 - Data collection
Trademarks are published in the public domain.
4.3.8 - Data access and storage
Data is stored with an automatic retention cycle to remove data after its operational.
4.3.9 - Data sharing agreements
N/A
Tier 2 - Operational Data Specification
4.4.1 - Data sources
The system is updated nightly with the latest published trademarks; users enter their candidate trademarks when using the service.
4.4.2 - Sensitive attributes
None - Users access the service anonymously, no IP data or other identifiers are stored.
4.4.3 - Data processing methods
N/A
4.4.4 - Data access and storage
Operational data, including logs and telemetry, may be collected and stored within secure, government-managed cloud environments. Access to this data is governed by internal policies and is restricted to authorised programme teams and stakeholders, with permissions managed through defined roles. Data retention and deletion practices follow organisational and regulatory requirements, ensuring that information is only kept for as long as necessary. Security and privacy measures include encryption, access controls, and regular reviews to safeguard sensitive information. Responsibility for data management is assigned to designated data owners within the programme.
4.4.5 - Data sharing agreements
N/A
Tier 2 - Risks, Mitigations and Impact Assessments
5.1 - Impact assessments
A DPIA screening questionnaire was completed which established that no personally identifiable information would be processed in the Pre-Apply tool.
5.2 - Risks and mitigations
No significant risks have been identified in relation to the use of the tool. Security Risks are minimal. The service does not collect or store personally identifiable information or personal data. It operates without authentication, requiring no login credentials or passwords.
Trademark databases, legal standards, and user expectations evolve over time. Without regular updates, the tool may present outdated or incomplete information, potentially leading to inaccurate guidance or missed risks. To aid mitigation nightly ingest schedules ensure the tool is refreshed with the latest trademark data and relevant regulatory updates. Monthly performance monitoring detects model degradation and triggers retraining or recalibration as needed. Version control and change logs provide transparency and traceability of updates to both data and algorithms.
Perception Risk is minor but exists around user interpretation of the tool’s results. The tool provides indicative guidance only and does not determine application outcomes. Users may mistakenly believe a negative result prevents them from applying. However, all trade mark applications are subject to full assessment and examination upon submission. To help mitigate this clear messaging is provided throughout the service to reinforce its advisory nature. Users seeking further support are directed to official guidance and help resources: https://www.gov.uk/defend-your-intellectual-property/get-help-and-advice.