
Automated patent classification is the use of machine learning to predict which technology category a patent application belongs to, a job that patent examiners otherwise do by hand for every application they receive. In the UK Intellectual Property Office feasibility study on AI patent prior art search, researchers trained classifiers on patents that examiners had already labelled, then measured how well different algorithms could assign International Patent Classification (IPC) subclasses to documents they had never seen. This chapter explains why classification is such a natural fit for machine learning, how the experiment was designed, and what it means for anyone who searches patents today.
Earlier chapters looked at how AI can help examiners build and expand a search query. Document classification tackles a different step in the same workflow: deciding where in the technology landscape an application sits before the detailed search even begins.
Why Patent Classification Is a Natural Fit for Machine Learning
To see why this task suits machine learning so well, it helps to look at the system patents already use. The International Patent Classification, administered by WIPO, is a hierarchical scheme that the study describes as having roughly 650 subclasses (Makarov, 2004). Its purpose is to classify patents in one consistent, standardised way across patent offices around the world.
Every patent receives at least one classification code that identifies the main technical subject of the invention. Further codes can be added to refine the picture and place the document more precisely. If you would like a full walkthrough of sections, classes, subclasses and groups, our guide to the International Patent Classification (IPC) covers the structure in detail, so it is not repeated here.
The important point for the study is that this work is manual. For each incoming application, an examiner reads the document and assigns the appropriate code by following established classification guidelines. That takes time and expertise, and it has to be repeated for every single filing.
A Ready-Made Training Set of Labelled Patents
This manual routine turned out to be a real advantage for the research team. Because every previously filed patent already carries the code an examiner gave it, the patent archive is in effect a very large collection of labelled examples.
Labelled data is exactly what supervised machine learning needs, and in most fields it is slow and expensive to produce. In the patent world it already exists as a by-product of everyday office work.
How Supervised Learning Predicts a Patent's Class
Supervised learning starts with training data in which each document has a known class label. The algorithm studies the relationship between features of the text and the labels attached to them, then generalises those relationships into a classification model. Depending on the method, that model may take the form of a mathematical function, a decision tree or a probabilistic model.
Once trained, the model can be given a new document and asked which class it most likely belongs to. In this scenario the new document is an incoming patent application. It is compared implicitly against the large body of existing patents, because the patterns found in those patents are already built into the model.
Choosing an Algorithm: The "No Free Lunch" Theorem
With the general approach settled, the researchers had to decide which supervised learning algorithm to use. They referred to a well-known result in machine learning theory, the "no free lunch" theorem (Wolpert, 1996). It states that when performance is averaged over every possible problem, any two learning algorithms are equivalent.
In plain terms, no algorithm is the best at everything. The right choice depends on how a method performs on the specific problem and on the properties of the data involved, not on its reputation or popularity. For that reason the team did not pick a favourite in advance. They compared a broad set of candidates:
- Support vector machines with a radial basis function kernel
- Decision tree induction
- Random forest
- AdaBoost
- Nearest neighbours
- Multilayer perceptron, a type of neural network
- Gaussian naive Bayes
- Bernoulli naive Bayes
Fair Testing With Ten-Fold Cross-Validation
Comparing algorithms fairly requires care. If a model is tested on the same documents it learned from, the score will look better than it really is. The researchers therefore used cross-validation, an established technique that estimates how an algorithm performs on unseen data and avoids that optimistic bias.
This matters most for complex, data-hungry methods. Deep learning is the classic example of an approach prone to overfitting, which means memorising the quirks of the training data instead of learning patterns that carry over to new examples.
In ten-fold cross-validation the data is divided into ten parts. The model is trained on nine of them and tested on the remaining one, and the process is repeated until every part has served as the test set. The averaged result gives a far more honest view of how the algorithm would behave on real incoming applications.
Classification is one of the signals that helps a search system narrow millions of documents down to the relevant few. To see what that looks like for your own idea, you can try a free AI-powered novelty search and review the closest prior art it identifies.
One Binary Classifier for Each IPC Subclass
The proposed design did not rely on a single model that chooses among hundreds of categories at once. Instead, a separate binary classifier was trained for each IPC subclass. Each one answers a simple yes or no question: does this patent belong to this subclass?
For a given subclass, patents that truly belong to it serve as positive training examples, while patents from all other subclasses serve as negative examples. Breaking one large problem into many small ones makes each classifier easier to train and evaluate. It also fits the way the IPC works, since one patent can carry more than one code.
Three Technology Domains, Three Dataset Sizes
To make the test demanding, the researchers chose two relatively similar subclasses within each of the study's three validation domains. Telling closely related categories apart is much harder than separating obviously different ones, so the results say more about real-world performance.
IPC subclasses also differ greatly in size, so the pairs were chosen to cover a small, a medium and a large collection. This allowed the team to observe how the amount of training data affects classification performance.
| Domain | IPC subclasses compared | Approximate size |
|---|---|---|
| Civil engineering | E03D (water-closets and urinals with flushing devices, flushing valves) and E03F (sewers, cesspools) | About 900 patents (small) |
| Transport | B62J (cycle saddles, seats and accessories) and B62K (cycles, cycle frames, steering devices, rider-operated controls, axle suspensions, sidecars) | About 3,000 patents (medium) |
| Computer technology | G06K (recognition and presentation of data, record carriers) and G06T (image data processing or generation) | About 20,000 patents (large) |
The computer technology pair shows why similar subclasses are a genuine challenge. Both G06K and G06T deal with data and images, so the vocabulary of the two groups overlaps heavily and a classifier cannot rely on a few obvious keywords.
Bag-of-Words Versus Word Embeddings
For each domain pair, two ways of representing the documents were tested. The first is the simple bag-of-words model, which treats a document as a collection of the words it contains. The second is a richer, domain-specific word embeddings model built with word2vec, which places words with related meanings close together in a vector space. Both are explained in the chapter on feature extraction from patent text.
Combining eight algorithms, three domains of different sizes and two representation models produces a detailed grid of results. The full report presents this comparison in a dedicated appendix, showing side by side which algorithms handled which classification challenges most effectively.
What Automated Classification Means for Patent Searchers
Classification is not only an administrative step. Codes help decide which examiner or technical unit handles an application, and they are one of the most reliable ways to narrow a prior art search, because they group inventions by technology and not by the words an author happened to choose. A system that can suggest the right subclass quickly helps in three ways:
- Speed. A suggested classification gives the examiner or searcher a starting point within seconds.
- Consistency. The same text receives the same suggestion, whoever submits it and whenever they do so.
- Better search scope. Knowing the likely subclasses early helps focus a search on the right part of the patent landscape.
The idea has moved well beyond research. WIPO offers a public tool called IPCCAT that takes a short text, such as an abstract, and proposes the most likely IPC symbols together with a confidence level. Tools like this are assistants. The official classification of an application remains an examiner's decision, and the proof of concept in the study was likewise designed to suggest classifications to examiners, not to replace their judgement.
For inventors and applicants, the practical lesson is that classification and keyword searching work best together. A keyword search can miss a document that uses unexpected wording, while a classification search can miss a document filed under a neighbouring code. You can read more about the different search types and when to use each one in our patent information search guide.
Classification Built Into Every WOIPS Search
WOIPS FeatureThis chapter ends with a simple lesson: classification and keyword searching work best together. WOIPS applies that lesson automatically.
WOIPS's workflow-engineered AI search process reads your invention description and identifies the IPC and CPC classes it most likely belongs to. These classes are stored and searched at subclass level, the third level of the hierarchy (for example C12P), so every group and subgroup beneath them is included and no relevant subgroup is missed.
WOIPS then combines those classes with your keywords in its search queries, so documents filed under the right class are found even when they use unexpected wording. You do not need to learn the classification scheme yourself. See how WOIPS combines classification and keywords in a single search.
For the overall results of the study, including how the combined system performed with real examiners, see can AI help search patent prior art?
Frequently Asked Questions
What is automated patent classification?
Automated patent classification uses machine learning to predict the technology category of a patent document, for example its IPC subclass. The model learns from patents that human examiners have already classified and then applies those patterns to new applications.
What does the "no free lunch" theorem mean for patent classification?
The theorem says that no learning algorithm is best across all possible problems. For patent classification this means the algorithm has to be chosen by testing candidates on real patent data, not by relying on reputation.
What is overfitting, and how does cross-validation help?
Overfitting happens when a model memorises its training data and then performs poorly on new documents. Ten-fold cross-validation repeatedly trains on part of the data and tests on the rest, which gives a realistic estimate of performance on unseen patents.
Can AI replace patent examiners in classifying applications?
No. In the study and in publicly available tools, AI proposes likely classifications and a human expert makes the final decision. The technology speeds up and standardises the work, but professional judgement stays in the loop.
Do I need to know IPC codes to run a prior art search?
Not necessarily. Classification codes improve search quality, but AI-powered search tools can analyse a plain-language description of an invention and find relevant documents for you.
This article is adapted from: UK Intellectual Property Office, AI-assisted patent prior art searching – feasibility study, April 2020, ISBN 978-1-910790-80-9, © Crown Copyright 2020, licensed under the Open Government Licence v3.0.
