
Patent similarity search is the use of AI to measure how close one patent document is to another, so that the most relevant prior art rises to the top of a long result list instead of staying buried in it. In the UK Intellectual Property Office feasibility study on AI patent prior art search, researchers at Cardiff University paired it with a second technique, topic modelling, which groups a large set of retrieved patents into themes that an examiner can browse. Together, the two methods answer a very practical question: once a search has returned dozens or even hundreds of candidate documents, how do you make sense of them all?
The previous chapter on automated patent classification showed how AI can assign documents to IPC subclasses. Topic modelling and document similarity pick up where classification stops, inside a subclass that is still far too large to read from start to finish.
Searching and Browsing: Two Ways to Explore Prior Art
There are two fundamentally different ways to navigate a large volume of text, and each serves a different purpose:
- Searching suits someone who already has a fairly specific idea of what they are looking for and wants to reach it directly.
- Browsing suits someone who is exploring more openly, hoping to discover information they did not know to look for in the first place.
For patent work, this distinction matters a great deal. Browsing supports what the researchers call "opportunistic exploration" of prior art, which is especially useful when clear search terms are hard to define at the start. A new technology often has no settled vocabulary yet, and an inventor may describe an idea in words that no earlier patent happens to use.
Browsing only works, however, if the documents have first been sorted into meaningful groups.
Why IPC Classes Alone Are Too Broad for Browsing
The International Patent Classification is the obvious candidate for those groups, because every patent already carries IPC codes. The practical problem is size. Each IPC category tends to be quite broad, so manually inspecting every document inside one of them is simply not realistic for a human examiner.
The researchers wanted finer-grained browsing inside an individual IPC subclass, which means discovering smaller themes automatically, since nobody has labelled them in advance.
How LDA Topic Modelling Groups Patents by Theme
To find those themes, the team used latent Dirichlet allocation, usually shortened to LDA (Blei et al., 2003). LDA is designed to discover abstract topics hidden within a collection of documents, without being told beforehand what the topics are.
Three features make the technique a good match for patent work:
- Each topic comes with keywords. A topic is characterised by the set of keywords that best distinguish it from the other topics.
- The keywords make topics readable. An examiner can look at the keywords and quickly judge whether the documents linked to that topic are likely to be relevant, without reading every single one.
- One document can belong to several topics. This is helpful when an invention touches on several distinct technical aspects at once.
To take a simple illustration, a patent for an electric bicycle could sit in one topic about battery mounting and in another about frame design.
If you would rather see AI-assisted searching in action than read about it, you can try a free AI-powered novelty search on your own idea and review the closest prior art it identifies.
Tuning the Model: No Single Setting Fits Every Class
Getting useful results from LDA is not a matter of running the algorithm once. Several parameters shape how helpful the resulting topics turn out to be:
- The number of topics to identify
- The number of keywords used to describe each topic
- The number of training iterations
- The minimum probability thresholds
The researchers therefore ran a series of systematic topic modelling experiments on their validation data. The aim was to find the best parameter settings for each individual IPC category, rather than relying on one generic configuration for all of them.
The Honest Limitation: Topics Are Hard to Evaluate
Topic modelling is an unsupervised technique, and that makes it notoriously difficult to evaluate. In supervised classification, predictions can be checked against known correct answers. With topic modelling there is no answer key, and there is no built-in guarantee that the topics an algorithm produces will feel coherent to a human reader.
Researchers in the field have developed topic coherence measures to put a number on this (Röder et al., 2015). The Cardiff team used these established measures and then went a step further by proposing an additional method of their own for measuring interpretability directly. When the task is triaging real, filed patent applications, a topic that a person cannot interpret is of little use.
Document Similarity: Comparing an Application With Earlier Patents
Alongside topic modelling, the study explored a more traditional information retrieval approach built on measuring document similarity directly.
In a conventional search, the examiner first maps the invention onto a suitable set of search terms, and the system then retrieves matching documents from the database efficiently. Document similarity takes a shortcut. The invention is already described in detail in the patent application being examined, so the application itself can be compared against previously filed patents to surface the closest matches.
Patent offices have since put this idea into daily practice. The USPTO, for example, gave its examiners an AI-based Similarity Search tool in 2022, which takes the text of the application under examination and returns a ranked list of similar documents.
The Vector Space Model and TF-IDF Weighting
Traditionally, this comparison is carried out with a vector space model. Every document is represented as a vector, which is essentially a long list of numbers. Each coordinate corresponds to an individual word, and its value reflects how frequently that word appears in the document relative to how it is distributed across the whole collection.
This weighting scheme is TF-IDF, the measure introduced earlier in this series (Spärck Jones, 1972). A word that appears often in one patent but rarely elsewhere receives a high weight, because it says something specific about that document.
Once documents are represented in this way, two of them can be compared by measuring the distance, or the similarity, between their vectors.
Why Cosine Similarity Gave the Best Results
The researchers experimented with several measures, summarised below.
| Measure | What it looks at | Behaviour with documents of different length |
|---|---|---|
| Jaccard similarity | The overlap between the sets of terms in two documents | Overlap shrinks when one document contains many more terms |
| Euclidean distance | The straight-line distance between two vectors | Sensitive, because longer documents produce larger vectors |
| Cosine similarity | The angle between two vectors | Robust, because only direction counts |
As expected, the best results came from cosine similarity. It measures the orientation of two vectors rather than their raw magnitude (Jurafsky & Martin, 2008), which makes it more robust to differences in document length. That matters for patents, where one specification may run to a few pages and another to several hundred.
From Whole Documents Down to Single Sentences
All of these measures can be applied at different levels of granularity: entire documents, individual sections, paragraphs, or even single sentences.
Most genuine inventions are incremental improvements on existing solutions, not something built from scratch with no relationship to earlier work. It therefore becomes important to pinpoint the specific paragraphs or sentences elsewhere that describe closely related ideas, instead of only comparing documents as undivided wholes.
Short texts create a technical problem, though. For a title or a single sentence, the standard vector space model tends to produce sparse feature vectors that struggle to discriminate between texts, even when dimensionality reduction is applied, because of the sheer number of dimensions involved (Houle et al., 2010).
To work around this, the researchers reused word embeddings to encode the meaning of sentences. The most direct approach represents a whole sentence by the centroid, or geometric average, of its word embeddings, and then measures the distance between the centroids of two sentences.
Word Mover's Distance: Bridging Different Vocabulary
A more fine-grained measure explored in the study is word mover's distance, or WMD (Kusner et al., 2015). It calculates the minimum cumulative distance that the words of one sentence would need to "travel" through the embedding space to reach the words of a second sentence.
The study gives a concrete example using the phrases "extendible umbrella handle" and "parasol foot with retractable point". After common stop words are removed, the distance is the sum of the distances between the closest matching word pairs:
| Word in the first phrase | Closest word in the second phrase |
|---|---|
| umbrella | parasol |
| extendible | retractable |
| handle | foot |
The two phrases do not share a single word, yet they describe closely related concepts. A purely keyword-based search would most likely miss the connection entirely. Word embeddings close exactly this kind of gap.
What This Means for Your Own Prior Art Search
You do not need to implement LDA or word mover's distance to benefit from the findings. Four practical lessons stand out:
- Retrieval is only half the job. A long result list has to be organised by theme or by similarity before a person can review it.
- A full description can beat a keyword list. Similarity-based searching starts from the text of the invention itself.
- Look at passages, not just documents. The closest prior art is often a single paragraph inside a longer patent.
- Human judgement stays in the loop. A person must be able to understand and trust what the system shows.
If you are still building your search skills, our prior art search resources explain databases, classification and search strategy step by step.
Similarity Analysis for Your Own Invention
WOIPS FeatureThe study found that comparing full descriptions works better than matching keyword lists. WOIPS is built around that finding.
Your search starts from a structured description of your invention, not just a few keywords. WOIPS retrieves candidate patents and non-patent publications, ranks them by relevance, and then analyses each document against your invention to show where they are similar and where they differ.
The result is an organised, reviewable list instead of a long, unsorted one, with the final judgement left to you. Discover how WOIPS compares documents with your invention.
Frequently Asked Questions
What is topic modelling in patent searching?
Topic modelling is an unsupervised AI technique that groups a collection of patents into themes automatically. Each theme is described by a set of keywords, so a searcher can decide quickly which groups are worth reading.
Why is cosine similarity used to compare patent documents?
Cosine similarity compares the direction of two document vectors, not their size. A short patent and a long patent on the same subject can therefore still be recognised as similar.
What is word mover's distance?
Word mover's distance measures how far the words of one sentence must travel in the word embedding space to match the words of another sentence. It can detect related meaning even when two texts share no words at all.
Do I need to understand these techniques to run a prior art search?
No. AI-powered search tools apply methods like these in the background. You describe your invention in plain language and then review the documents the tool identifies as closest.
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.
