AI patent prior art search depends on query expansion to find relevant documents that describe the same idea in different words. Query expansion adds related terms, such as synonyms, narrower terms and domain-specific neighbours, to the examiner’s starting search terms so that fewer relevant patents are missed. The UK Intellectual Property Office feasibility study tested two complementary methods, a thesaurus called WordNet and word embeddings, and then ran the finished query through a local Elasticsearch engine.
Choosing good initial search terms is only the first step. Patent applicants and authors often describe the same concept with completely different vocabulary, so a search built on a single set of words can easily overlook a highly relevant document. The study treated this wording gap as a core problem and explored two ways of closing it.
Lexical Relationships: Expanding Queries with WordNet
The traditional route to query expansion is a thesaurus, a resource that groups words according to relationships such as synonymy, meronymy and hyponymy. The researchers used WordNet, which is widely considered the largest lexical database of English. It sorts nouns, verbs, adjectives and adverbs into synsets, short for “sets of synonyms”, and each synset represents one distinct concept.
These synsets are connected by lexical relations. Hyponymy means “is a kind of”, so a sedan is a kind of car, while meronymy means “is a part of”, so an engine is a part of a car. Starting from any word, the system can follow these links to surface related terms and offer them to the examiner as sensible additions to a query.
To make this practical, the team did not ask examiners to look words up in a separate tool. Instead, they used the NLTK WordNet API to read WordNet’s data programmatically and built the results directly into the system’s own interface.
Semantic Relationships: Expanding Queries with Word Embeddings
WordNet is easy to use and broad, but it is a general-purpose resource, so its coverage varies from one technical domain to another. The study offers a telling example: WordNet knows “tablet” as a small pellet of medicine, yet it does not capture the far more common modern meaning of a mobile computing device. Word embeddings fill this gap, because they can be trained on a domain-specific collection of documents and therefore learn the meanings that actually apply in that field.
Once words are placed in a vector space, the system can retrieve a word’s nearest neighbours and offer them as expansion candidates. In the transport domain, the neighbours of “driver” include synonyms such as “vehicle-operator” and narrower terms such as “cyclist” and “motorman”. In other words, embeddings are not limited to one kind of relationship.
That flexibility has a trade-off. On their own, embeddings can show that two words are related, but not how they are related. The distance between words still works as a rough indicator of relationship strength, and adjusting a distance threshold lets the system control how many alternative terms the examiner sees at one time.
Want to see how expanded queries translate into real results? Yes. You can explore the WOIPS Novelty Search service and see how relevant documents and their similarities and differences are identified.
How Query Expansion Changes the Examiner’s Role
Together, the two approaches turn a basic query into a steadily improving search strategy. WordNet contributes structured, labelled relationships, while embeddings contribute flexible, domain-aware suggestions. The result is a change in daily practice: instead of recalling every possible term from memory and personal experience, the examiner reviews and refines a list the system has already proposed.
The study highlights two benefits. Building a query becomes more efficient, and results become more consistent between different examiners handling similar applications. Consistency supports reproducibility, which matters a great deal in a legal and regulatory setting. To keep humans in control, the researchers created an interface where examiners can select or deselect any suggested term rather than accept the whole list.
Local Search Engine: Elasticsearch and Its Suggesters
After the query is finalised, whether it was expanded through WordNet, word embeddings or both, the search itself runs on Elasticsearch, a very popular search and analytics engine. The team also used three built-in Elasticsearch features called suggesters: the term suggester, the phrase suggester and the completion suggester.
The term suggester proposes alternative words one token at a time within a set edit distance, meaning the number of single-character changes between two words. This helps with spelling mistakes and regional spelling variants. The phrase suggester goes further and uses an n-gram model to propose whole corrected phrases based on how words normally appear together, which can help an examiner decide which terms to keep. Both power familiar “did you mean” suggestions.
The completion suggester provides search-as-you-type behaviour using data structures built for very fast lookups. Examiners get instant feedback while typing, which can steer them toward relevant documents before the query is complete and improve precision.
Plug-ins, Python Integration and the Local Setup
Elasticsearch can be extended with plug-ins. One example is the International Components for Unicode (ICU) plug-in, which improves support for Asian languages, Unicode normalisation, case folding, text collation and transliteration. This matters for patent work, because documents come from many jurisdictions and scripts.
Elasticsearch is normally reached through a REST API. The research team instead used Python bindings so that it became a closely integrated part of the proposed system rather than a loosely connected outside service. For this feasibility study they installed a local Elasticsearch server and stored all validation data separately from the other components, with further technical detail provided in an appendix of the full report.
What This Means for Modern Prior Art Searching
The main lesson is that no single technique covers every need. A thesaurus gives clear, explainable relationships, embeddings give domain sensitivity, and a capable search engine adds spelling tolerance and fast feedback. Combining them, while leaving the final choice of terms to a human expert, balances automation with professional judgement. For broader background, see our guide to prior art search, and for official reference material on patent searching, see the WIPO patents resources.
Frequently Asked Questions
What is query expansion in patent searching?
Query expansion is the process of adding related words to an initial set of search terms so that documents using different wording for the same idea are also found. It reduces the risk of missing relevant prior art.
What is the difference between WordNet and word embeddings?
WordNet is a hand-built thesaurus with labelled relationships such as synonymy, hyponymy and meronymy, but its coverage of specialised domains can be limited. Word embeddings are learned from domain-specific text and capture field-specific meanings, although they cannot state what type of relationship links two words.
Why did the researchers use Elasticsearch?
Elasticsearch is a widely used search and analytics engine, and its built-in suggesters add spelling correction, phrase suggestions and auto-complete. These features help examiners refine queries and improve precision. You can read more about how searches are structured in our prior art search hub.
Do examiners have to accept every suggested term?
No. The study’s interface lets examiners select or deselect each suggested term, so the system supports their judgement rather than replacing it.
Can I see AI-assisted prior art search in action?
Yes. You can submit your invention for a free search with WOIPS and see how relevant documents and their similarities and differences are identified.
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.