Access Official Technical PDF

Share Summary:LinkedInEmail

Patent trend analysis can help industrial organizations examine how technical priorities change over time. Patent records contain more than descriptions of inventions: across many documents, recurring terms and phrases can reveal the topics an organization is pursuing. The challenge is turning millions of patent abstracts into structured information that engineers, technology strategists and operations leaders can use.

An approach developed for U.S. patent abstracts addresses this challenge by automatically identifying and ranking “catchphrases”—short expressions that represent a document’s technical subject. The method is unsupervised, meaning it does not require a manually labelled training set for the core extraction process. Applied across an organization’s patents, these phrases can support analysis of topic continuity, topic shifts and citation patterns.

Why patent abstracts need structured analysis

Patent documents are detailed, technical and numerous. Reviewing them individually can make it difficult to identify broad changes in an organization’s innovation activity. A concise, searchable representation of each document can make the collection easier to index, categorize, compare and explore.

Catchphrases are intended to capture meaningful concepts rather than simply list frequent words. They can range from a single technical term to multiword expressions such as “integrated circuit device.” That distinction matters in industrial settings: a phrase may express a particular component, process or system more clearly than an isolated word.

Automated extraction is useful because only a small proportion of documents have catchphrases assigned by their authors, while adding them manually across a large patent archive is labor-intensive. A consistent extraction process can make it more practical to search and compare patent collections at scale.

How patent catchphrase extraction works

The method described in the study uses two stages: candidate selection and candidate ranking. It was applied to USPTO patents granted between January 2003 and May 2018, using abstracts where available. The researchers also used a non-patent text collection to help distinguish language that is particularly characteristic of patent documents.

1. Select candidate phrases

First, natural language processing identifies noun phrases in each abstract. Candidates include one-word terms and phrases of up to four words. The researchers found that retaining stopwords—the common words often removed during text processing—produced better extraction results in their experiments, because such words can be part of meaningful technical phrases.

2. Rank candidates by relevance and specificity

Next, each candidate is scored using two kinds of evidence. One estimates how well the phrase represents the content of that specific abstract. The other estimates how distinctive its terms are to patent text compared with non-patent text. Combining these signals produces a ranked list, from which a user or analysis pipeline can select the highest-ranked phrases.

This design addresses two different questions: “Does this phrase describe the patent?” and “Is this phrase informative in the patent domain?” Considering both can help distinguish a useful technical concept from a phrase that appears frequently but carries little topical meaning.

Using extracted phrases for patent trend analysis

Once patents have comparable phrase representations, organizations can examine how their technical topics evolve. In the study, researchers compared the sets of two-word catchphrases appearing in patents filed by the same organization in consecutive years. They used Jaccard similarity, a measure of overlap between two sets: greater overlap indicates more shared phrases, while lower overlap indicates less overlap.

Interpreted carefully, this provides a way to explore whether an organization’s patent topics remain relatively consistent or shift over time. The study’s analysis of selected software, hardware and mobile-phone companies found different patterns, including profiles with multiple periods of higher topical similarity and profiles with lower similarity across years. A low-overlap period may prompt further investigation into emerging technical areas; it does not, by itself, explain why a company changed direction.

To make comparisons more interpretable, the researchers grouped similarity profiles according to their shapes, including profiles that generally increased, peaked later or contained multiple peaks. They also normalized company-name variations and removed frequently occurring phrases that could create noisy or redundant signals. These steps matter when building an organization-level view from records that may use different names or contain generic patent language.

What citation patterns can—and cannot—show

The analysis also compared company-level topic similarity patterns with the citation counts received by patents. Patents were grouped into citation ranges, and the researchers examined the share of patents in each range for different similarity-profile categories. Some categories showed a larger share in the medium and high citation ranges than others.

This is an observed relationship, not proof that topic diversity causes higher citation counts. Citations are one indicator of how later patent documents refer to earlier work; they are not a complete measure of commercial success, technical quality or operational value. Citation patterns can be used to formulate questions for further review, but they should not be treated as a standalone performance measure.

Practical considerations for technical teams

For engineers, system integrators and technology decision-makers, patent catchphrase analysis can complement other sources of technology intelligence. It can help organize large patent collections and provide a starting point for questions such as:

  • Which technical phrases recur across an organization’s patents?
  • When do the topics in a patent portfolio appear to shift?
  • Which phrases distinguish one organization’s patent language from another’s?
  • Do changes in topic overlap coincide with changes in citation patterns?

Results depend on the documents and processing choices used. Abstracts do not contain every detail in a full patent, and automated phrase rankings should be reviewed in context before informing engineering or investment decisions. The study evaluated its method against established extraction approaches and reported higher precision and recall in its tests, using phrases from the International Patent Classification as evaluation references. Those findings describe the study’s evaluation setup, not a guarantee for every patent collection or application.

Conclusion

Patent trend analysis based on catchphrase extraction offers a structured way to explore technical topics across large patent archives. By identifying candidate noun phrases, ranking them for document relevance and patent-domain specificity, and comparing their use over time, organizations can make patent text more searchable and examine changes in innovation focus. The approach is most useful as an analytical aid: combine its signals with expert review and other evidence, and treat citation relationships as correlations rather than causal conclusions.

Access Official Technical PDF

About The Author