The Language of Innovation

8.45 million patent abstracts provide a rich corpus for examining the thematic structure of US patenting activity. By applying NMFNon-negative Matrix Factorization — a matrix decomposition method used for topic modeling that produces interpretable, additive topic representations. topic modeling to every utility patent abstract filed with the USPTO since 1976, this analysis uncovers the latent themes of US patenting. Using TF-IDFTerm Frequency–Inverse Document Frequency — a numerical statistic that reflects how important a word is to a document relative to a larger collection. Common words are down-weighted. to convert raw text into numerical features, and non-negative matrix factorization to discover 25 latent topics, the results indicate which themes are rising, which are declining, and how they correspond to the formal technology classification system.

The stacked area chart below illustrates how the share of each topic has evolved over time. Computing, semiconductor, and communications topics have expanded substantially, while traditional mechanical and chemical engineering topics have experienced a decline in relative share, though not necessarily in absolute volume.

To visualize the full semantic landscape of patents, a stratified sample of 5,000 patents is projected from high-dimensional TF-IDF space into two dimensions using UMAP. Each point represents a patent, colored by its dominant topic. Clusters indicate families of related inventions, and overlapping regions suggest technology convergence. Note that the axes of a UMAP projection are unitless — only the relative distances and clustering patterns are interpretable, not the absolute positions.

The relationship between discovered topics and the formal CPC classification system warrants examination. The chart below cross-tabulates the eight most prevalent topics against CPC sections. Some topics align closely with a single CPC section (for example, chemistry-related topics correspond to Section C), while others, particularly computing, span multiple sections.

The degree to which contemporary patents differ from those of earlier decades constitutes an important empirical question. This analysis measures novelty as the Shannon entropy of each patent's topic distribution: patents that draw approximately equally from many topics (high entropy) are more thematically diverse, and may be considered more novel, than patents concentrated in a single topic (low entropy).

Having uncovered the latent thematic structure of patent language, the analysis turns next to the legal and policy framework governing the patent system. The topics and trends identified in this chapter provide essential context for understanding how legislative and judicial decisions have shaped the direction and character of US patenting activity.