How AI Semantic Search Is Changing Patent Discovery

Patent discovery is often hindered by the challenge of understanding what terminology is used to describe a technology, even when the fundamental concept is known. Semantic search attempts to address this issue by looking for relationships based on meaning rather than exact word matches. This is achieved through the use of text embeddings, where both queries and documents are converted into numerical representations that capture their content aspects. These representations can then be compared to retrieve semantically related documents.

For instance, when searching for patents related to reducing heat in electric vehicle batteries, a semantic search system might return results discussing heat dissipation, temperature regulation, or cooling arrangements described using different terminology. While these additional results can potentially guide researchers in new directions, they are not necessarily technically equivalent to the initial query. Therefore, researchers must examine the retrieved documents in context to determine their relevance.

Two patents related to battery cooling illustrate the challenge. One describes a liquid circulation system to regulate battery temperature, while the other focuses on directing airflow around individual battery cells. Although both patents address thermal management, a semantic search system might consider them related, but it does not guarantee that the latter patent is relevant to the first research question.

Evaluating the effectiveness of semantic search involves comparing results obtained using conventional keywords and natural-language descriptions of the same problem. This helps identify if the search uncovers relevant documents using different terminology and assesses whether the additional results improve the research process or merely increase the amount of material to review. Real-world evaluation can be done using a small dataset of already-reviewed patents.

Despite the benefits of semantic search, it is not a panacea for patent discovery. It complements human expertise by helping users understand what they need without knowing the exact terminology used in the source material. Large language models can further assist by summarizing documents, explaining unfamiliar terms, and suggesting alternative ways to express technical concepts. This can be particularly useful when the preliminary search yields little or when the researcher is unfamiliar with the field.