My feeling lately is that Boolean searching will very likely be abandoned within five to ten years.
Keywords: one technical concept has too many possible expressions. However many synonyms you add, something is always missed and every new term brings new noise. Abstracts also miss much of the technical content, and searching the description explodes it.
Patent Classifications: the system exists to route work to examiners, not for patent analysis. It is complex, layered on top of itself, useless for emerging technologies, and distorted by human error. Use it only as a support.
It is still mainstream, mainly because of cost.
- A thorough novelty search benefits neither side: if the invention is already disclosed, the firm cannot close the case and the client has trouble explaining it internally. Some clients treat patents as a vanity project and will not pay for meaningful claim scope, so the firm has no incentive to search. Bad practice drives out good.
- Sometimes traditional searching is enough. Apart from FTO, most searches can close without being exhaustive, especially when the client cannot tell good work from bad.
- There are no other tools. Academia is ahead of industry on search technique, but those methods never reached the search systems.
- Experts patch the gaps themselves, with unions of dozens of queries for recall, then heavy manual reading for precision. How much they actually read is left to conscience.
Recall and precision cannot both be had on limited resources. Once recall passes 80%, precision usually drops below 30%. A search returning 10,000 patents still leaves thousands to read by hand, which the fees make impossible.
A search also does not audit itself, since that invites trouble. Recall and precision are usually only demanded for landscape work, and even there only precision can be accurate. Recall requires counting what the search missed, and what you never retrieved cannot be the standard. The usual substitute is a gold standard built from a few companies’ patents, which is prone to bias: the ones chosen are usually the most relevant patents, not the borderline cases that get missed. Most recall figures in reports are custom made.
High quality has only one path: enough people and enough money. The hours cannot be escaped.
So what does AI bring? Speed in reading large volumes of patents. In my own implementation, high accuracy still costs close to what humans cost. But if token costs come down, the old workflow can be waved goodbye.
Overturning a method I have used for over ten years is uncomfortable. But since it lifts quality to another level, I have to admit I simply did not know what I did not know. Keyword searching will not be useless. Its role moves earlier, to rough screening ahead of AI search.
*I am a Taiwan patent practitioner, and this reflects my experience there.