What information would you like to extract from patents?

Question

I am a Natural Language Processing / Machine Learning person and I'm starting to explore the patent world.

Since there are huge repositories for patents, I was wondering if any other useful information could be extracted from there. Which patent-related question might be interesting for you? This can be anything from "Is there any trend/new concept/new industry which is emerging?" to "Can I know the intended use of the invention automatically?" to anything, really. Do you think something is lacking or is still not good enough?

I'm looking for any answer, no matter how simple or overcomplicated.

Thanks!

Answer 1

Some ideas:

1. Every patent description contains a list of references to Related patents and other prior art. A complete survey of the relationships between patents and other prior art could produce a fascinating and useful interactive explorable map.

2. The claims themselves are composed in dense language, and at first blush appear to be ill suited to machine learning. But if you could parse and interpret the claims via references first to supporting language in the patent description and second to similar language in the Related patents, you might derive some machine understanding of the claims. This would be a big research project though.

3. Qualifications such as "means or step plus function" limitations in a claim have specific significance in patent law. Just determining and presenting those qualifications would yield a valuable tool.

4. Each patent's prosecution history contains important information, ranging from derivation from a provisional to correspondence with the examiner to inter partes reviews. A tool that lets you see the highlights of that history at a glance would be awesome.

5. A patent often relies on assumptions about prior art that have been disproven in the prosecution or litigation of other patents. Digging that up is time consuming and complicated; a dedicated tool could help.

6. Federal court docket trackers like DocketNavigator and DrugPatentWatch let you explore events in patent litigation. But they don't provide insights. AI could help with that.

Answer 2

Usually when I'm searching patents, it is because I'm looking for what is called "prior art". In other words I have a technology in mind and I'm looking for patents or applications that are highly relevant. This is tricky to do with keyword searches because not everyone uses the same keywords. Indeed, sometimes patent lawyers intentionally use uncommon terms to describe the invention. There may be good legal reasons for this obfuscation, but it is annoying.

My process is generally iterative where I search broadly, quickly scan the results and then use the most relevant identified documents to help target the search by doing things like searching for other patents/applications by the same inventors, classification codes, assignees, etc. I once used a software package called by Invention Machine Goldfire that purported to use AI to provide natural language "semantic" search for patents. Try as I might, I didn't find that aspect of the package very helpful.

Answer 3

This is more modest than some of the other suggestions.

Much of the patent search and examination process revolves around classes. For example new applications are assigned to examiners based on determining a class/subclass for the application. A former USPTO examiner explained that this is now done automatically based on the wording of claim 1. I think it might be useful to patent searchers to have a way to find a class from the text of a proposed claim. This can aid in searching by class.

It could be approached using the published definitions of the classes or just by digesting a lot of claims and the class designation the applications were assigned to.