December 3-4, 2021 Royal Sonesta Hotel | On Zoom
Session + participant details on the NBER site | Edit this agenda
Adam Jaffe, Brandeis University and NBER
The Rise of Process Claims: Evidence from a Century of U.S. Patents (slides)
Bernhard Ganglmair, ZEW Mannheim
W. Keith Robinson, Wake Forest Univesity
Michael Seeligson, Southern Methodist University
An approach to classifying claims, beyond keyword searches. We will be sharing code & data (on Github & Zenodo) for others to use and adapt.
Q: When you’re relying on textual models, what problems come up w/ formulas and schematics, given how they are represented in those patents?
A: (Bruno + Bernhard looked at eachother and both rolled their eyes, indicating this is a hard problem! <laughter>) Bruno: This is not something we are picking up, and these are important elements. Even if you think about Josh Greer’s work in pharma… molecules, and how they are picked up… if it’s not expressed in words, we will miss it for now. That’s a challenge, more in some areas than in others.
Bernhard: Looking at our coverage, how well we do for claims in chem and biology, we’re missing a lot. I’ve just recently come across a dataset of a few hundred thousand manually classified claims in chemicals. If you are in the audience, please call me — I would love to use the dataset just to see how well we do w/ a large sample of chemicals, to improve on the approach.
Modeling Patent Clarity
Jonathan Ashtor, Cardozo School of Law
Can the definiteness requirement be used to model claim clarity, at issuance and publication? This work modeled linguistic features of claims, and trained an ML model on rejections using these features.
Q: Do we know anything about which firms write better patents?
A: Portfolio size may bear on this; I don’t look at the division or subsidiary level, just at the general size of the portfolio, so it’s a bit rough.
Technology Differentiation and Firm Performance
Bruno Cassiman, IESE
Sam Arts, KU Leuven
Jianan Hou, KU Leuven
How do we characterize the tech portfolio of a firm, and how does that correlates with firm performance? How can we use text to characterize the portfolio?
There is some work on measuring similarity and complementarity of the tech of different firms, but little theory + formal models. For a given year, we can plot firms in technology space.
Measures: portfolio similarity, technology differentiation (very different from patent similarity and citations)
We are sharing this method and open data, which could be used to look at other questions — characterizing portfolios by region, or by inventor.
Q: How do you cluster / classify firms that work in many classes?
(E.g.: firms compete at business unit level, not at corporate level.)
A: We don’t cluster, trying to get away from classification…
Comment: Can we draw connections w/ Bernhard’s paper around differentiation and competitive advantage? Process vs product inventions are on a continuum, but if you think about process patents as cost-reducing and product patents as value creating, that might help distinguish different ways that firms try to differentiate.
Closing comment: we had 2 papers that take the words in the patents at face value: “what does this tell us about the firm’s inventions?” and 1 paper that asks “how do firms write the patents they write?” This area is exciting – as someone who started w/ classifications, this clearly tells us much more. About both the invention and the applicant’s strategy.
Introducing OpenAlex – maintaining an open replacement for MAG (outline)
Jason Priem, OurResearch
Heather Piwowar, OurResearch
Open Alex will provide a replacement for MAG, as of Jan. 3. The code and data will all be open. Here we describe 4 types of data you might be looking for: what was in MAG but is going away (patents!); what is frozen w/ no updates; what has ongoing support; what is new.
Feedback welcome once this is out in January, please share how you are using it and what else you would like to see.
Matt: Thank you! In some ways MAG going away was a good thing, as we have no documentation of what MAG did; no code, no benchmarks. Thanks for open sourcing everything: a quantum leap. We want to help; I may have data to share with you after Jan 3.
Q: What are your thoughts on licensing the data? ODC-BY, other?
A: our current license is ‘none’ as with Crossref data. (CC0; facts are not copyrightable). Lawyers are arguing about this, we will keep our ears open if people think this is not true for some subset; but we don’t think we have the right to apply a [new] license to it.
Agnes Cameron, Knowledge Futures Group
A searchable update to the original I3 datasets index.
Linked open data for each dataset, with annotations: metadata, examples, superceding data.
You can edit metadata in the original Google sheet, and can add longer-form markdown notes that display as usage guides alongside each entry.
We built this largely to include datasets and metadata that often don’t get captured on other platforms — links of inheritance, dependence, and supercession; related datasets; and timelines.
You can now add curated collections: a way to index thematic lists for a specific purpose. Items in a collection are often in the index as individual entries, but collections may include both larger and more granular elements.
Feedback is warmly welcome! If you have ideas about how to approach dataset-relatedness, or other facets you would like to see, please share — as issues on the github repo or by email.
Measuring Firms' Technology Use with Employees' Job Data
Tania Babina, Columbia University
Anastassia Fedyk, University of California at Berkeley
Alex Xi He, University of Maryland
James Hodson, Jozef Stefan International Postgraduate School
Using employment and employee data from Burning Glass + Cognism
Before we had papers; now we have resumes: for firm-worker match data
We use this to infer AI skills in particular.
Q: do you cluster resumes to find duplicates for the same person?
A: In 2018 there were 165M workers, we have resumes for 100M
Q: Is open data available?
Cognism: Firm-level measures + geographies will be released.
Burning Glass: may be able to post aggregated data, need to ask | TBD
Q: Have you thought about generalizing this method to firms that use tech from elsewhere to do their work?
Comment - Really cool data! I thought an interesting result was that you find a positive correlation between the AI measure and the Hoberg-Phillips fluidity measure. What's your interpretation of this? I guess you could imagine that firms are trying to maintain a competitive advantage, or are responding to a changing competitive landscape
Matching Patent Assignees to Startups
Matt Marx, Cornell University and NBER
Michael Ewens, California Institute of Technology and NBER
NB: Small != Startup. Some orgs stay small.
Youth != Startup. Some established orgs spin out branches that are ‘old’. We don’t want to reinvent the wheel, there are other efforts to match assignees to startups.
Harmonizing a few sources:
– OpenCorporates data: they are transparent about provenance, access is free; but no data on: headcounts, sales, industry ratings, industry field codes.
– Form D filings: doesn’t get all startups, but has a lot of them
Comment: If you have inventor names, and employment, and know that an inventor took out a patent that led to a product at that company, that could be a useful sign.
Intellectual Property Theft (slides)
Britta Glennon, University of Pennsylvania and NBER
Daniel P. Gross, Duke University and NBER
Lia Sheer, Tel-Aviv University
IPR analysis: how do firms respond to (theft cases, theft trends)?
How does theft affect their H1B hires, incl. specifically from China?
How does it affect the # of immigrants they hire?
After completing the work and putting out a first pub, will start sharing data privately with other researchers.
This is the hardest paper we’ve worked on yet, for reasons Britta notes
Comment: MSU group in this area: https://a-capp.msu.edu/
Q: Curious about how many of the people involved in the thefts were permanent residents. And how many first came to the US on a student visa. Are firms hiring different types of immigrants rather than fewer immigrants?
A: That is a great point. once we collect the additional data we hope to have a better understanding on the background of these workers.
Q: are cases initiated by firms or by DOJ? (Or is this known?) I’m trying to understand what share of theft might be captured by cases, and it seems like some firms may have incentives not to make it public they were successfully victimized.
Q: Can we learn anything from the strategies of the thieves themselves? Are they stealing frontier tech?
Q: this is really cool! Two thoughts... 1. How do you guys know what 'event' to code for the diff-in-diff: there's the news catching the theft, the indictment of the spies, the veredict.. etc... and 2. would be interesting to separate the cases where the defendant is ultimately found innocent vs guilty..
A: these are great questions. We ideally want to be able to measure when the firm itself learned of the theft. At the moment we're working with the earliest date we can get our hands on, which isn't always that far back, but we have an RA trying to push deeper into the timelines as we speak.
Q: @dan @britta @lia - what is the underlying theory behind the matched sample analysis? If a firm has not been affected should one think of this as a persistently lucky firm (which does not need to take any action to be safe) or should we think of it as a firm that has taken appropriate actions?
In other words, I am not sure what story would generate a difference between the treated and control firms
A: we started looking also into spillovers to rival firms, but first want to understand the treaded firm's action. Indeed a lot to think about!
Comment: Can H1B can be considered immigrants as this is a kind of non-immigrant visa? Comment: In the sense that people are moving to the US (for their jobs), they are immigrants, even if it's not clear how long they will stay
U.S. Entrepreneurship over the Long-run: New Data and Approaches to Measurement (slides)
Daniel P. Gross, Duke University and NBER
Jorge Guzman, Columbia University
Innessa Colaiacovo, Harvard Business School
Q: is the real outlier the inter war years?
A: in the usual log-log GDP graphs, there is an outlier in the interwar and then a return to trend.. in our case, we find a change in trend after the war. we have some hypothesis about what this means, still prelim so I’ll hold off :)
Q: to what extent can you look beyond firm names? can you look at firm descriptions ultimately?
A: We’d love to, but can’t too much. one option could be to try to look at what the 'filings' say but they are mostly not digitized and are handwritten (in cursive!) when digitzed for the earlier years in our sample....
Q2: I wonder if you can capture related meanings beyond the words? For example, firm names might include “code” related to “program”?
Q: with D&B data do you have a sense of establishment longevity? or places going from single establishment to multi establishment firms?
A: we have thought about this quite a bit, but matching the two and thinking about selection differences can be tricky... happy to hear your thoughts
Comment: Thanks very much for another heroic data project! The change of ownerships would be interesting to trace too.
Matt Marx, Cornell University and NBER
Frame: 21 datasets used by 27 articles on patent-paper pairs. Most by hand, only one is open (PubMed, 15 pairs). Fiona, et al match to Web of Science; most don’t have full access to allow clean reuse. Ths is where we were w/ patent-paper cites 3+ years ago.
How do we define a PPP? settle on a definition.
Aim to produce a broad public dataset. Link to OpenAlex, cover all fields (not just med), include advanced data: geography, affiliation
Comment: It could be illegal to put CEO on patent when they were not involved --> patent is worthless if challenged. But there are probably grey areas in both cases - defining what is a substantive/large enough contribution is often a debate. See https://www.science.org/doi/10.1126/sciadv.1700404
Q: How do you decide what to work on?
A: When you see lots of papers and no sharing of source datasets - the 21 sources for 27 papers - that’s an indication that there’s a role for a public good
Q: For temporal specificity - you can publish a paper up to 1 year after, measured from absolute priority date
Q: For pairs w/ no overlap of author+inventor: is that still a pair?
A: Don’t only look at similarity; do you also require some similarity of content?
Andy T: authors and inventors should overlap by at least one person. if another person published something that is then in a patent app, it becomes prior art and should block the patent from grant. the author would have 1 year to submit before they undermine themselves for getting a patent
Emilio R: We are currently working on a project that exploits the funding source to identify the patent-paper pairs. It's going to be limited in size but maybe not terrible.
Bhaven: (noting a gold-standard set you can measure against)
Q: Can you separate different families?
Comment: can you look at both cosign similarity, not just topic pairs?
Matt: I hear people asking for more info about different fields, not just one ranking of “likelihood of being a PPP”
BP: Lots of