Oren Etzioni doesn’t think the rapidly evolving academic search industry is going to make anyone rich. But he does hope it will lead to big breakthroughs by helping researchers sift through millions of papers, with artificial intelligence connecting dots within and between disciplines.
Etzioni has been a professor of computer science at the University of Washington since 1991, but he’s had an impact far beyond the classroom. The late Paul Allen, Microsoft’s co-founder, picked him in 2013 to be the first CEO of the Allen Institute for Artificial Intelligence. Before that, Etzioni created MetaCrawler, one of the first search engines, and he has also cofounded and sold a series of tech companies, including NetBot, the first commercial comparison shopping engine; FareCast, which predicts travel prices; and Decide.com, a shopping adviser for electronics. Etzioni is also a partner in the Madrona venture-capital group in Seattle.
When I was 18, I thought that by now we would have machines with human-level intelligence.
One of the Allen Institute’s priorities is an academically oriented search engine, established in 2015, called Semantic Scholar (slogan: “Cut through the clutter”). The need is great, with more than 34,000 peer-reviewed journals publishing 2.5 million articles a year. “What if a cure for an intractable cancer is hidden within the tedious reports on thousands of clinical studies?,” Etzioni once said.
Although Semantic Scholar has focused so far on computer and biomedical sciences, Etzioni says that the engine will soon push into the social sciences and the humanities as well. The Chronicle spoke with him about information overload, impact factors’ imperfect inevitability, and the promise and perils of AI.
•
What is Semantic Scholar, and how does it differ from Google Scholar, PubMed, and other academic search engines?
All these engines are fighting the problem of information overload. Across many disciplines, it’s been documented that the number of publications is doubling every few years, and this has been going on for decades. Academic searches become more and more of a needle-in-a-haystack problem, and not nearly as much money or as many resources have been dedicated to this problem as to finding a good product on Amazon or a good video on YouTube.
Where Semantic Scholar is different is that we’re committed to using our AI capabilities, specifically around processing natural language and images — the figures and diagrams and so on. Specifically, we use AI to rank, to filter, to get rid of extraneous information, and then to extract, whether it’s the results, clinical-trial databases, connections between different papers or between papers and blog posts that summarize them. The ultimate in cutting through the clutter is, Hey, what if I don’t have to read these 10 papers at all because I have a handy blog post or video that summarizes the relevant information? Everything we do falls under the heading of helping scientists be much more efficient.
In terms of how other search engines are using AI, is it a matter of degree that separates you?
Yes. We’re built on AI from the bottom up. Whereas they have a sprinkling of AI, for us it’s the meat and potatoes. PubMed, Google Scholar, and others are fundamentally keyword search engines. Our fundamental hypothesis is that as research grows exponentially, keyword searches aren’t going to cut it anymore.
Which search engines do you consider Semantic Scholar’s main competitors?
The ones I tend to think about are PubMed and Google Scholar. In other disciplines, there’s Scopus. And Chan Zuckerberg acquired a search engine called Meta, which they’re getting ready to launch publicly.
But I don’t really think of these as competitors. We have put in open source various capabilities and we share our data. And we’ve actually created a cooperative group called Open Academic Search, and we have people on there from Baidu, from Google, and elsewhere. If there’s any competition here, it’s just healthy in pursuit of better scientific information.
I read that Semantic Scholar is concentrating on more accessible abstracts. Is the goal to cross over into more of a lay audience?
Very much so. The world would be a better place if we were all better educated about science. What’s a better way — particularly in this world of fake news, and debates, whether they’re about climate science or vaccines — than allowing educated consumers access to primary sources?
If you’re going to talk about science, perhaps you should be scientific.
Even if you’re not a medical researcher, you might be looking up information on your knee or your back or, God forbid, your heart. People used to look up information on PubMed, and now they’re finding that Semantic Scholar is more accessible.
You’ve concentrated so far on biomedical research. Do you plan to push into the social sciences and humanities too?
Our trajectory has been to cover computer science, neuroscience, then biomed. In 2019 — we haven’t yet announced this, but I’m happy to share it in the context of this interview — we’re planning to do a major expansion with the goal of ultimately covering the social sciences and the humanities as well.
Because of impact factors, search engines are inextricably linked to rankings of individual professors, departments, and institutions. You’ve been a professor. How do you feel about that kind of power being bound up in this technology?
I think it’s a problem, but it’s an inescapable problem in the sense that people are hungry for data, imperfect as that may be. I have sat in a lot of faculty meetings where Google Scholar’s raw citation data was used. There were caveats and provisos, but look, this is what we’ve got. Semantic Scholar’s highly-influential-citation metrics is part of the solution. It’s far from perfect, but we think it’s a lot more indicative of actual influence than raw citations.
For example, we don’t include self-citations. I think most people would agree that if you’re citing yourself, which people do with alarming frequency, that’s not a basis for saying you’re an influential scholar.
There are both for-profits like Google and Elsevier in the search-and-ranking game, and nonprofits, albeit some very wealthy nonprofits, like your institute. Is someone going to make a big profit in scholarly search engines?
I may be naïve, but I don’t think anyone’s going to make a lot of money. It’s noteworthy that even Google is not monetizing their effort. They don’t show ads on Google Scholar.
I’ve known you since you were a budding computer-science wonk in high school and college. Could you imagine then how much AI has changed the world in intervening decades?
I got into AI in high school with the book Gödel, Escher, Bach. So in a funny way, our hopes were higher back then. Thirty years of working on these projects has taught me a measure of humility and a measure of immense admiration for the human mind. When I was 18, I thought that by now we would have machines with human-level intelligence.
You’ve discussed the ways in which we both underestimate and overestimate, or hype, the powers of AI. What are some of the myths and exaggerations you see surrounding it?
The biggest myth and exaggeration comes from Elon Musk, who says that AI is an existential threat to humanity. I would invite him to use Semantic Scholar to search for any scientific evidence to back up that audacious claim. Claims like, with AI, “we are summoning the demon.” If you’re going to talk about science, perhaps you should be scientific.
What are your greatest fears about AI and where do you see its greatest potential?
My greatest fear is AI’s impact on jobs. Particularly employment of the most vulnerable people in society, people who didn’t finish high school or college. That’s a very real and imminent issue.
On the positive side, the most tangible thing is AI’s ability to save lives — reducing transportation accidents, curing diseases, confronting climate change, enabling scientists, and in hospitals improving care.
What advice would you give computer- or data-science undergraduates?
I would remind them that this is a fast-moving arena, so to focus on learning the fundamentals — the math, the computer science, the statistics. Those are invariant and the basis of all this. The latest software framework is going to change three times between when you’re an undergrad and when you’re a seasoned professional 10, 15 years later. So get your fundamentals right.
This interview has been edited for length and clarity.
Alexander C. Kafka is a Chronicle senior editor. Follow him on Twitter @AlexanderKafka, or email him at alexander.kafka@chronicle.com.
