> No one will be allowed to read or download work from the repository, because that would breach publishers’ copyright. Instead, Malamud envisages, researchers could crawl over its text and data with computer software, scanning through the world’s scientific literature to pull out insights without actually reading the text.
I find this totally unconvincing. The average scientific article isn't any good, and the NLP algorithms that do tasks like this are even worse.
For scientific literature to be useful to those of us outside the academy, we need to be able to see the full document - what methodology was employed, what assumptions were made, how the data was gathered - just to be able to gauge whether the authors had any idea what they were doing. Ideally we would also be able to search over documents, explore the citation tree in some sort of UI, and access articles in multiple formats (sometimes you want a PDF, sometimes you want plaintext, and I imagine that the mathy-types might like LaTeX source).
I applaud Malamud's efforts to overcome this problem, and I hope his results prove me wrong. I just think it's sad that we have to resort to hacks like this to overcome what is obviously enormous scam that is stealing our tax dollars and stifling academic and economic creativity.
> The average scientific article isn't any good, and the NLP algorithms that do tasks like this are even worse.
In my day job, I'm often tasked with implementing algorithms from recently published physics papers. In order to do so, I normally have to read through at least 10 related papers (both cited papers and cited by papers) in order to have a clear idea in my head about what is going on. Even then, I often have to discuss what I have read with 2 or 3 other people and try out many different approaches before we finally understand what the paper meant.
This is because few papers include enough details to reproduce what they are doing. Many of these papers have mistakes or typos in their equations. Many of the math equations are also under specified. For example, an equation may list a sum over an index, but then in the text there may be a whole paragraph that describes what that index means (there is nothing wrong with this, but it would make it hard for AI to "just use the equations").
At the end of this whole process, which can take a week, about half of the time we choose not to implement the algorithm for one of several reasons:
* The authors misrepresented their work and it does not perform as well as claimed (e.g. the chosen examples are special cases that make their approach look better than the state-of-the-art approach)
* The work is not reproducible from the information in the paper
* The amount of work to implement it is far greater than an initial reading of the paper would suggest, due to additional details that were left out of their discussion
So, all of this is to conclude that understanding scientific papers is at the very limit of human ability for a group of PhDs in the field. I do not think that until we have much more powerful AI that it has any hope of making sense of this mess.
edit:
P.S. I am guilty of these same mistakes when publishing. I understand the deadlines and pressure to publish that leads to these issues. It is a huge amount of work to fully document and publish all the details needed to reproduce some new algorithm .
This sums up my experience pretty well. I used to work in a field of physics where a lot of projects were not open-sourced, which is probably your experience too. Now, I work in computer vision, where there is a lot of open sourced work and I _still_ find myself coming up against most of these problems when dealing with implementations of cutting edge results.
>In my day job, I'm often tasked with implementing algorithms from recently published physics papers. In order to do so, I normally have to read through at least 10 related papers (both cited papers and cited by papers) in order to have a clear idea in my head about what is going on. Even then, I often have to discuss what I have read with 2 or 3 other people and try out many different approaches before we finally understand what the paper meant
This post makes me feel slightly better about how often I come away from reading a paper feeling like I only have a shallow understanding of the content.
Yes. I worked with a company in India looking to mine biology papers with information on genes and automatically populate a database from that. Given the various ways in which people wrote (required to pass anti-plagiarism checks as well as different writing styles) it turned out that any kind of automated annotation was rife with errors. Given that it was supposed to be used for developing drugs they dropped it in favor of hiring Master students part time at $200/month and annotating manually.
Great anecdote. I'm increasingly of the opinion that "pay some humans to do it" is the most underrated data product engineering pattern out there. It's well known that FAANG invest heavily in human annotators (it's not a coincidence that Mechanical Turk was developed at Amazon) and it's unclear why anyone else building data products shouldn't have to.
Yeah, this is silly. Elsevier and the like should be burned down for crimes against humanity. I mean, it generally is illegal to price gouge during famines.
Elsevier is a useless profiteer, but let's please remember that scientists are voluntarily submitting their articles to Elsevier journals, and the funding agencies (and universities) are doing nothing to stop them. I find it a little frustrating to read the wailing from institutions like UC (where I used to work) about subscription prices when they are part of the problem in the first place.
> A trigger for this mission came from a landmark Delhi High Court judgment in 2016. The case revolved around Rameshwari Photocopy Services, a shop on the campus of the University of Delhi. For years, the business had been preparing course packs for students by photocopying pages from expensive textbooks. With prices ranging between 500 and 19,000 rupees (US$7–277), these textbooks were out of reach for many students. In 2012, Oxford University Press, Cambridge University Press and Taylor and Francis filed a lawsuit against the university, demanding that it buy a license to reproduce a portion of each text. But the Delhi High Court dismissed the suit. In its judgment, the court cited section 52 of India’s 1957 Copyright Act, which allows the reproduction of copyrighted works for education. Another provision in the same section allows reproduction for research purposes.
It would be interesting to see the server getting hacked with all the papers in the open, and then the researchers sending a sorry email to the publishers.
> Over the past year, Malamud has — without asking publishers — teamed up with Indian researchers to build a gigantic store of text and images extracted from 73 million journal articles dating from 1847 up to the present day.
Maybe I missed it, but the article doesn't seem to explain exactly how Malamud's group compiled its database.
Throttling is a major problem with the naive approach of throwing wget on a publisher site. The publisher detects a bot on its network downloading everything in sight and either slows data transfer to a trickle or just shuts down access to it.
The publishers may not win on copyright, but they may try to make a case based on criminality if Malamud's team actively took steps to circumvent throttling and defeat the defenses of the hosting sites. Especially if the publisher knows what to look for in its logs.
I'm a little late to this party, but here's a mandatory research plug:
My lab works in the area of evidence based medicine. My research focuses on Randomized Controlled Trials, and the overall goal is to automate (fully or partially) the meta-analysis of medical interventions. To that end, we collected a dataset of intervention pairs, statistical significance findings about these pairs, and a minimal rationale supporting the significance finding.
We're working on collecting more complete annotation. We hope to facilitate meta-analysis, evidence based medicine, and long document natural language inference. We might even succeed (somewhat) since this is a very targeted effort, as opposed to something more broad.
This is good, more because it will make searching for information easier than because it will avoid copyright. Searching through journal articles to find information about a given topic is very hard, even at a university with more or less universal library access to the online literature.
Authoring papers of uneven detail and quality and then publishing them in whatever prestigious journal that will have them is a terrible way to document the progress of science.
Instead, new results should be added to an open science information base, fully linked to all previous results that bear upon the new results, either in support or contradiction. This is what some of the bioinformatic data bases attempt to achieve by scanning articles, but it would be better to omit the article step.
Readit News
I find this totally unconvincing. The average scientific article isn't any good, and the NLP algorithms that do tasks like this are even worse.
For scientific literature to be useful to those of us outside the academy, we need to be able to see the full document - what methodology was employed, what assumptions were made, how the data was gathered - just to be able to gauge whether the authors had any idea what they were doing. Ideally we would also be able to search over documents, explore the citation tree in some sort of UI, and access articles in multiple formats (sometimes you want a PDF, sometimes you want plaintext, and I imagine that the mathy-types might like LaTeX source).
I applaud Malamud's efforts to overcome this problem, and I hope his results prove me wrong. I just think it's sad that we have to resort to hacks like this to overcome what is obviously enormous scam that is stealing our tax dollars and stifling academic and economic creativity.
In my day job, I'm often tasked with implementing algorithms from recently published physics papers. In order to do so, I normally have to read through at least 10 related papers (both cited papers and cited by papers) in order to have a clear idea in my head about what is going on. Even then, I often have to discuss what I have read with 2 or 3 other people and try out many different approaches before we finally understand what the paper meant.
This is because few papers include enough details to reproduce what they are doing. Many of these papers have mistakes or typos in their equations. Many of the math equations are also under specified. For example, an equation may list a sum over an index, but then in the text there may be a whole paragraph that describes what that index means (there is nothing wrong with this, but it would make it hard for AI to "just use the equations").
At the end of this whole process, which can take a week, about half of the time we choose not to implement the algorithm for one of several reasons:
* The authors misrepresented their work and it does not perform as well as claimed (e.g. the chosen examples are special cases that make their approach look better than the state-of-the-art approach)
* The work is not reproducible from the information in the paper
* The amount of work to implement it is far greater than an initial reading of the paper would suggest, due to additional details that were left out of their discussion
So, all of this is to conclude that understanding scientific papers is at the very limit of human ability for a group of PhDs in the field. I do not think that until we have much more powerful AI that it has any hope of making sense of this mess.
edit: P.S. I am guilty of these same mistakes when publishing. I understand the deadlines and pressure to publish that leads to these issues. It is a huge amount of work to fully document and publish all the details needed to reproduce some new algorithm .
This post makes me feel slightly better about how often I come away from reading a paper feeling like I only have a shallow understanding of the content.
Good job India.
Dead Comment
Maybe I missed it, but the article doesn't seem to explain exactly how Malamud's group compiled its database.
Throttling is a major problem with the naive approach of throwing wget on a publisher site. The publisher detects a bot on its network downloading everything in sight and either slows data transfer to a trickle or just shuts down access to it.
The publishers may not win on copyright, but they may try to make a case based on criminality if Malamud's team actively took steps to circumvent throttling and defeat the defenses of the hosting sites. Especially if the publisher knows what to look for in its logs.
My lab works in the area of evidence based medicine. My research focuses on Randomized Controlled Trials, and the overall goal is to automate (fully or partially) the meta-analysis of medical interventions. To that end, we collected a dataset of intervention pairs, statistical significance findings about these pairs, and a minimal rationale supporting the significance finding.
Since these annotations were collected on full text PubMed Open Access articles, we can distribute both the articles and the annotations: https://github.com/jayded/evidence-inference ; paper: https://arxiv.org/abs/1904.01606 ; website: http://evidence-inference.ebm-nlp.com/
We're working on collecting more complete annotation. We hope to facilitate meta-analysis, evidence based medicine, and long document natural language inference. We might even succeed (somewhat) since this is a very targeted effort, as opposed to something more broad.
Authoring papers of uneven detail and quality and then publishing them in whatever prestigious journal that will have them is a terrible way to document the progress of science.
Instead, new results should be added to an open science information base, fully linked to all previous results that bear upon the new results, either in support or contradiction. This is what some of the bioinformatic data bases attempt to achieve by scanning articles, but it would be better to omit the article step.
Deleted Comment