Are You a Victim of 'TL,DR'?

Related articles

Until a few moments ago, I, too, had been a victim of "tl,dr." In fact, I had the disease but did not know its name. Tl,dr is an acronym to “Too long, didn’t read.” Admit it, you have suffered from the same malady, but help is on the way.

Truthfully, writing about science and health often involves more reading than writing. With the infodemic on the COVID-19 pandemic underway now for ten months, there are an enormous number of articles to sift through. A new and useful AI tool makes the search a bit easier.

SPOILER ALERT How I make “the sausage” of my writing follows. 

As a science “journalist,” I can get feeds of upcoming articles across a wide range of topics. Each piece competes with the others for my attention with a snappy line or two, which entices more than inform. Those that have piques my interest lead to a page description that can lead to the articles themselves if I am still awake and paying attention. As a result, I read far more science now than at any time when I was an active clinician. For scientists to keep up, you have to read at least a paper’s abstract, the author’s summary; and those can be a full page long. Bottom line, like the frog-prince, you have to read/kiss many abstracts to find a prince of information.

A new software tool, aptly named TLDR, takes the abstract and generates a one-sentence summary – several paragraphs collapse into a single thought. It was released this week on an academic search engine, Semantic Scholar. If you have a scientific article hanging around, you can try the tool out yourself here. The tool currently works best on artificial intelligence papers, but it is a matter of time before it is rolled out for other disciplines.

“I predict that this kind of tool will become a standard feature of scholarly search in the near future. Actually, given the need, I am amazed it has taken this long to see it in practice.”

Jevin West, information scientist University of Washington

The AI builds upon natural-language processing where thousands of papers act as examples to train the algorithm and then are run through a test set to look at the results. The goal is to compact the information into a salient sentence without sacrificing its accuracy. As one might suspect, giving the algorithm more information, abstract, introduction, and summary resulted in a slightly more informative sentence, but with a great deal more compression – reducing the verbiage 47-fold. Given the digital nature of publishing, this would not be too difficult to automate for journals. The authors note that introductions and summaries are often hidden behind pay-walls, making their system a little bit less informative but still crunching the words; in this instance, 7-fold.

The words used in scientific articles vary with the discipline, and the research group has developed algorithmic variations for additional fields. They have indicated that biomedicine “is likely to come first.” Of course, all of this work will be monetized licensed to publishers and may become a personalized service for individual scientists.

Bottom-line, this would be a great tool for its intended purpose, to quickly skim looking for relevant material. It would still require reading the scientific paper but might quickly separate the chaff from the wheat. It would also eliminate those cutesy lines by academic, public relations personnel who create those science feeds for journalists.

Note: I couldn’t help myself but submitted this entire article as a summary to the tool; here is what I got back:

“Semantic Scholar: An AI-based search engine for reading scientific articles.”

I would have preferred a bit more, but I love my words and find it hard, in the words of Stephen King, “kill my little darlings.”

 

Source: tl;dr: this AI sums up research papers in a sentence Nature DOI: 10.1038/d41586-020-03277-2