Can science writing be automated?

Reminds me of this earlier work from students also at MIT, called SCIGen [1]. You can type in author names and it’ll generate some seemingly coherent text for you as a “research paper”. They even submitted it to a conference.

[1] https://pdos.csail.mit.edu/archive/scigen/

Can deep learning paper writing be automated? :P

These neural network hype articles are barely one step down from Buzzfeed listicles in what they project versus what they claim versus what they have actually done. Scan from a list of really ambitious goals, scan among list of available datasets (God forbid, no scientist should have to actually collect and clean new data pertaining to the problem!), throw a bunch of neural networks architectures at it leveraging cloud compute, and publish a paper based on whatever sticks. The process isn't complete without a title involving the word "deep" and a PR announcement generated by a neural network trained on Buzzfeed articles.

Maybe I'm slightly too harsh. The actual paper in this case is not making crazy grandiose claims, but this article is certainly doing no favors. If this is the quality of research reporting, then it might as well be automated.

Maybe it's because I've seen too much of the masses of automatically generated "articles" that pollute my search results, but I really hope stuff like this doesn't proliferate, because I think it'll just lead to even more of that spam --- and be even harder to separate from the useful stuff.

In general, I have the same opinion of this as I do things like automatically generated comments in code: "If it can be generated by a tool, it's not worth reading."

Could this very article have been written simply by applying some model to the paper text? I doubt it. Look at it, it's full of quotes from the authors, full of commentary and explaining the context to a lay audience (with a healthy dose of self-promotion from the side of the university).

Although research papers do contain some context (on a more specialized level, and with many references to other papers instead of repeating the idea), they are more about the technical approach, which is less useful for direct use in the popular article.

One goal of popular science writing is to inspire people, to tell us why we should be excited about a topic, to see what's going on in fields distant from ours etc. This is far from "research paper in -> pop sci article out".

TL;DR: Summarizing a research paper is not the same as writing an interesting pop sci article about a research project.

It would be interesting if authors used a tool like this, to see if the automated summary matches the point they are trying to make. If the tool gets it wrong, maybe your paper needs improving.

Summarizing scientific articles and writing them are not the same thing. I think this is very interesting work (although I have not read the paper and my knowledge of AI and NLP is very superficial), but the headline really puts me back.

Every scientific paper already contains a summary. It’s called the abstract.

That( there) leaves "science writing" ill defined and as such this( here) might seem to yield a problem which consists NOT of well posedness.

Anyway. Here, a very, very much related conference paper from 2017 which I found last night:

https://dl.acm.org/citation.cfm?id=3107452 "Discovering Inconsistencies in PubMed Abstracts through Ontology-Based Information Extraction"

Found via the Google Scholar profile of Nisansa de Silva:

https://scholar.google.com/citations?user=OgPD66oAAAAJ

Whose works I looked into after finding this:

https://cs.uoregon.edu/area-exam/sensing-puns-its-own-reword...

Because I suspect multilingual Paronomasia, Polysemy & Syllepsis can serve to decrease the minimum description length of highly complex topics.

For example:

The German noun "Begier" ('Desire', but with an element of greed, i.e. 'greed' with the prefix be-, so, literally, "begreed" or possibly "bygreed") sounds almost equal to a a word easily constructible in English, from the stem "beg" (as in "to beg") and the suffix -ia (as in, "Paronomasia"): Beg-ia, or perhaps Beggia. Such phonetic similarity can serve as a powerful mnemonic, something widely exploited in the Torah, and I'll leave explaining that to a Brazilian statistics professor:

https://link.springer.com/article/10.1007/s11841-017-0592-y

(Open Access. Read the PDF version if you want to actually stand a chance of visually parsing the Hebrew/Aramaic characters.)

Note however that some scholars might find the claim given in that paper controversial - mostly because one can't find the proposedly alluding elision (as opposed to repetition) in this list of types/kinds of paronomasia & polysemy from Noegel's Encyclopedia of Hebrew Language & Linguistics:

For Paronomasia:

* Homoeoprophoron

* Homoioteleuton

* Epanastrophe

* Parasonance

* Homonymic paronomasia

* Bilingual paronomasia

* Anagramic paronomasia

* Hendiadic paronomasia

* Rhyme

* Farrago

* Geminate parallelism and clustering

For Polysemy:

* Contronymic Polysemy

* Double entendres

* Antanaclasis

* Unidirectional polysemy

* Janus parallelism (pivotal polysemy)

* Double polysemy

* Bilingual polysemy

* Polysemy clustering

* Numerical polysemy

* Gematria

* Notarikon

* Acronymy

* Acrostics [abacedarius] (also Telestichs and Menostichs)

* Atbash

* Amphiboly (Amphibology)

Edit #1:

Ugh, accidentally hit the submit button before finishing the comment. I'll edit the rest in.

Edit #2:

Still editing stuff in.

Edit #3: OK, there.