‘Exaggeration Detector’ Could Lead to More Accurate Health Science Journalism

It would be an exaggeration to say youll never ever once again read a news short article overhyping a medical development. However, thanks to scientists at the University of Copenhagen, finding embellishment might one day get more manageable.

Read the full paper here: https://arxiv.org/pdf/2108.13493.pdf.

The paper comes amid a pandemic that has fueled demand for reasonable, precise info. And social media has actually made health misinformation more extensive.

Research like Wright and Augensteins could speed more precise health sciences news to more individuals.

In a new paper, Dustin Wright and Isabelle Augenstein explain how they utilized NVIDIA GPUs to train an “exaggeration detection system” to identify overenthusiastic claims in health science reporting.

A Sobering Realization

To be sure, putting this research study to work would need investment in production, functionality and marketing, Wright states.

Initially, seeing the strength of claims made in press releases and the scientific papers they summarized. Then, identifying the level of exaggeration.

That makes the issue of detecting exaggeration in health sciences press releases an excellent “few-shot knowing” usage case.

News release convey info. But they also need to be strong adequate to create interest from reporters. Not constantly easy.

They then broke the issue of finding exaggeration into two associated concerns.

Such technology might permit scientists to spot exaggeration in fields with a restricted quantity of know-how to classify training data.

Wright and Augenstein broadened on the idea to train a PET design to both detect the strength of claims made in press releases and to examine whether a news release overemphasizes a documents claims.

If they respond to “yellow,” the instructor understands they understand what they see. If not, the teacher knows the student needs more aid.

AI-enabled grammar checkers can already assist writers polish the quality of their prose.

Each set, or “tuple,” has annotations from experts comparing claims made in the papers with those in corresponding news release.


Researchers Dustin Wright and Isabel Augenstein created complementary pattern-verbalizer sets for a primary job and an auxiliary task. These pairs are then utilized to train a machine learning design on information from both tasks (source: https://arxiv.org/pdf/2108.13493.pdf).

Instructors PET

You can capture Dustin Wright and Isabella Augenstein on Twitter at @dustin_wright37 and @IAugenstein. Read their full paper, “Semi-Supervised Exaggeration Detection of Health Science Press Releases,” here: https://arxiv.org/pdf/2108.13493.pdf.Featured image credit: Vintage postcard, copyright expired..

The co-authors started by curating a collection that included both the releases and the papers they were summarizing.

That task falls on the press offices of universities and research study organizations. They utilize authors to produce news release– brief, news-style summaries– relied on by news outlets.

” Part of the reason things in popular journalism tend to get sensationalized is some of the journalists do not check out the documents theyre discussing,” Wright states. “Its a little bit of a sobering awareness.”

Wright and Augensteins contribution is to reframe the problem and apply a novel, multitask-capable version of a strategy called Pattern Exploiting Training, which they called MT-PET.

” Whenever I tweet about stuff, I believe, how can I get this tweet out without exaggeration,” Wright says. “Its hard.”

Shot On

Wrights also reasonable about the human factors that can cause exaggeration.

Its not the very first time researchers have put natural language techniques to work identifying buzz. Wright indicates the earlier work of associates in scientific exaggeration detection and false information.

The training procedure depends on cloze-style expressions– expressions that mask a keyword an AI needs to fill– to guarantee it comprehends a job.

They then ran this information through an unique sort of PET design, which discovers much the method some second-grade students discover reading understanding.

dont have the time to dig much deeper.

A teacher may ask a trainee to fill in the blanks in a sentence such as “I ride a big ____ bus to school.”

As an outcome, Wright and Augenstein were able to demonstrate how MT-PET surpasses PET and supervised knowing.

These 563 tuples offered them a strong base of training data.

Its tough to blame them. Numerous journalists require to summarize a great deal of details quickly and frequently

The scientists trained their designs on a shared computing cluster, utilizing 4 Intel Xeon CPUs and a single NVIDIA TITAN X GPU.

University of Copenhagen researcher Dustin Wright.

Few-shot knowing methods can train AI in areas where information isnt plentiful– there are only a couple of products to find out from.

One day, comparable tools might assist journalists sum up brand-new findings more accurately, Wright states.

Leave a Reply

Your email address will not be published.