This AI System Will Help You Avoid Spoilers Hidden In Online Reviews

Some researchers have been working on an AI system that could identify spoiler in those online reviews about TV shows and books.

As per the author Ndapa Nakashole, spoilers are all over the Internet. They are even more common on the various social media platforms. As an Internet user himself, he also understands the pain and destruction of a spoiler on a particular experience.

Currently, several websites allow users to flag their online posts manually that become kind of spoiler warning signs. However, that is not always the case. Therefore, a group of researchers has been working on an AI tool that utilizes neural networks to detect spoilers automatically. They called it SpoilerNet.

On the theory level, the researchers would like to understand better the way people write those spoilers. Plus which common knowledge and linguistic patterns people use to write those sentences with spoilers.

A new dataset

That tool could be used in making a browser extension that helps people avoid spoilers. To train SpoilerNet, the researchers created their datasets of 1.3 million reviews of books with book reviewers' spoiler tags. These tags come with sentences that hide spoilers in a link to view spoiler. They collected these reviews from Goodreads.

They found that sentences with spoilers tend to appear in the end part of the review. However, different users have different standards when tagging spoilers, so the neural networks must be carefully calibrated to take that into account.

Mengting Wan, the first author of the study, said that this is the very first dataset with annotations for spoilers at this granularity and scale. Wan also added that a word in different contexts might have different meanings. Like "red" is a color in one review, but it could be a character's name in another. Understanding and identifying those differences is very challenging.

The accuracy rate

The searchers use 80% of Goodreads' reviews to training SpoilerNet, running them through several neural networks layers. As of the moment, SpoilerNet could detect spoilers with 89-92% accuracy. The searchers also used a dataset of 16,000 reviews ( single-sentence ones) of 880 TV shows to train SpoilerNet. The tool reaches 74-80% spoiler detecting accuracy.