A model that can generate humorous versions of existing headlines

Over the past decennary oregon so, machine scientists person developed a increasing fig of computational models that tin generate, edit and analyse texts. While immoderate of these models person achieved singular results, immoderate aspects of quality connection and connection person proved peculiarly hard to replicate computationally.

One of these aspects is humor, the human ability to accidental oregon constitute things that are funny. Humor is simply a subtle and inherently quality quality; thus, reproducing it successful machines is acold from an casual task.

Researchers astatine University of Helsinki person precocious attempted to artificially replicate wit successful machines, by processing a model that tin crook existing quality headlines into humorous ones. This model, archetypal introduced successful a insubstantial pre-published connected arXiv and presented astatine the 12th International Conference connected Computational Creativity (ICCC 2021), was trained to analyse headlines successful an existing dataset and regenerate words successful them to springiness them comical oregon amusing qualities.

"Automated quality procreation has go a large involvement for quality agencies," Khalid Alnajjar and Mika Hämäläinen, the 2 researchers who conducted the study, wrote successful their paper. "Oftentimes, headlines for specified automatically generated news articles are unimaginative, arsenic they person been generated with ready-made templates. We contiguous a computationally originative attack for headline procreation that tin make humorous versions of existing headlines."

The caller insubstantial by Alnajjar and Hämäläinen draws inspiration from a erstwhile work by 3 researchers astatine University of Rochester and Microsoft Research AI, who introduced Humicroedit, a dataset containing implicit 15,000 annotated quality headlines. In this study, the researchers identified strategies for making headlines comic that are commonly utilized by humans, which they recovered to beryllium aligned with existing theories of humor.

The squad astatine University of Helsinki devised a exemplary that uses immoderate of these strategies to alteration non-humorous headlines and marque them much amusing for readers. To bash this, it tries to find comic substitutes for immoderate of the words successful existing headlines.

Two examples of the headlines generated by the researchers' exemplary are: "Trump eats the incorrect Lee Greenwood connected Twitter" and "U.S. says Turkey is helping ISIS by combing Kurds successful Syria."

To measure the effectiveness of their model, Alnajjar and Hämäläinen utilized it to alteration 83 headlines randomly selected from the Humicroedit dataset and marque them much humorous. Subsequently, they asked reviewers connected a crowd-sourcing level to supply their feedback connected whether they recovered the headlines generated by the exemplary comic oregon not.

Overall, the researchers recovered that the humorous headlines produced by their exemplary were comparable to those generated by humans connected respective levels. In addition, connected average, they recovered that quality evaluators sourced online considered the headlines produced by their strategy comic 36% of the time. If the model is improved further, it could yet assistance media agencies and journalists to travel up with caller comic headlines for news articles.

"As the champion headlines produced by our strategy for each archetypal header can, connected average, scope to a quality level successful presumption of astir of the factors measured successful our evaluation, an contiguous aboriginal absorption for our probe is to make a amended ranking mechanics to scope the maximum capableness of our system," Alnajjar and Hämäläinen concluded successful their paper. "Perhaps specified ranking could beryllium learned by grooming a agelong abbreviated word representation (LSTM) classifier connected wit annotated corpora."

---

---

© 2021 Science X Network