Computer scientists have been developing a variety of models that can create, edit, and analyze text over the past decade. Some of these models have produced remarkable results, but some aspects of communication and human language are particularly difficult to reproduce computationally.
Humor is one of these. It’s the human abilityto say or write funny things. Humor is an intrinsic human quality. It is difficult to reproduce it in machines.
University of Helsinki researchers have created a framework to make news headlines more humorous by artificially replicating humor in machines. The model was first presented in a paper that was prepublished on arXiv. It was then used to analyze headlines from an existing dataset, and to replace words to make them funny or amusing.
Khalid Alnajjar, Mika Hamalainen and the other researchers who carried out the study wrote that automated news generation was a growing interest in news agencies. These news articles generated automatically are often unimaginative because they are generated from pre-made templates. This paper presents a computationally inventive approach to headline generation. It can create humorous versions of already-generated headlines.
Alnajjar, Hamalainen’s recent paper draws on , a previous work by three researchers from University of Rochester and Microsoft Research AI. They introduced Humicroedit, which is a data set that contains over 15,000 annotated headlines. The researchers found strategies to make headlines funny that were used frequently by humans. These strategies were aligned with existing theories about humor.
University of Helsinki’s team devised a model to make non-humorous headlines more humorous and amusing. It attempts to find humorous substitutes for certain words in headlines.
The headlines generated by the model include “Trump eats the wrong Lee Greenwood tweet” and “U.S. claims Turkey is aiding ISIS by combing Kurds into Syria.”
Alnajjar & Hamalainen used the model to make humorous headlines from 83 randomly chosen Humicroedit dataset headlines. They then asked crowd-sourcing platform users to give their feedback about whether or not they found the headlines generated using the model humorous.
The researchers concluded that humorous headlines generated by their model were similar to human-generated headlines on multiple levels. They also found that 36% of human evaluators who accessed their model online thought the headlines produced were funny. The model could be improved to allow journalists and media agencies to create new headlines for news stories.
Alnajjar-Hamalainen concluded that the headlines generated by our system for each headline averaged to human level in terms most of the factors in our evaluation. Therefore, an immediate direction for their research is to create a better ranking mechanism to maximize our system’s potential. Perhaps such ranking can be achieved by learning a long-term memory (LSTM), classifier for humor annotated corpora.