2024.Mar.28

REPORTS

Report on the 3rd BAIRAL Research Meeting for 2023
“Understanding and Generating Humor by Computers: Toward Smooth Communication between Human and AI”

The 3rd BAIRAL was held online on July 31, 2023. At this session, Kohtaro Tanaka, a Master’s Student at Graduate School of Information Science and Technology at The University of Tokyo, was invited to speak on the theme of “Understanding and Generating Humor by Computers: Toward Smooth Communication between Human and AI.”

Humour is essential in communication, but research on its understanding and generation by computers has yet to progress due to its high subjectivity and the need for background knowledge. Tanaka has focused on memes (“bokete”) that effectively convey messages on social media and is researching to create AI to generate and understand funny memes. In this presentation, he discussed his research, the humour capabilities of large-scale generative AI, and the current state and challenges of computer humour processing.

In his research on humour evaluation methods and datasets, Mr. Tanaka created a consistent meme dataset. The creation method involved human annotation using Best-Worst Scaling (BWS), calculating the degree of agreement in the evaluations, and excluding memes with inconsistent labels. He proposed a module that extracts mismatches between images and captions to create an evaluation model based on humour theory. This model performed better than other comparative models in the humour evaluation task but it has not yet achieved complete humour understanding.

In recent years, large-scale language models such as GPT-4 have been attracting attention, and these models are beginning to have the essential ability to understand humour. When GPT-4 was asked to explain the funniness of a joke, it correctly extracted keywords such as time leap and the Honnoji Incident and interpreted the images by the text and background knowledge. However, humour’s understanding of the latest trends and current affairs still needs to be improved. It also needs to be resolved that it becomes difficult to interpret images from text and background knowledge when the situation becomes complex. Furthermore, considering individual differences and generating content that combines image understanding, text understanding, and background knowledge are issues that need to be sufficiently discussed in humour research and need to be addressed in the future.

The discussion raised a question about the propriety of generating content that constitutes defamation as humour. As AI models become more advanced, ethical discussions on what kind of content should be generated become important, and technologies for filtering out dangerous content, such as how to make bombs, are already being researched and implemented in language models. Mr. Tanaka stated that even for image content, it is necessary to develop methods to discriminate ethically appropriate content. On the other hand, joke content, such as a cat swimming, is funny precisely because there is a certain amount of mismatch. The ethical tolerance for such content varies by community. Therefore, when implementing AI models in society, it is necessary to develop models that can be designed to match the ethical sense of each community.

As for the application examples of humour in communication, Tanaka gave a scenario in which humans and robots cooperate to perform tasks. For example, when doing household chores, the robot can recognise the necessary functions from the conversation and smoothly move on to the work. Existing research has shown that productivity increases through humorous communication when multiple people work together. If humorous communication can be realised in scenarios where humans and robots collaborate, it may lead to smoother work.

In the discussion about robots changing human behaviour, there was a question about whether humour also has the potential to change human behaviour, such as humans behaving in a way that is compatible with machines in order to obtain good results from machines. Tanaka presented both the possibility of positive changes, such as having robots serve as practice partners for communication through humour and the possibility of negative changes, such as humans not finding originality in humour and refraining from using it in communication. Participants pointed out the possibility of humans becoming dependent on robot activities, the concern that humans may stop using their communication skills, and the possibility of changes in economic value and society as humans change with the spread of AI. In addition, perspectives such as AI-generated diverse humour being a reference for people who are not good at communication and the possibility of reducing divisions through humour in today’s society where there are many divisions provided an opportunity to think about communication not only between humans and AI but also between humans.