Teaching technology to resolve conversational conflict
An EU-funded project has produced award-winning research on the way people naturally avoid arguments in conversation, with researchers hoping that better understanding could help social media platforms foster more positive discussions online.
© JEGAS RA #256571434, source: stock.adobe.com 2019
When the World Wide Web was first invented 30 years ago, early users expected the internet to bring humans together, enabling discussion between people who might never meet in real life. But today, those discussions have become angry and polarised. On social media, people use language as a weapon especially against those they consider their political adversaries.
As technology companies face growing pressure to tackle the ways their platforms are dividing societies, many hope artificial intelligence can foster less volatile online conversation. But researchers believe there is a lot more to learn about human dialogue before automatic content moderation can be effective.
The EU-funded MULTISIMO project made a major effort to fill that knowledge gap by studying face-to-face conversations that could evolve into arguments. In just two years, the project produced award-winning research that provides new insight into the way people intuitively use speech and body language to try to resolve conflict.
If you hope to defuse conflict, I think its important to first understand what successful communication looks like, says project coordinator Carl Vogel, of Trinity College Dublin in the Republic of Ireland. Using state-of-the-art equipment and research techniques, we analysed group dialogue and our findings will be indispensable for technologies trying to echo human communication.
Detecting personality types
By organising 46 trial participants into pairs and asking them to play a version of the game-show Family Feud, researchers watched as the groups tried to agree how best to answer a series of questions. Under the guidance of three facilitators, the teams conferred with one another as their body language was scrutinised to see how speech, facial expression, eye contact and gestures can defuse disagreement.
As well as learning more about human conversation, the MULTISIMO project aimed to develop a system that could automatically detect personality types. This is important because imagine you are designing a bot to moderate dialogue on social media, for example, says Vogel. You would want that bot to be able to detect when one user is more dominant so it could adopt strategies to facilitate better conversation.
To design this system, the project team first had to learn how humans perceived each other. They enlisted 13 observers to rate each of the players level of dominance. The first five were asked to make their decisions by watching a video of the entire experiment. The other group were asked to assume the role of bots, making snap judgements by just listening to seconds-long snippets of the footage.
Maria Koutsombogera, research fellow within MULTISIMO, says: Humans show a remarkable ability to make quick and accurate assumptions about other humans with a minimal amount of information about them. In a similar manner, bots can be trained to automatically deduce personality types from short audio excerpts.
Making chatbots more human
MULTISIMOs research might offer insights on how technology can ease tensions between social media users, but it can also help make chatbots or avatars conversation style more human. The teams conclusions on body language will be especially useful when chatbots eventually evolve into holograms.
Additionally, in an era of fake news, a better understanding of human interaction could help fact-checkers differentiate between staged and authentic videos spreading online.
Finally, footage of the trial groups playing Family Feud filmed on 360-degree cameras have been shared on the projects website, so they can be easily accessed by other academics. To make this data available is extremely important, says Vogel. Its expensive and extremely time-consuming to do this type of research. While we studied the problems were interested in, we also want to help other people pursuing research in this same field.
This means Vogel expects more innovative conclusions to emerge from this dataset. He says: Youll see this project as a slow burn because other people will come asking questions that we havent even imagined.
The experiment resulted in ground-breaking research on dominant personality types that won the Best Paper Award at an info-communications conference in Hungary. The team also produced nine peer-reviewed research papers and shared their findings at six conferences. MULTISIMO received funding from the EUs Marie Skłodowska-Curie actions programme.