While many expect chat bots to become a huge part of our AI-dominated future, they have been rather underwhelming so far. Behind almost any online bot are drop-down menus, slightly tweaked to appear as questions. This makes the whole experience way too robotic. In order to reach new heights, chatbots need new skills. They need to get cognitive abilities to reason with people. And these abilities might not be as far as you think.
Facebook has been a major player in this field for a long time. They already launched their chatbot assistant, simply known as M. Yes, it’s true that it’s still humans who do all the difficult tasks, but it’s a step in the right direction. Facebook also created an open-source tool, designed to teach bots.
Yesterday, the social media giant’s researchers from the FAIR lab showed a new way of training chatbots to be better negotiators. The scope is pretty narrow so far. It concentrates only on one negotiation scenario, but it could pave the way for more capable bots in the future. The testing results were quite interesting. First, people often mistakenly thought they were talking to humans. Bots also managed to develop complicated negotiation strategies, similar to game theories like the prisoner’s dilemma. The AI reached this impressive level without any help from humans.
AI Bots learn complicated strategies
The technology, used in the research, is a neural network that learned how to reason with people from a dataset of recorded negotiations. In this particular case, the single negotiation scenario was ‘multi-issue bargaining’. Researchers asked a group of people from Mechanical Turk to divide a number of items between them. Each participant valued the items differently and was asked to maximize the number of points they received.
This was the foundation of the neural network, but there was more to it. The researchers added an innovative technique, called ‘dialogue rollouts’. They asked the bots to ‘think ahead’ by simulating the course of future negotiations and choose the best strategy to get what they want. People often use these techniques while playing chess or similar games, but it was the first time they were used for dialogue as well.
The implementation of these rollouts brought ‘significant improvements’, compared to similar experiments. After the training, the bots started negotiations in Mechanical Turk. One of the researchers, Dhruv Batra, says people mostly didn’t notice they weren’t talking to humans. The best results, demonstrated by the AI, was pretty close to the scores the people were getting. ‘Certainly not better, but certainly not much worse too’, Batra said.
While negotiating, the bots even managed to learn how to trick one another. They discovered a good strategy, where they ‘passionately’ pursued an item they really didn’t want, to give it up at the last minute to get a better compromise. Nobody had programmed this strategy, and the researchers say, the bots adopted it themselves.
Far from perfect still
Yet there’s room for improvement, as the chatbots are far from working ideally at this current stage. Their results weren’t consistent, and we are still talking about this one very specific scenario. It’s not clear if they’ll be able to use these skills in different situations.
‘It’s a good step forward, but not a breakthrough’, said Kaheer Suleman, a researcher at AI Company Maluuba, owned by Microsoft. He believes, the main problem here is the fact that the bots got their training data from Mechanical Turk. People on this online crowdsourcing marketplace often want to finish their talks as quickly as they can. They are unlikely to be artistic with their language, says Suleman, whose company also works on chatbots. This means the sentences the bots learn from are really ‘elementary’.
The experiment pretty much confirms this is the case. The bots were indeed using very simple sentences, like ‘I want the plate and the lamp, you get the keys’. There was no subtlety here, as they were very direct in describing what they wanted and how much they wanted it. This is an issue with any conversation, generated by AI. Robots are still robots, and they yet to learn how to make sentences that aren’t robotic.
The distant future
But even with all these flaws, the technological process is moving forward. The massive amounts of user data might help major companies make the things better. The FAIR team is already thinking of different cases the chatbots could be used in. They could haggle, arrange meeting times, and people could program them to try to get the results they want out of the negotiations.
Some believe, though, the only side that will benefit from the new technologies are corporations, not consumers. Since bots train on data, big companies have access to vast amounts of it, says Trim CEO Thomas Smyth. The more data they have, the better their bots will be at negotiating.
But it’s still a distant future. Right now, we are far from seeing what impact chatbots could really have on our lives. Every AI research still remains a work-in-progress, without concrete plans to ever become a product for people to use.
Source: The Verge