In part 1 of this blog we discussed AI’s terminology and limitation. Now we can dive straight into the Peace Machine.
Timo Honkela describes three concepts in his book called the Peace Machine. Each of them are ideas on how artificial intelligence (AI) could help human communication.
- Meaning negotiating machine is about recognizing that people use different words to mean the same concepts and aid them to get to the same page.
- The feeling machine identifies feelings and helps people to understand others as we as themselves better.
- The million people meeting splits discussions from mass events to small groups that have the discussions simultaneously and then collects the information using data mining.
There’s two things that Timo Honkela is very clear about in his book. Firstly, the peace machine is not a black box put on the table at crisis resolution negotiations. And secondly, none of the three concepts can be implemented with today’s technology. With that in mind, how far could we get with today’s AI technology?
The underlying technology of meaning negotiation is tensor analysis. Tensors are multidimensional arrays that in our use case would hold information about concepts and their connections.
To use meaning negotiation in a generic discussion we would need to have information about the individuals, their knowledge, their languages and maybe even their values. And then we should be able to connect that information to the subject of the discussion to understand the possible differences in meaning.
The amount of data and calculation power required is simply out of today’s scope. But could we build a subset of meaning negotiation? As mentioned earlier in part 1, AI works best in scenarios that are well defined. So, we could change the setting from plugging into a conversation into analysing content that we already have.
Honkela describes in his book an example of a case where research had been done on how US democrats and republicans talk about healthcare. From the large amount of speeches it can be seen that the two parties connect very different ideas and terms to one subject.
While this kind of analysis is pretty far from translating meanings from one individual to another, it can be useful in many practical solutions. Especially in politics it’s important to look behind the rethorics to understand what are the true differences between political parties.
Sentiment analysis is at the same time the most and the least advanced of the three concepts. While there are several services and frameworks that can be used to do sentiment analysis right now, there are very few practical applications that could take use of it.
The current services take use of machine learning. They have been fed with a large amount of text data together with emotion labels. With this information the AI system can identify feelings from the text.
However, there are several reasons why sentiment analysis is far from perfect. Sentiment is hard to identify from written text even for humans. It lacks all the non-verbal parts of the communication, like the tone of voice and the rhythm. These restrictions make it practically impossible to identify, for example, sarcastic notes or irony.
Another restriction to the sentiment analysis comes from the context. AI systems don’t have it. If the teaching data is too general some topics might be rendered to certain feelings regardless of what has been said about them.
For example, “war” is a difficult word. Usually discussions about war include sadness or frustration. If we analyse a text that is very specific about war and peace there might be a lot of hope and happiness related to the sentences even if they contain the word “war”. AI would probably get confused and find sadness and frustration just because it sees the word “war”.
An important finding is that sentiment analysis tools should be taught using context specific data. If we want to analyse text about peace, war and security, our teaching data should also talk about peace, war and security.
This far we have talked about technical restrictions. But when feelings are concerned there’s always the human factor as well. People don’t like their feelings to be labeled. Think about a machine that says “I noticed that you are getting a bit angry”. This probably would make you even angrier.
When Honkela talks about feeling analysis, he talks about learning. Teaching us about our own feelings and reactions as well as how our actions affect others. And that’s where sentiment analysis for peace tech is at its best. As a tool of reflection.
Reflection is never easy for anyone. You have to question your own decisions and feelings. And there a machine just might be the right tool. Machine is neutral partner to talk to. So, you might be more acceptive for ideas that the machine gives you. And for that, the current services just might be enough.
Million people meeting
A million people meeting is simple enough as an idea. Our political system is based on chosen few who make the decisions for the rest of us. If people could discuss topics simultaneously in smaller group we could collect more of people’s collective knowledge to benefit everyone.
The idea is that a machine would mine information from millions of simultaneous discussions. The system would look for common patterns in the opinions and facts and try to summarize the collective knowledge. The machine could also feed the trends found back into the discussion.
How possible is this idea? Gathering a million people to talk about one topic is hard enough. Not to mention getting them to do that in one platform that could reach all the discussions.
Sounds hard, but this is actually happening everyday. On Twitter and Facebook trending topics definitely have millions of simultaneous discussions on-going. So, it seems that the data is already there.
Next we need something called data mining. It combines statistics and machine learning to find patterns in data.
One example of data mining is to find the most commonly used words and how they connect to each other. You might remember tensors from meaning negotiation. That’s what we are looking at here as well.
Based on those connections, through some kind of visualisation, we could show what has been discussed. From this point on we need a human to figure out how the terms and relations connect to the topic discussed. To put it in another way: Machines could visualise the discussion topic and structures, but they couldn’t form opinions or suggestions.
Another approach would be to automatically summarize the discussions. The problem here would be that a million people produce a lot of discussion and it would be hard for a machine to make a compact and meaningful summary out of it. An expert approach would be more likely to find interesting point of views from the data.
Timo Honkela describes an utopia. The Peace Machine just might be part of the future, but taking machines to be a part of human communications have significant risks as well.
Today’s AI technology performs best in specific problems. If we are able to narrow down the questions that we want to answer, we can utilize AI efficiently. But for generic cases humans still outperform machines.