Dr. John Quinn, Artificial Intelligence Advisor for UN Global Pulse and the technical lead of the radio content analysis project developed by Pulse Lab Kampala, talks about the development of the tool and next steps. John has a background in machine learning and computational statistics and did his PhD at the University of Edinburgh, looking at machine learning for physiological monitoring. Since 2007, he has also been a lecturer in Computer Science at Makerere University in Uganda.
You have been heavily involved in developing the radio content analysis tool. Can you please explain how it came to be and why radio was chosen as a source of big data?
Radio was always one of the things we were most interested in at Pulse Lab Kampala. When the Lab got started in late 2013, I was the first team member recruited, and colleagues in Pulse Lab New York and Pulse Lab Jakarta were doing a lot of work with Twitter data at the time. There are interesting things that can be done with social media data, but the big problem, especially in Uganda, is the demographic issue. It’s mainly used by young people in cities, people who are affluent, people with smartphones, which makes it a limited source of information on development and humanitarian issues. Radio on the other hand is an incredibly rich source of information and it’s actually one of the main mediums of communication in the country. People phone in to give their opinions, and you get a range of discussion that you don’t easily get from other sources.
There was, therefore, an opportunity to really listen to the voices of the people that we have never been able to do before – although we knew the process of mining radio data was going to involve a lot of work. With most social media analysis, people type in text so it’s relatively easy to analyse. Radio analysis involves speech recognition processes to make sense of what is being uttered. Since there isn’t speech recognition software for Uganda languages, we had to
In general, people think of speech technology as easy because our brains are adapted to listening to language, but the same process is much harder to re-create with computers. You sometimes get a feel of how difficult this really is when you visit a country you have never been before and start trying to listen out for different words. To get our system running, we needed lots of training data with examples of people speaking the language, to be able to train the software to identify discussions about disease, disasters, refugees and other topics of interest to the SDGs.
In your opinion, what was the recipe for success, so to speak, for developing the tool?
One of the major factors was having a great network of partners, particularly the Stellenbosch University in South Africa. Prof. Thomas Niesler and his group have had experience of working with speech recognition for African languages. In 2014, they worked with us to help develop our first speech recognition programme for Ugandan English. They were able to use South African English and adapt it to Ugandan English so that we could get the variation in language, pronunciation and the like.
Then we moved to Luganda and Acholi languages, and more recently Lugbara and Rutooro. Our speech recognition for local languages isn’t perfect and we are continuously working to improve its functionality. For example, if people are discussing a topic like public health, they use several words that we might pick up: “malaria”, “hospital”, “doctor”, or “medicine” in the course of the conversation. Even if we don’t pick up every word, it is enough for the programme to recognize one or two words in a phrase to still identify the discussion as relevant.
Another piece of the puzzle that helped make this endeavor successful was the ongoing collaboration with our UN network and being able to pass on the issues which are coming up, whether it’s about health, or education, or refugees, to the UN agencies that work in those areas.
How easy was it for you to get radio stations on board? Were there reservations or questions regarding data privacy and security and how did you address them?
Data privacy is a very serious concern even though the data we collect is public. If someone calls in to the radio station and is talking about a topic, they might think it’s only going to be aired in their local community but may not wish for this information to be broadcast anywhere else – so we really have to be responsible with this. We had to consider topic appropriateness and this guided us in selection of topics. For example, natural disasters was a good topic because people talk about broad things like their crops, how the disaster has affected them – those kind of things. Hence topic selection is based on things that are neutral and practical.
When we were developing the tool, we held public consultation meetings with various stakeholders, including the media, to get their inputs. We also worked with radio stations to get insights from them into their broadcast schedules. They helped guide us on what times and days would be best for phone-in discussions, for example.
Can we then look at the radio content analysis tool as a worldwide innovation?
Absolutely. What makes it so interesting is how rich the data is especially in Uganda where radio is a very important tool for communication about issues going on in the community. Some of our international collaborators were surprised at how big a deal radio is in Uganda. I think we were kind of in the right place at the right time. Radio in Uganda is a treasure trove of information, and we just need a way to search it and organize it. There is of course tremendous potential for replicating what we have done in other countries, and other regions where radio is a rich source of information.
How would you explain the radio content analysis tool to a non-technical audience?
Think of the Internet, which has all sorts of information on different web pages, but imagine that there wasn’t any search engine like Google to help find what you’re looking for. If you just had to look at random web pages until you found what you wanted, then it would be pretty difficult to use! That’s a bit like how radio is now. If you want to listen to what’s being said on natural disasters, say, you might listen to the radio for a very long time to get the information you need. What we are trying to do for radio is similar to what Google did for the Internet. We are making radio content categorized and searchable, so that we can find the information we need to advance the Global Goals.
Lastly, what’s next for the radio content analysis tool?
One thing to understand is that the tool is not a one-size fits all type of solution. One limitation of the project is that every time we want to start analyzing radio in a different language, we have to train a new speech recognition system. In Uganda, there are more than 40 languages, which means we have to find ways to make the technology more easily scalable.
Another thing is that we’ve done a lot of work on the technology side, but we still have to factor in the human component of the analysis and interpretation of what comes out. People with the right experience and training can be better than computers at assessing bias, interpreting things in context, and translating and summarizing.
Finally, as mentioned above, we must also ensure we constantly protect the privacy of individuals who are speaking on the radio, even though this is a public source of information. That is why we apply a set of data privacy and data protection principles set out by UN Global Pulse.
This interview was conducted by Annette Akimanzi, Communications Officer at Pulse Lab Kampala.