A month ago, I met up with two friends from the international development world, and we talked about one of their research projects into the refugee crisis on the Myanmar-Bangladesh border. Even if it’s no longer in the headlines, Rohingya and other refugees are in a precarious, flood-prone camp on the Bangladeshi side of the border. Local protectionist policies prohibit refugees from regular employment, so there are smaller jobs inside the camp which pay in redeemable coupons, and a cryptocurrency project which serves a similar purpose.
One idea was whether a refugee code school could be started in Bangladesh. It would be great to set up a training program and get refugees into remote-paid jobs. At first I was optimistic because I’d done workshops at refugee code schools in Turkey and Kurdistan, and met with the founder of Code to Inspire (Afghanistan), and heard about their work finding remote internships and jobs. But as we talked about the circumstances of the Rohingya, there were clear differences:
- the refugees have few or no regular schools in the camp, so we can’t partner with one or use their facility
- everyone has smartphones, but we can’t expect students to have conventional computer skills or secure a laptop at home
- any local applications should use Rohingya language and script, which has few technical resources; with the local jobs ban, Bangla language is not prioritized
The schools which I visited last year were handling difficult circumstances, but they were based in established cities and could screen students for preexisting language and computer skills.
Can you teach programming on a smartphone?
In the has been a question from parents in the US, too:
This one is a little more recent and focused on adults learning:
There were a variety of responses, which I want to try out. Some are IDEs that ‘work’ on mobile, some have a Duolingo-like model. I can’t think of how a phone could be used for bigger, projects-based learning.
Whatever angle you look at this, there isn’t a clear leader in this space. And I worry the lack of a leader means that these programs don’t measure up, even for students in the best and most stable circumstances.
What about labeling/tagging for machine learning?
In Kurdistan, I met two bootcamp graduates who work at Taqadam, a company which employees refugees in labeling maps and other images for machine learning. After chatting with my friend about this, I also had a Skype call with Lorraine from Na’amal who is familiar with a few of these initiatives.
Labeling/tagging is limited, it doesn’t elevate workers to highly-skilled, well-paid jobs. But when you’re starting with nothing (no school, no laptops, no tech-savvy), it’s a way to try out building a team, training on technology, using smartphones, and getting paid. If you’re not able to run this pilot project for some reason, if the workers’ payments can’t be withdrawn and spent in the camp, that’s a significant failure.
How about chat bots?
I was starting to think about this — if we can’t code a full app on a phone, maybe it would be possible to do chat bot programming and training through messaging a programer chat bot? Suppose we present a situation (‘I noticed ___, should I talk to a doctor?’, ‘when you had your first baby, ___ ?’) and ask 10 students to fill in the blank with ten different questions, and grade some of the other students’ messages. These questions are going to take expert input to answer, but the chatbot brain needs to be trained on real questions that we can classify and repeat expert answers
At first this is more of a machine learning labeling project and not coding, but the same model could be applied to virtually any project, and I could see expanding on the mobile platform and programmer chat bot for students to make their own projects.
I spoke with an engineer who uses the Rasa platform for chat bots in a women’s health program.
Rasa makes a lot of sense because there is some in-depth programming and config files and training.
For English chatbots, you can configure different messaging, a tokenizer, and pretrained word embeddings. For this particular project, Rohingya language is not going to be supported out of the box.
- I asked around if there is a large corpus of Rohingya text somewhere in any script, or even movie subtitles. Right now this seems unlikely to appear to make my own ELMO or BERT.
- I could see building a vocabulary list and term frequency (TF-IDF) — old school, pre-word2vec NLP — but no one has code or info for tokenizing Rohingya language. It might be a cool research project to use HuggingFace tokenizers, or settle for simple whitespace and punctuation splits.
- If we could screen for students with English expertise, it makes the NLP part easier and less experimental. Even if this limits the impact of the program, it means there’s more time to focus on the school portion. It would also be good to pick up on dialect, local words, and texting conventions of the students to make chat bots attuned to their needs.
These are the suggestions that I’ve been sending to my friend while I watch the RASA class and get a better feel for the platform.