The Write & Improve programme, an AI-driven language-learning service launched last year, is run by start-up English Language iTutoring. Co-founder and c.t.o. Paul Butcher will present the product at the FutureBook EdTech stream on 1st December, during the panel “How can publishers harness the power of AI?”.
Butcher has held senior technical roles at several high-tech start-ups. Most recently before co-founding English Language iTutoring, he was chief software architect of SwiftKey, the mobile phone keyboard that uses natural language processing and machine learning to make text entry painless. SwiftKey was the best-selling Android application worldwide for several years and was recently acquired by Microsoft. Butcher is also the author of two well-regarded books on software engineering, and is currently working on a third.
We sat down with Butcher to discuss that most fashionable of acronyms - and how his team are actually turning it into a useful tool for language learning.
Who are English Language iTutoring?
We are a joint venture between exams group Cambridge Assessment, Cambridge University Press and a group of technologists all based around Cambridge. There is an academic institute within the University of Cambridge called the ALTA (Automated Language Teaching and Assessment) Institute. It’s first and foremost about doing research, but obviously the research that comes out of it is very relevant to the sort of things that Cambridge Assessment English are up to. We basically exist in order to take that research and productise it.
We are slightly unusual compared to the average start-up, which has to create the technology it is going to use, and create a brand to market the products under, as well as all the operational things to make it actually work. For us, the underlying technology has come out of the ALTA Institute and we have very high levels of confidence in it because we’ve got a bunch of extremely clever academics who have been creating it over 15 years, and we don’t have to create our own brand because we are piggybacking on the Cambridge Assessment brand. So what we’re about is that third part – making sure that it has a nice user interface, operates 24 hours a day all around the week, the marketing and payments and product design.
Tell us more about the AI at the core of Write & Improve.
It's a natural language processing engine, the application of artificial intelligence techniques to human language. The engine we have got is aimed at assessing the quality of the English created by people who are learning English as a foreign language and then giving them feedback which will help them improve. People tend to immediately think of a product like the grammar checker in Microsoft Word or something like [platform] Grammarly – but their goal is very much to just get you to a correct piece of writing. Whereas what we’re focused on is the pedagogical aspect of feedback, we try to give feedback which doesn’t tell the user to change this word to that, but to give them pointers as to which aspects of English they should concentrate on in order to improve in the run-up to, say, taking an IELTS exam.
How does the AI do its job?
It’s based on training data; basically a very, very big statistical model. We train that statistical model with many, many hundreds of thousands of examples and it learns to understand the kind of mistakes that a typical English language learner makes while they are learning the language. People use the tool and they get feedback, but we also capture what it is they have written and “label" it by putting it in front of trained English language teachers, who notice places where the automated engine is either giving incorrect or incomplete feedback. They correct it and those corrections, or “labels", get sent back into the system, and the system learns to give better feedback over time. The way these things work it’s about the quantity of training data that you can apply. At the moment we are knocking at the door of 200,000 submissions a month.
What is new and different about Write & Improve?
The core technology it’s based on was around back in the 1970s, but what’s changed is the level of computer performance that’s available and the understanding of how to apply that technology to problems, in particular how to get access to the very large quantities of training data that are necessary to make it work. Where it’s interesting is making sure we present it to potential users in a way that is appealing to them. We are using a very data-driven approach: we do an awful lot of A/B tests where we will try launching some functionality or modification, and we measure the results on our user base, and see whether they use the technology more or less, and come to a judgement whether the tool we have just released makes things better or worse by looking at the way the users interact with the site.
See Butcher and other edtech innovators live at FutureBook 2017 - tickets on sale here until Friday.