A research team from the University of Sheffield in the UK has been awarded a £1 million grant by the Leverhulme Trust to carry out a major research project with an ambitious aim: make language learning more natural for adults, Claudia Civinini writes. Team members Dr Dagmar Divjak and Dr Petar Milin represent expertise from a wide range of fields, from linguistics and psychology to machine learning, and are supported by research software engineer Dr Mike Croucher.
In their project, they aim to reach a deeper understanding of what speakers know about their first language and, with the aid of machine-learning techniques that mimic the way humans learn, find a new way to teach foreign languages to adults. They will concentrate on the two most widely spoken languages in the UK, Polish and English.
The study will be carried out over five years and will consist of three components. The first three-year component will gather data on the linguistic knowledge of native speakers. What do speakers know about their own language? This will be conducted through a series of experiments, from cloze (gap-fill) and correction exercises to other lab-based techniques such as eye-tracking. The second component will focus on machine learning and run parallel to the first. The two components will cross-inform one another. The patterns of learning that emerge from the tests performed by native speakers will inform a series of algorithms. These will be tested by using them to predict the outcome of the experiments from the first component of the study, to try and replicate the way humans learn.
When asked how they would account for the machines’ lack of understanding of pragmatics and semantics, Dr Divjak explained that their previous work had shown that a standard statistical classifier can actually predict as well as a native speaker which of six synonyms to choose, without the semantic knowledge. Dr Milin explained that the learning algorithms they plan to use are all biologically (or psychologically) plausible. For example, one of the core algorithms, which was co-developed by Dr Milin, is similar to conditioning, referring to the famous Pavlov’s Dog experiment. Cues such as words that co-occur in a context can allow machines to predict what comes next – in the same way that humans do.
The third and final component, of about three years, will be dedicated to the development of teaching materials based on the patterns that the machines have flagged up as important. The materials will be tested through classroom intervention studies. The aim is to mirror natural language learning, moving towards replacing explicit grammar instruction with implicit learning. This will be achieved, for example, by exposing students to comprehensible input packed with examples of the pattern they intend to learn. ‘We’ll try to stay as close as possible to how people learn their first language,’ the team said. ‘Learning may not be easier, but more natural.’ We take the opportunity to congratulate the team on the achievement and wish them well for their project. Since last year, we have started to interview academics again with our long-time favourite question: what would you do with a £1 million grant? Since then, two research teams have won such grants in ELT – one in the US (January 2017 Gazette, we’ll interview them soon) and one in Sheffield. We start to sense a pattern here.
Picture: University Sheffield Courtesy: Bea