With the help of such a programme, students can improve their essays before they hand them in. And for teachers, this will make checking up on them a lighter task. The knife cuts both ways. Almost two years ago, Liqin Zhang started his PhD research at TELI, the research group for Technology-Enhanced Learning Innovations of the Welten Institute, with Professor Marco Kalz as a supervisor. Zhang's interest in language in combination with artificial intelligence (AI) brought him from the province of Guangdong in China via the universities of Twente and Groningen to the Open University in Heerlen.
Assessing argumentation quality
Zhang calls it 'challenging' to develop a tool that can determine the text quality for the teacher. In the first instance, his goal is a little closer: in the next two years, he wants to develop a model that works to analyse the quality of the argumentation in essays. Such a model could provide students with automated feedback that would help them to improve their texts.
Argumentation mining
Liqin Zhang began his research with a continuous survey into teachers' views on essays and the difficulties in assessing them. In the meantime, he is mainly busy building up a large stock of annotated Dutch-language texts by secondary school students. With the help of colleagues, he has now compiled a database of some thirty essays. The annotations indicate where, for example, an argument begins or ends, or what the central theme of the text is. The identification of this structure is called argumentation mining and helps computers to analyse natural language (natural language processing).
Algorithms
The field of argumentation mining has only recently developed and is especially new for languages such as Dutch, says Zhang. 'You have to have enough data to be able to do this properly. The corpus I now have is sufficient to test whether the model works. But to improve the quality, you need a lot more data, thousands of texts.' With the current text corpus, Zhang can at least 'feed' the algorithms of the program. The more the algorithms 'learn' about the structure of the argumentation in a text, the better they can analyse and assess it.
Multilingual application
Zhang does not speak Dutch fluently, but that is not a problem for his research, he says. He bases his work on research by German scientists who made a working model for the analysis of English-language texts. This model uses artificial neural networks, which mimic the functioning of the brain, and Word2vec. With this technique, words are marked with characteristics that allow them to be placed and recognised in a context automatically. Programming with this multilingual application can be done for any language, such as Spanish or Chinese. 'And Google Translate helps to find out the global meaning of a text', Zhang laughs.
