Newly appointed Associate Professor will use machine learning to explain gender bias towards politicians

One of the most hyped research projects carried out at DIKU in 2019 was Isabelle Augenstein’s gender bias detection project, which got a lot of media attention in both Danish and international media, including Forbes. The public showed a great interest in Isabelle’s and her colleague’s extensive findings on how women and men are typically described in literature, and now, Isabelle and her research group have received funding to continue the project to new areas.

The extended project expands the gender bias analysis to social media – namely Twitter – to multiple languages, and certain entities so that the group can investigate how biases towards entities differ in different countries, for instance, biases towards politicians. They will also be able to investigate the relationship between different languages and gender bias.

- Are female politicians ascribed to different characteristics than male politicians? Is there more gender bias towards politicians in conservative countries rather than in liberal countries? Do certain languages create more gender bias towards specific kinds of persons simply because of the linguistics of those languages? And which attitudes hold the most gender biases in general? These are some of the questions that we’re interested in answering, Isabelle says.

Techniques like these are sought after by many companies. Especially in the light of the #MeToo movement, there is a great interest in gaining a more in-depth understanding of the scale of gender issues on social media.

- Even the University of Copenhagen would like to know how people perceive the University. Do people think there is a lot of gender bias here? Do people perceive our atmosphere as diverse? The nice thing about this project is that it will offer some explanations as well. It won’t just conclude “there is gender bias towards this entity or not”, but we will try to tie that to specific attitudes that other people referring to that entity exhibit, for example, attitudes about parental leave or employment quotas, Isabelle explains.   

The original project was about detecting patterns from language that quantify gender bias in our society. Isabelle, together with researchers at the University of Cambridge, University College London, Johns Hopkins University, and Microsoft Research, used machine learning methods to analyse 3.5 million books and found out that men are typically described by words that refer to behaviour, while adjectives ascribed to women tend to be associated with physical appearance.

- It started as an informal collaboration without funding; it was simply something that we were interested in and investigated when we had the time. So, I’m very excited that we have now received funding from the Independent Research Fund Denmark to continue the project for four years, Isabelle says.

Analysing natural language

Isabelle is head of the CopeNLU research group, which investigates subjects within Natural Language Understanding (NLU). They build machine learning models that can automatically understand and analyse natural language, which means any language produced by humans. Besides the gender bias detection project, other projects that are in the pipeline for Isabelle and the CopeNLU group is a project about fact-checking and a project about cross-lingual and multilingual learning which are also some of Isabelle’s core research areas.

The fact-checking project has two PhD students, both funded by a Marie Curie fellowship program. One is investigating explainability in fact-checking, and the other is investigating fact-checking for scientific claims, which is very challenging. She also collaborates with one of DIKU’s other researchers, Professor Christina Lioma, whose team has recently won a prize for fact-checking in online debates.

In the cross-lingual and multilingual learning project, Isabelle’s group is finding similarities and differences between high resource languages such as English and low resource languages such as Icelandic. The aim is to train models to work well for multiple languages, including low-resource ones.

Supervising is my favourite part

Isabelle is happy to have completed her tenure track position and to have been promoted to Associate Professor. In her new position, she will be able to take part in more exciting research projects.

- There is quite a big difference in being an Assistant Professor and an Associate Professor in terms of what one is allowed to do, at least in Denmark. The most important change is that I am now allowed to formally supervise PhD students, which I really like. It’s actually my favourite part of my job, so it means a lot to me to be able to do that for sure now, Isabelle says.

This is because the research culture in Machine Learning is different from the more theoretical fields of Computer Science. A lot of Machine Learning research is experimental, which means one has to implement different models and run many experiments with them on servers. Supervising PhD students and Postdocs is, therefore, a big part of what it means to do research in this field.

- Our work requires a lot of “helpers” who can either be PhD students or Postdocs. My colleagues and I would not be able to write papers on our own without our PhDs and Postdocs. Therefore, Professors in the Machine Learning section usually supervise a lot of PhD students and Postdocs – I have 8 myself – and we are very actively involved in forming their ideas and coming up with good hypotheses.

Last but not least, Isabelle will continue teaching courses in natural language processing and machine learning as part of DIKU’s Master’s and Bachelor’s programmes, which she also enjoys:

- I like teaching here in the Danish University system because we have time to prepare and reflect on lectures. In that way, we don’t end up merely churning out lectures, as opposed to some other countries I know of. To me, that preserves my joy in teaching.