CLASP
The Centre for Linguistic Theory and Studies in Probability

Bias and Methods of AI technology studying Political Science

In this talk I will describe cross disciplinary project involving computer scientists and political scientists, funded by WASP-HS. Political scientists have an increasing interest in using AI methods in their research. In particular, the “text-as-data” paradigm is growing and require knowledge about techniques from machine learning, data science and natural language processing. Various sources of data is used, from short informal twitter posts, to transcripts of speeches or written parliamentary motions. Furthermore, the language of the country of study will influence the choice of AI techniques. For instance, a lot more data is available for English than for smaller languages, which means that powerful but data-hungry deep neural network methods might not always be the first choice.

In our project, we have chosen to initially focus on written motions from the Swedish parliament, after also exploring transcripts of speeches from the US Senate. I will briefly describe some preliminary results where we compare the word embeddings resulting from training machine learning models on data from different parties. Is there a difference in how opposing parties use language about controversial issues which is reflected in word embeddings? Can this hint at differences in policy and ideology?

This project is joint work between political scientists at Karlstad University (Annika Fredén and Pasko Kisic Merino) and computer scientists at Chalmers (Moa Johansson and Denitsa Saynova).