PhD defence by Victor Petrén Bach Hansen
Towards Fairness in Conversational Natural Language Processing
With the emergence of deep learning-based Natural Language Processing (NLP), the field of conversational Artificial Intelligence (AI) has gone from a mere pipe-dream to full-blown commercial integration into our society in the span of less than a decade. The recent advances of conversational systems are primarily driven by data-hungry models that require vast quantities of data, which poses several challenges in terms of the models' performance and their societal impacts.
This thesis presents work that contributes to the field of NLP and conversational AI in multiple ways. The first part explores methods for building more intelligent dialogue systems. Here we examine how user feedback can be incorporated to transfer knowledge from one domain to another more efficiently, how to resolve elliptical structures in a conversational context and how bias in dialogue-based data collection guidelines can manifest itself in the resulting corpora.
The second part looks at how NLP models adhere to socio-demographic fairness principles under different constraints, namely compression and privacy. While compressing neural models for the sake of a reduced memory footprint and inference cost is an attractive trait, we find that pruning methods in text classification systems lead to an increase in disparity of performance among different groups. Similarly, model privacy is also shown to be at odds with fairness principles, but we find that combined with distributionally robust optimization, it can lead to both private and fair models.
- Chair: Professor, Christian Igel, Department of Computer Science, UCPH
- Research Scientist, Katja Filippova, Google
- Professor, Sebastian Pado Computer Science Department, Stuttgart University
Professor, Anders Søgaard, Department of Computer Science, UCPH
Moderator at this defence
Assistant professor, Desmond Elliott, Department of Computer Science,
University of Copenhagen
For an electronic copy of the thesis, please go to: https://di.ku.dk/english/research/phd/.
This defence will be conducted online. Click here to join.