Human-Centered Research
In human-centered data science, we focus on developing models that account for the diversity in perspectives and experiences of the people who are impacted by them. While data science traditionally operates with numbers, humans communicate with language. The idea of mapping symbolic language input into quantitative models evolved from an academic niche interest to a pervasive technology that is used worldwide. We rely on language technology to search for information, to communicate across language boundaries, and to refine our writing style. The most recent models can assist us during brainstorming or even write draft versions of entire documents. Integrating their usage into our daily routines will soon feel as ordinary as using a calculator. These developments are fueled by large language models that optimize billions of parameters using terabytes of training data. The enormous computing resources required for their development can currently only be afforded by a handful of dominant companies whose engineering objectives are driven by economic targets. As a consequence, language model development currently prioritizes the English language and a mainstream audience which creates a inbalance in the access to information for already marginalized user groups. In our lab, we work on natural language processing research with a human-centered perspective.
Projects
Our projects focus on cross-lingual and cognitively inspired research questions that we explore with computational models.
- Interpretability of transfer in multilingual models
- Cognitive plausibility in NLP
- BabyLM - cognitively inspired representation learning for more efficient language models
- Robustness of interpretability methods
- Educational NLP