TABU Dag 2025

The PAGINA team has presented their ongoing work at TABU Dag 2025.

TABU Dag is a well-established, international linguistics conference organised at the University of Groningen with a varied programme and guest speakers from different fields.

In our presentation, we stressed the importance of text simplification and its challenges, especially regarding evaluation, advocating the need for human-centered evaluation where the readers are put first.

We described our dataset with aggregated reading times from thousands of actual readers and how it can be used to train a reading time predictor. Such a predictor could be useful for evaluating simplified texts, where lower predicted reading time implies a text is easier to read, assuming that people read faster through simple(r) texts.

Next to the predictor, we asked three large language models (LLM) to directly assess the reading time and complexity of several articles from the dataset.

We find good features at different linguistic levels that are informative to predict reading time and that LLMs can look into more subtle features such as style, tone, comprehension, and cognitive load.