Fairness of Extractive Text Summarization

We propose to evaluate extractive summarization algorithms from a completely new perspective. Considering that an extractive summarization algorithm selects a subset of the textual units in the input data for inclusion in the summary, we investigate whether this selection is fair. We use several summarization algorithms over datasets that have a sensitive attribute (e.g., gender, political leaning) associated with the textual units, and find that the generated summaries often have very different distributions of the said attribute. Specifically, some classes of the textual units are under-represented in the summaries according to the fairness notion of adverse impact. To our knowledge, this is the first work on fairness of summarization, and is likely to open up interesting research problems.