Automated content moderation is widely deployed, but the systems behind it are often black boxes, and their coverage skews heavily toward English. A new preprint on arXiv addresses both problems at once by focusing on Urdu written in the Nastaliq script, a language and script that existing moderation tools rarely handle well.

The paper describes an explainable system built around three components: a curated corpus for training and evaluation, benchmarking of transformer-based models, and word-level attribution that highlights which parts of a text drive a moderation decision. That last piece is meant to make the system's output understandable to the people affected by it, rather than presenting an opaque verdict.

Because the abstract is truncated, specific results and dataset details are not available in the source. The contribution, as stated, is in extending explainable moderation beyond English and into a linguistically distinct script. The paper is a single source, so there are no conflicting findings to compare, but the framing suggests a broader push toward more transparent and inclusive moderation tools.