Elicitation of Requirements for a NLP-Model Store for Abusive Language Detection

Abstract

While in social media users most commonly interact with each other without a guiding entity, the discussion space of newspaper platforms is moderated by community managers, who invest considerable amounts of time and effort in keeping comment sections clean. To reduce this effort and allow community managers again to more freely interact with their users, automated comment moderation systems (ACMS) be be utilized. However, most newspapers do not have the expertise to create, update, and maintain machine learning (ML)-models. Thus, they are forced to rely on proprietary off-the-shelf solutions. However, if they want to keep their sovereignty over their data and systems, they would need access to models which could be integrated within their current moderation or content management systems. One option could be a platform for newspapers and data scientists where the data scientists could sell their pre-trained models and where newspapers could hire data scientists to create tailor-made models for them. In order to identify the requirements, community managers have for such systems, we conducted a series of semi-structured interviews with community managers of newspapers of varying size (from local to national). Furthermore, the information was enriched by the participation in multiple workshops on content moderation. We were able to elicit five major technical requirements necessary to create the described design artifact.