AUTOMATIC MODERATION FOR IBIS FORUM

Abstract

Painstaking manual moderation work in IBIS (Issue Based Information System) forums is needed to maintain the requisite structure and requires careful human reading and intelligence. The core moderation task is to identify whether a post is “structurally incorrect”. We are the first to automate this with a learning task. We clean and analyze the raw data from a particular IBIS forum sourced from the MIT Deliberatorium, and in doing so, create a dataset appropriate for devising a machine learned solution. We engineer six sets of features: bag of words, type, length, selected word(s), structure, similarity, then provision these features to a SVM classifier to create a learned automatic moderator. Our system achieves an approximate AUC of 0.8, which we feel is a good first step towards full automation. For in depth analysis, we further perform ablation tests and micro analyses to understand the effect of each feature set on performance.