Detecting Harmful Comments on Social Media a Machine-Learning Framework Using Text Features, Apriori Analysis and Classifiers

Comment sections on social media are where most public conversation now happens, and they are also where abuse, harassment, hate speech and spam spread fastest.Human moderators cannot read everything, and simple keyword filters are easy to dodge with a misspelling or a symbol.In this paper we propose a detector for harmful comments built on machine learning.It looks at three kinds of evidence: the words and phrasing of the comment, a few stylistic and structural cues such as capital letters, repeated characters and links, and light context such as whether the comment is a reply.We will also use the Apriori algorithm to find which cues tend to show up together, and then train Logistic Regression, Random Forest, XGBoost and a fine-tuned transformer model and compare them on the same split.The tool is meant to return a label, a confidence score and a short note on which parts of the comment pushed the score up.The Jigsaw Toxic Comment dataset will serve as a first benchmark, followed by a second dataset from a different platform to check how well the models carry over.This is a proposal.We have not run the experiments yet, so no accuracy figures of our own are reported here.