Distinguishing “good” from “bad” Arguments in Online Debates & Feature Analysis using Feed-Forward Neural Networks

Argument extraction and analysis in online content has gained significant interest in the past few years. The amount of information on the web increases daily with an increasing portion of information and opinion exchange occurring in online interactions on social media. Suitably mined and analysed, it could provide a lot of insight into the beliefs and reasoning of people about problems that are affecting our society. Traditional methods based on computation expensive text pre-processing and large amount of feature extraction are neither necessary nor suitable for this sort of domain. Online language in social media does not follow the usual grammar and stylistic rules as taught in school books, which makes it questionable whether complex feature vectors or word embeddings yield the desired results. Based on this, this study proposes a method with which arguments in a given online debate can be assessed by comparing individual argument features against the average argument in the debate, therefore making the values dependent on the quality of the overall debate, using only simple string manipulation and light linguistic feature extraction and a standard feed-forward neural network. By comparing all arguments against each other a ranking can be retrieved and the arguments of the debated sorted according to quality which makes such an application suitable for the task of filtering out the best and most valuable arguments in any online argumentation.