SEMPLICe: um modelo sequencial de proficiência em comunidades online para aprendizado de idioma

Abstract

Reddit is an online social network where users interested in a common subject may interact with each other through subreddits. Subreddits for language learning have been attracting users of various proficiency levels each year, interested in boosting their learning. In particular, on subreddit German, users are advised to inform their proficiency level when writing a post. Yet only 20% of the posts have such tags. In this paper we address the problem of classifying users’ proficiency from their publications. We conduct experiments which show that classifiers that treat publications as independent observations perform poorly. We then propose a new model dubbed SEMPLICe, which uses both textual features and the publication history of an user to classify her proficiency level over time. By assuming that proficiency is monotonically non-decreasing as long as the user remains active, SEMPLICe yields a weighted F1 score up to 29.6% higher than previous methods. SEMPLICe uses dynamic programming to achieve linear complexity on each user’s history size.