Polygloss - A conversational agent for language practice

Abstract

This paper explores the impact on language proficiency of comprehensible output applied in computer assisted language learning (CALL). Targeting speakers of intermediate level, we adapted a visually-grounded dialogue task, optimizing for language acquisition. The task was implemented as a mobile application where learners are organized in pairs and write short texts to play an image-guessing game, producing samples in a wide variety of languages. Following a framework for CALL evaluation, we conducted an analysis of the game and players’ gains through time, including the measure of pre-trained XLM-r cross-lingual transformers’ acceptability score of the samples. The results confirm the intended fit for intermediate speakers as well as reveal possible benefits for other levels. This research provides a successful case study of a multilingual CALL design where users have the autonomy to generate output creatively.