VALICO-UD: Treebanking an Italian Learner Corpus in Universal Dependencies

This article describes an ongoing project for the development of a novel Italian treebank in Universal Dependencies format: VALICO-UD. It consists of texts written by Italian L2 learners of different mother tongues (German, French, Spanish and English) drawn from VALICO, an Italian learner corpus elicited by comic strips. Aiming at building a parallel treebank currently missing for Italian L2, comparable with those exploited in Natural Language Processing tasks, we associated each learner sentence with a target hypothesis (i.e. a corrected version of the learner sentence written by an Italian native speaker), which is in turn annotated in Universal Dependencies. The treebank VALICO-UD is composed of 237 texts written by non-native speakers of Italian (2,234 sentences) and the related target hypotheses, all automatically annotated using UDPipe. A portion of this resource (36 texts corresponding to 398 learner sentences and related target hypotheses)—firstly released on May 2021 in the Universal Dependencies repository—is associated with error annotation and the automatic output is fully manually checked. In this article, we focus especially on the challenges addressed in treebanking a resource composed of learner texts. In addition, we report on a preliminary data exploration that makes use of three quantitative measures for assessing the quality of the data and for better understanding the role that this resource can play in tasks lying at the intersection of Computational Linguistics and learner corpus studies