Zero-shot Cross-Lingual Switch for Artificial Knowledge Technology in Grammatical Error Detection
Authors: Gaetan Lopez Latouche, Marc-André Carbonneau, Ben Swanson
Summary: Grammatical Error Detection (GED) strategies rely closely on human annotated error corpora. Nevertheless, these annotations are unavailable in lots of low-resource languages. On this paper, we examine GED on this context. Leveraging the zero-shot cross-lingual switch capabilities of multilingual pre-trained language fashions, we prepare a mannequin utilizing information from a various set of languages to generate artificial errors in different languages. These artificial error corpora are then used to coach a GED mannequin. Particularly we suggest a two-stage fine-tuning pipeline the place the GED mannequin is first fine-tuned on multilingual artificial information from goal languages adopted by fine-tuning on human-annotated GED corpora from supply languages. This strategy outperforms present state-of-the-art annotation-free GED strategies. We additionally analyse the errors produced by our technique and different sturdy baselines, discovering that our strategy produces errors which might be extra various and extra much like human error
