Italian Counter Narrative Generation to Fight Online Hate Speech

Abstract

English. Counter Narratives are textual responses meant to withstand online hatred and prevent its spreading. The use of neural architectures for the generation of Counter Narratives (CNs) is beginning to be investigated by the NLP community. Still, the efforts were solely targeting English. In this paper, we try to fill the gap for Italian, studying how to implement CN generation approaches effectively. We experiment with an existing dataset of CNs and a novel language model, recently released for Italian, under several configurations, including zero and few shot learning. Results show that even for underresourced languages, data augmentation strategies paired with large unsupervised LMs can held promising results. Italiano. Le Contro Narrative sono risposte testuali volte a contrastare l’odio online e a prevenirne la diffusione. La comunità di NLP ha iniziato a studiare l’uso di architetture neurali per la generazione di CN. Tuttavia, gli sforzi sono stati rivolti esclusivamente all’inglese. In questo lavoro, cerchiamo di colmare la lacuna per l’italiano, mostrando come implementare efficacemente approcci di generazione di CN. Sperimentiamo con un dataset esistente di CN e un modello del linguaggio per l’italiano recentemente rilasciato, in diverse configurazioni, tra cui zero e few shot learning. I risultati mostrano che anche per lingue con poche risorse, strategie di data augmentation abbinate a potenti modelli del linguaggio possono offrire risultati promettenti. Copyright ©2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).