Multilingual and cross-lingual document classification: A meta-learning approach

The great majority of languages in the world are considered under-resourced\nfor the successful application of deep learning methods. In this work, we\npropose a meta-learning approach to document classification in limited-resource\nsetting and demonstrate its effectiveness in two different settings: few-shot,\ncross-lingual adaptation to previously unseen languages; and multilingual joint\ntraining when limited target-language data is available during training. We\nconduct a systematic comparison of several meta-learning methods, investigate\nmultiple settings in terms of data availability and show that meta-learning\nthrives in settings with a heterogeneous task distribution. We propose a\nsimple, yet effective adjustment to existing meta-learning methods which allows\nfor better and more stable learning, and set a new state of the art on several\nlanguages while performing on-par on others, using only a small amount of\nlabeled data.\n