Cross-lingual text classification alleviates the need for manually labeled\ndocuments in a target language by leveraging labeled documents from other\nlanguages. Existing approaches for transferring supervision across languages\nrequire expensive cross-lingual resources, such as parallel corpora, while less\nexpensive cross-lingual representation learning approaches train classifiers\nwithout target labeled documents. In this work, we propose a cross-lingual\nteacher-student method, CLTS, that generates "weak" supervision in the target\nlanguage using minimal cross-lingual resources, in the form of a small number\nof word translations. Given a limited translation budget, CLTS extracts and\ntransfers only the most important task-specific seed words across languages and\ninitializes a teacher classifier based on the translated seed words. Then, CLTS\niteratively trains a more powerful student that also exploits the context of\nthe seed words in unlabeled target documents and outperforms the teacher. CLTS\nis simple and surprisingly effective in 18 diverse languages: by transferring\njust 20 seed words, even a bag-of-words logistic regression student outperforms\nstate-of-the-art cross-lingual methods (e.g., based on multilingual BERT).\nMoreover, CLTS can accommodate any type of student classifier: leveraging a\nmonolingual BERT student leads to further improvements and outperforms even\nmore expensive approaches by up to 12% in accuracy. Finally, CLTS addresses\nemerging tasks in low-resource languages using just a small number of word\ntranslations.\n