In this paper, we propose a cross-lingual convolutional neural network (CNN) model that is based on word and phrase embeddings learned from unlabeled data in two languages and dependency grammar. Compared to traditional machine translation (MT) based methods for cross lingual sentence modeling, our model is much simpler and does not need parallel corpora or language specific features. We only use a bilingual dictionary and dependency parser. This makes our model particularly appealing for resource poor languages. We evaluate our model using English and Chinese data on several sentence classification tasks. We show that our model achieves a comparable and even better performance than the traditional MT-based method.
展开▼
