Toward any-language zero-shot topic classification of textual documents

In this paper, we present a zero-shot classification approach to document classification in any language into topics which can be described by English keywords. This is done by embedding both labels and documents into a shared semantic space that allows one to compute meaningful semantic similarity between a document and a potential label. The embedding space can be created by either mapping into a Wikipedia-based semantic representation or learning cross-lingual embeddings. But if the Wikipedia in the target language is small or there is not enough training corpus to train a good embedding space for low-resource languages, then performance can suffer. Thus, for low-resource languages, we further use a word-level dictionary to convert documents into a high-resource language, and then perform classification based on the high-resource language. This approach can be applied to thousands of languages, which can be contrasted with machine translation, which is a supervision-heavy approach feasible for about 100 languages. We also develop a ranking algorithm that makes use of language similarity metrics to automatically select a good pivot or bridging high-resource language, and show that this significantly improves classification of low-resource language documents, performing comparably to the best bridge possible.