Embedding wikipedia title based on its wikipedia text and categories

摘要

Distributed word representation is widely used in many NLP tasks and knowledge-based resources also provide valuable information. Comparing to conventional knowledge bases, Wikipedia provides semi-structural data other than structural data. We argue that a Wikipedia title's categories can help complement the title's meaning besides Wikipedia text, so the categories should be utilized to improve the title's embedding. We propose two directions of using categories, cooperating with conventional context-based approaches, to generate embeddings of Wikipedia titles. We conduct extensively large scale experiments on the generated title embeddings on Chinese Wikipedia. Experiments on word similarity task and analogical reasoning task show that our approaches significantly outperform conventional context-based approaches.