We propose a novel generative model to explore both local and global context\nfor joint learning topics and topic-specific word embeddings. In particular, we\nassume that global latent topics are shared across documents, a word is\ngenerated by a hidden semantic vector encoding its contextual semantic meaning,\nand its context words are generated conditional on both the hidden semantic\nvector and global latent topics. Topics are trained jointly with the word\nembeddings. The trained model maps words to topic-dependent embeddings, which\nnaturally addresses the issue of word polysemy. Experimental results show that\nthe proposed model outperforms the word-level embedding methods in both word\nsimilarity evaluation and word sense disambiguation. Furthermore, the model\nalso extracts more coherent topics compared with existing neural topic models\nor other models for joint learning of topics and word embeddings. Finally, the\nmodel can be easily integrated with existing deep contextualized word embedding\nlearning methods to further improve the performance of downstream tasks such as\nsentiment classification.\n