Topic Modeling Using Collapsed Typed Dependency Relations

Topic modeling is a powerful tool to uncover hidden thematic structures of documents. Many conventional topic models represent documents as a bag-of-words, where the important linguistic structures of documents are neglected. In this paper, we propose a novel topic model that enriches text documents with collapsed typed dependency relations to effectively acquire syntactic and semantic dependencies between consecutive and nonconsecutive words of text documents. In addition, we propose to enforce coherent topic assignments for conceptually similar words by generalizing words with their synonyms. Our experimental studies show that the proposed model and strategy outperform the original LDA model and the Bigram Topic Model in terms of perplexity; and our performance is comparable to other models in terms of stability, coherence, and accuracy.