Topic discovery by spectral decomposition and clustering with coordinated global and local contexts

Topic modeling is an active research field due to its broad applications such as information retrieval, opinion extraction and authorship identification. It aims to discover topic structures from a collection of documents. Significant progress have been made by the latent dirichlet allocation (LDA) and its variants. However, the “bag-of-words” assumption is usually made for the whole document by conventional methods, which ignores the semantics of local context that play crucial roles in topic modeling and document understanding. In this paper, we propose a novel coordinated embedding topic model (CETM), which incorporates spectral decomposition and clustering technique by leveraging both global and local context information to discover topics. In particular, CETM learns coordinated embeddings by using spectral decomposition, capturing the word semantic relations effectively. To infer the topic distribution, we employ a clustering algorithm to capture semantic centroids of coordinated embeddings and derive a fast algorithm to obtain the topic structures. We conduct extensive experiments on three real-world datasets to evaluate the effectiveness of CETM. Quantitatively, compared to state-of-the-art topic modeling approaches, CETM achieves significantly better performance in terms of topic coherence and text classification. Qualitatively, CETM is able to learn more coherent topics and more accurate word distributions for each topic.