Topic similarity networks: Visual analytics for large document sets

Abstract

We investigate ways in which to improve the interpretability of LDA topic models by better analyzing and visualizing their outputs. We focus on examining what we refer to as topic similarity networks: graphs in which nodes represent latent topics in text collections and links represent similarity among topics. We describe efficient and effective approaches to both building and labeling such networks. Visualizations of topic models based on these networks are shown to be a powerful means of exploring, characterizing, and summarizing large collections of unstructured text documents. They help to "tease out" non-obvious connections among different sets of documents and provide insights into how topics form larger themes. We demonstrate the efficacy and practicality of these approaches through two case studies: 1) NSF grants for basic research spanning a 14 year period and 2) the entire English portion of Wikipedia.

Evaluating Similarity Metrics for Latent Twitter Topics

“…Gretarsson et al [9] estimated the similarity of latent topics by computing the L 1 distance. Later, Maiya et al [5] adopted the Hellinger distance metric to calculate the similarity between topics. On the other hand, the most common distance metric used in the literature is the cosine similarity [6,10,11].…”

“…It is often used as a topic similarity metric in prior work [6,9]. The Hellinger distance-based metric (HDM) is often used to quantify the similarity between a pair of probability distributions, as in [5].…”

“…Commonly, the similarities of topics can be computed by using the distribution of topics over the vocabulary. Previous work has applied metrics such as the Hellinger distance (HD) [5], the Jensen Shannon (JS) divergence [6] or the cosine similarity (CS) [1,7] to measure the similarity between topics.…”

“…However, Ramage et al[1] argued that the top-ranked words might often be similar. Hence, we choose to use the top 15 words in this work 5. We use Gibbs sampling as it can still generate topics that connect well to the real topics (see[2]).…”

Vector-based similarity measurements for historical figures

“…Network-based similarity measurements can be used for classification on large scale document collections [15], to provide unsupervised hierarchic similarity structure [16] and also to benefit visualizations to better understand "similarity" [17]. Deepwalk [18] seeks vector representations using recent advancements in language modeling and unsupervised feature learning (or deep learning) from sequences of words to graphs.…”