Dylan Baker

Harvey Mudd College Mathematics 2017

Dylan Baker Photo
Thesis Advisor: Prof. Talithia Williams
Second Reader: Prof. Tanja Srebotnjak
E-Mail: dbaker@hmc.edu

The Document Similarity Network: A Novel Technique for Visualizing Relationships in Text Corpora

With the abundance of written information available online, it is useful to be able to automatically synthesize and extract meaningful information from text corpora. We present a unique method for visualizing relationships between documents in a text corpus. By using Latent Dirichlet Allocation to extract topics from the corpus, we create a graph whose nodes represent individual documents and whose edgeweights indicate the distance between topic distributions in documents. These edge lengths are then scaled using multidimensional scaling techniques, such that more similar documents are clustered together. Applying this method to several datasets, we demonstrate that these graphs are useful in visually representing high-dimensional document clustering in topic-space.

Sample Document Similarity Network

An example Document Similarity Network, constructed over 60 subreddits from the website Reddit.com