Communications of the ACM
Q&A: Finding Themes
Credit: Denise Applewhite / Princeton University
In announcing David Blei as the latest recipient of the ACM-Infosys Foundation Award in the Computing Sciences, ACM president Vint Cerf said Blei's contributions provided a basic framework "for an entire generation of researchers." Blei's seminal 2003 paper on latent Dirichlet allocation (LDA, co-authored with Andrew Ng and Michael Jordan while still a graduate student at the University of California, Berkeley) presented a way to uncover document topics within large bodies of data; it has become one of the most influential in all of computer science, with more Google Scholar citations than, for example, the earlier PageRank paper that launched Google. Blei's approach and its extensions have found wide-ranging uses in fields as varied as e-commerce, legal pretrial discovery, literary studies, history, and more. At Princeton since 2006, Blei begins this fall as a professor at Columbia University, with joint appointments in the two departments his work bridges, computer science and statistics.
How did the idea for LDA come about?
The idea behind topic modeling is that we can take a big collection of documents and learn that there are topics inside that collection—like sports or health or business—and that some documents exhibit a pattern of words around that topic and others don't. That idea really began in the late '80s and early '90s with latent semantic analysis (LSA), from people like Susan Dumais at Microsoft Research. Then Thomas Hoffman developed probabilistic latent semantic analysis (PLSA), taking the ideas of LSA and embedding them in a probability model.
No entries found