IBM Israel Research Seminars

There have been recent attempts to produce trainable (unsupervised) models of human-language syntax and semantics, as well as morphology. To our knowledge, there has not been an attempt to produce a generative model that incorporates semantic, syntactic, and morphological elements. Some immediate applications of this tool are stemming, word clustering by root, and disambiguation (at the syntactic, semantic, and morphological levels).

In this work, we propose a hierarchical topics-syntax-morphology model. We provide the variational inference and update rules for this model (exact inference is intractable). We show some preliminary results on segmentation tasks.

Joint work with John Lafferty and David Blei.

Speaker bio
Leo Kontorovich earned a bachelor's degree in 2001 from Princeton University in pure mathematics, with a minor in applied mathematics. He interned at Bell Labs in 1999 and 2000. In 2001-2002, he was a visiting Research Fellow at the Hebrew University of Jerusalem. Leo is currently a Ph.D candidate at Carnegie Mellon University in the Center for Automated Learning and Discovery (CALD) program. His advisor is John Lafferty.