With the field of computational linguistics' empirical revolution of
the 1990s came the realization that human intuitions about language are insufficient for accurate and robust natural language technologies. The move from hand-written, rule-based models to data-driven techniques led to huge advances, yet we still leaned on
human intuition for constructing annotated linguistic datasets.
Despite major advances in this paradigm (some of which we'll discuss in this talk), we now know that, in the wild world of real and diverse linguistic data, natural language technology raised on expert-made annotations remains insufficient for real, robust applications.
In this talk we adopt the premise that unsupervised learning will, in the long run, be the way forward for learning computational models of language cheaply. We focus on dependency syntax learning without trees, beginning with the classic EM algorithm and presenting several ways to alter EM for drastically improved performance using crudely represented "knowledge" of linguistic universals. We then present
more recent work in the empirical Bayesian paradigm, where we encode
our background knowledge as a prior over grammars, applying inference to obtain hidden structure. Of course, "background knowledge" is still human intuition. We argue, however, that by representing this knowledge compactly in a prior distribution--far more compactly than the many decisions made in building treebanks--we can experimentally explore the connection between proposed linguistic universals and unsupervised learning.
This talk includes discussion of joint work with Shay Cohen, Dipanjan Das, Jason Eisner, Kevin Gimpel, Andre Martins, and Eric Xing.