« September 4, 2009
Adam Bargteil, Assistant Professor | HOME

November 6, 2009
Ellen Riloff

Title: Toward the Automatic Acquisition of Semantic Taxonomies from Web Text

Semantic knowledge is enormously helpful for natural language processing, but difficult to obtain. Some manually created semantic resources exist, but they are far from complete and are often woefully inadequate for specialized domains. In this talk, we will describe bootstrapping algorithms that read Web texts to automatically acquire semantic taxonomic knowledge. The key to our approach is the use of a doubly-anchored hyponym pattern that can identify terms that participate in hypernym/hyponym (isa) relations (e.g., "dog" is a "mammal", and "mammal" is an "animal").

First, we will describe a bootstrapping algorithm that begins with one semantic class name (hypernym) (e.g., "mammal") and one seed hyponym term (e.g., "dog") and iteratively learns more hyponym terms that belong to the semantic class. This algorithm uses a graph structure to identify the most reliable hyponym terms. Second, we will describe a bootstrapping algorithm that uses the doubly-anchored hyponym pattern to learn new high-level conceptual terms (i.e., "vertebrate" or "predator") and iteratively use these new terms to fill out the concept space and generate taxonomic relations between words. Along the way, we will discuss some of the (immense!) challenges in formally evaluating the correctness of concepts and taxonomic relations.

(This is joint work with Eduard Hovy and Zornitsa Kozareva at the USC Information Sciences Institute.)

Posted by Admin at November 5, 2009 04:30 PM