1 Knowledge-Guided Schema Discovery from Semi-Structured Text

A significant amount of information is expressed as the semi-structured, non-grammatical text found in auction listings and classified advertisements. There would be value in automatically fitting this type of information to relational database schemas. Work has been conducted on automatically populating such a database, and automatic schema mapping is a large area of research, but the problem of automatically generating such a schema is relatively unaddressed. Schema Discovery is the recent research thrust that addresses this. In this paper, we introduce the TESS system for knowledge-guided schema discovery from semi-structured text. TESS performs term extraction over listing text and applies semantic reasoning to differentiate common terms into categories. This approach differs from existing schema discovery work in that it requires no hand-labeling or training examples, and it takes advantage of existing large databases of collected semantic knowledge. Problem Description The goal of this project will be to extract database schemata from semi-structured text, specifically from auction and classified listings, which lend themselves well to a relational representation. The work is inspired by efforts going on in the field to automatically populate such a database in order to allow improved searching and some automated reasoning; in particular, Michelson & Knoblock [5] have created a tool named Phoebus, which tries to match cars listed on craigslist1 with an entity table that knows what attributes should appear in any listing for a car, enabling intelligent querying. While Michelson & Knoblock mention the possibility of expanding the search domain automatically, their DB schema are actually created by hand, rather than discovered automatically. Such discovery would require, or at least be assisted by, some semantic understanding of the content of the target text. Tools that do large-scale knowledge representation, such as TEXTRUNNER [1], Cyc [3], and OpenMind Common Sense [7], may provide that background knowledge