📄 OntoNotes: The 90% Solution

Summary

This NAACL HLT 2009 tutorial abstract presents OntoNotes, a resource that gives NLP research a shallow but wide-ranging representation of sentence meaning. It reports no experiments. Instead, it describes a corpus built by a five-year collaboration among BBN Technologies, USC's Information Sciences Institute, the University of Colorado, the University of Pennsylvania and Brandeis University. The data combine several annotation layers: Treebank syntax, PropBank predicate-argument structure, ontology-linked word senses, names and coreference. They cover English, Chinese and Arabic across four genres. Annotation design sought a balance between inter-annotator agreement, annotation throughput and depth of representation. Many applications need several layers at once, so the corpus is stored in a relational (MySQL) database that keeps layers consistent while letting each one exist on its own. An object-oriented Python API, with classes mirroring the database tables, wraps the database. The tutorial covers the layers, cross-language differences, the schema and the API, and ends with hands-on cross-layer querying. Its significance is that it offers large, high-quality, multilingual, multi-layer data in an integrated, programmable form.

Key points

• The document is a two-page tutorial abstract (Tutorials, pp. 11–12). It was written when the DARPA GALE OntoNotes project was in its third year, and it contains no quantitative results.
• Languages: English, Chinese and Arabic.
• Genres: four text types, namely weblogs, broadcast conversation, broadcast news and newswire.
• Annotation layers: Treebank, PropBank, word sense, names, coreference and an ontology. Question answering and machine translation are named as target applications.
• Infrastructure: a MySQL database schema (with ER diagram) interconnects the layers while keeping each one independent, which supports multi-site work. A Python API maps classes to tables and can be extended with new layers.
• Hands-on session: participants load a database from a MySQL dump, build Python objects for the layers, run cross-layer queries and write modified data back.