Principles of Template Design

PRINCIPLES OF TEMPLATE DESIGN Jerry Hobbs, David Israel Artificial Intelligence Center SRI International Menlo Park CA 94025 ABSTRACT The functionality of systems that extract information from texts can be specified quite simply: the input is a stream of texts and the output is some representation of the informa- tion to be extracted. Hence, the problem of template design is an instance of the problem of knowledge representation. In particular, it is the problem of representing essential facts about situations in a way that can mediate between texts that describe those situations and a vaxiety of applications that involve reasoning about them. The research on which we report here is directed at eluci- dating principles of template design and at compiling these, with examples, in a manual for template designers. 1. Introduction The functionality of systems that extract information from texts can be specified quite simply: the input is a stream of texts and the output is some representation of the informa- tion to be extracted. In the message understanding research promoted by ARPA through its Human Language Technol- ogy initiative, the form of this output has been templates (feature-structures), with complex path-names (slots) and various constraints on fillers. The design of these templates, especially considered as concrete data structures, has been determined to some degree at least by considerations having to do with automatic scoring. Beyond that, it has not been made clear what principles have driven or should drive the design of these output forms; but it has become clear that serious defects in the form of the output can undermine the utility of an information extraction system. If the output is unusable, or not easily usable, the breadth and reliability of coverage of the natural language analysis component will be of little value. As part of the DASH research project on Data Access for Situation Handling, we axe attempting to elucidate principles of template design and at compiling these, with examples, in • a manual for template designers. Our methodology has in- eluded detailed critical analysis of the templates from a vari- ety of information extraction tasks (MUC-4, MUG-5, Tipster- 1, the Waxbreaker Message Handling [WBMH] tasks), to- gether with the creation of templates for the TREC topic descriptions and narratives. The design of templates, or more generally, abstract data structures, as output forms for automatic information ex- traction systems must be sensitive to three different but in- teracting considerations: 1. the template as representational device 2. the template as generated from input 3. the template as input to further processing, by humans or programs or both. The central consideration in our research is that of the tem- plate as a representational device. The problem of template design is a special case of the general problem of knowledge representation. In particular, it is the problem of represent- ing, within a constrained formalism, essential facts about sit- uations in a way that can mediate between texts that describe those situations and a variety of applications that involve rea- soning …

Content truncated to 500 words