Our goal is the automatic abstraction of journal articles, initially in the field of crop protection. We build a set of templates against which the original text is compared. The templates are designed so that they match the text at points of high information content, where inferences can be made about which expressions best reflect the content of the document. Strings found by matching templates are assigned roles specific to each template. These roles correspond to slots in a frame which is used to represent the document as a whole. An abstract is generated which contains the concept-strings selected from the text.
Author and article information
Computing Department, Lancaster University