Merging information by discourse processing for information extraction

In information extraction tasks, a finite-state pattern matcher is widely used to identify individual pieces of information in a sentence. Merging related pieces of information scattered throughout a text is usually difficult, however, since semantic relations across sentences cannot be captured by the sentence level processing. The purpose of the discourse processing described in this paper is to link individual pieces of information identified by the sentence level processing. In the Tipster information extraction domains, correct identification of company names is the key to achieving a high level of system performance. Therefore, the discourse processor in the Textract information extraction system keeps track of missing, abbreviated, and referenced company names in order to correlate individual pieces of information throughout the text. Furthermore, the discourse is segmented, so that data can be extracted from relevant portions of the text containing information of interest related to a particular tie-up relationship.<>