NatS Oberseminar 14. April 2005

Automatic Rule Acquirement for financial Event Extraction

Zhou Jianhui

Information Extraction (IE) can extract desired contents from free or semi-structured texts on the Internet or in the other information streams, and then organizes the extracted information into structured framework for applications to use. Event extraction, which is located at the upper level of the hierarchy of an IE system, is primarily responsible for identifying and sifting the information associated with the interested events out of the parsed texts.

Our system concentrates on the event extraction in financial news articles. Our system employs traditional rule-based approach and makes uses of machine learning to gain extraction rule set automatically. During event extraction, it takes our event information directly from the text with identified NP, PP and VP, without involving complicated S-V-O analysis, which greatly reduces the complexity implementation. And also, in the rules training process, it gradually unfastens constraints of the initial rules among five levels of strictness in order to ensure to gain as much information as possible from train corpus. Applying the generated rule set to the pre-defined events,our system achieves the best result in "dividend" event with 92.9% in close test and 78.3% in open test by F-Measure and the worst result in "loan" event, which reaches 80.3% in closed test and 69.9% in open test.