Reasearch On The Optimization Of Extracting Pattern In Slot Filling

Slot Filling is one of KBP (Knowledge Base Population) evaluation task introducingby TAC (Text Analysis Conference) in2009. As the continuation and development ofquestion answering and entity relation extraction in information extraction, the main pur-pose of Slot Filling is to explore how to extract attribute information of entities with thespecific type from the open data source.In the current Slot Filling task, the pattern matching method based on dependencypath is used by many systems, and the dependency path is mainly built with the dependen-cies defined by Stanford. Based on the pattern matching approach on dependency path, thisdissertation explores its typical problems and then provides the corresponding solutions.The main work includes the following aspects:Firstly, this dissertation studies several different implementation methods for the SlotFilling task, and implements the baseline with the evaluation resources provided by KBP,natural language processing tools provided by Stanford, and WordNet, a semantic dictio-nary.Secondly, the dependencies defined by Stanford are not meeting the need of buildingdependency path for the Slot Filling task, which includes many issues, such as high de-gree of redundancy, low reliability, and low discrimination. To solve these problems, thisdissertation redefines most of the Stanford dependencies, removes all the useless depen-dencies, refines and transforms several kinds of dependencies, and finally applies thesemethods to the baseline to optimize its performance.Finally, to solve the problem of low precision of the patterns, this dissertation builds akey semantic set for more than half of all the slots, and explains the filtering methods based on the possessive relationship and subject predicate relationship. Our method furtherimproves the F1-measure by10.3%, from14.4%to24.7%.