CUNY-BLENDER TAC-KBP2010 Entity Linking and Slot Filling System Description

摘要

The CUNY-BLENDER team participated in the following tasks in TAC-KBP2010: Regular Entity Linking, Regular Slot Filling and Surprise Slot Filling task (per: disease slot). In the TAC-KBP program, the entity linking task is considered as independent from or a pre-processing step of the slot filling task. Previous efforts on this task mainly focus on utilizing the entity surface information and the sentence/document-level contextual information of the entity. Very little work has attempted using the slot filling results as feedback features to enhance entity linking. In the KBP2010 evaluation, the CUNY-BLENDER entity linking system explored the slot filling attributes that may potentially help disambiguate entity mentions. Evaluation results show that this feedback approach can achieve 9.1percent absolute improvement on micro-average accuracy over the baseline using vector space model. For Regular Slot Filling we describe two bottom-up Information Extraction style pipelines and a top-down Question Answering style pipeline. Experiment results have shown that these pipelines are complementary and can be combined in a statistical re-ranking model. In addition, we present several novel approaches to enhance these pipelines, including query expansion, Markov Logic Networks based cross-slot/cross-system reasoning. Finally, as a diagnostic test, we also measured the impact of using external knowledge base and Wikipedia text mining on Slot Filling.