Collective Information Extraction Using First-Order Probabilistic Models

ABSTRACT Traditional information extraction (IE) tasks roughly consist of named-entity recognition, relation extraction and coreference resolution. Much work in this area focuses primarily on separate subtasks where best performance can be achieved only on specialized domains. In this paper we present a collective IE approach combining all three tasks by employing linear-chain conditional random fields. The usage of probabilistic models enables us to easily communicate between tasks on the fly and error correction during the iterative process execution. We introduce a novel iterative-based IE system architecture with additional semantic and collective feature functions. Proposed system is evaluated against real-world data set, introduced in the paper, and results are better over traditional approaches on two tested tasks by error reduction and performance improvements