IR-916: (2012) Dalton, J. and Dietz, L., "Bi-directional Linkability From Wikipedia to Documents and Back Again: UMass at TREC 2012 Knowledge Base Acceleration Track," Notebook Proceedings of The Twenty first Text REtrieval Conference (TREC 2012), Gaithersburg, MD, USA, November 7-9, 2012. [View bibtex]

Abstract

UMass' objective is to introduce a single model for Knowledge Base Entity Linking and KB Acceleration stream filtering using bi-directional linkability between knowledge base (KB) entries and mentions of the entities in documents. The goal is to identify documents from a stream that are central for a given entity. Our submissions to the TREC 2012 KBA Track consists of three stages: First, potentially relevant documents are retrieved from the stream. Second, potential mentions of the target entity are identified in the retrieved documents. Third, bi-directional links between the potential mentions and the target entity are established or dismissed, giving rise to a filtered set of central documents. Notice, that the third stage is closely related to the entity linking problem of TAC KBP. The baseline run gathers name variations from the Wikipedia KB entry and incorporates them into the probablistic retrieval of stream documents. Our experimental runs further include important NER spans and contextual terms using Latent Concept Expansion from annotated documents from the training time range. Also some experimental runs leverages bi-directional linkability using a supervised re-ranking approach trained on TAC KBP entity linking data as a measure on how likely potential mentions in the stream document refer to the target KB entry. Our experiments show that incorporating entity context from query expansion methods provides significant gains both in precision and recall over the baseline, with all of our experimental runs outperforming the median. Further, our best performing run uses linkability evidence from both directions by using the TAC Entity Linking model.