A Wikipedia-LDA Model for Entity Linking with Batch Size Changing Instance Selection

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 562–570, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP A Wikipedia-LDA Model for Entity Linking with Batch Size Changing Instance Selection Wei Zhang† Jian Su‡ Chew Lim Tan† †School of Computing ‡ Institute for Infocomm Research National University of Singapore sujian@i2r.a-star.edu.sg {z-wei,tancl}@comp.nus.edu.sg Abstract Entity linking maps name mentions in context to entries in a knowledge base through resolv- ing the name variations and ambiguities. In this paper, we propose two advancements for entity linking. First, a Wikipedia-LDA method is proposed to model the contexts as the prob- ability distributions over Wikipedia categories, which allows the context similarity being measured in a semantic space instead of literal term space used by other studies for the dis- ambiguation. Furthermore, to automate the training instance annotation without compro- mising the accuracy, an instance selection strategy is proposed to select an informative, representative and diverse subset from an au- to-generated dataset. During the iterative se- lection process, the batch sizes at each itera- tion change according to the variance of clas- sifier’s confidence or accuracy between batches in sequence, which not only makes the selection insensitive to the initial batch size, but also leads to a better performance. The above two advancements give significant im- provements to entity linking individually. Col- lectively they lead the highest performance on KBP-10 task. Being a generic approach, the batch size changing method can also benefit active learning for other tasks. 1 Introduction Knowledge base population (KBP)1 involves ga- thering information scattered among the docu- ments of a large collection to populate a know- ledge base (KB) (e.g. Wikipedia). This requires either linking entity mentions in the documents with entries in the KB or highlighting these men- tions as new entries to current KB. Entity linking (McNamee and Dang, 2009) in- volves both finding name variants (e.g. both “George H. W. Bush” and “George Bush Se- nior” refer to the 41st U.S. president) and name disambiguation (e.g. given “George Bush” and 1 http://nlp.cs.qc.cuny.edu/kbp/2010/ its context, we should be able to disambiguate which president it is referring to). Compared with Cross-Document Coreference (Bagga and Baldwin, 1998) which clusters the articles according to the entity mentioned, entity linking has a given entity list (i.e. the reference KB) to which we disambiguate the entity men- tions. Moreover, in the articles, there are new entities not present in KB. For name disambiguation in entity linking, there has been much previous work which de- monstrates modeling context is an important part of measuring document similarity. However, the traditional approach for entity linking treats the context as a bag of words, n-grams, noun phrases or/and co-occurring named entities, and meas- ures context similarity by the comparison of the weighted literal term vectors (Varma et al., 2009; Li et al., 2009; McNamee et al., 2009; Zhang et al., 2010; Zheng et al., 2010; Dredze et al., 2010). Such literal matching suffers from sparse- ness issue. For example, consider the following four …

Content truncated to 500 words