IR-823: (2011) Dalton, J.,  Allan, J. and Smith, D., "Passage Retrieval for Incorporating Global Evidence in Sequence Labeling," Proceedings of the 20th ACM Conference on Information and Knowledge Management (CIKM '11), pp. 355-364. [View bibtex]

Many current methods for addressing feature sparsity and incorporating non-local features are ad-hoc and very domain and task depen- dent. We introduce a principled method of ad- dressing feature sparsity using non-local evi- dence based on passage retrieval. We apply re- trieval methods to a local context before infer- ence time, which allows the use of Viterbi de- coding while incorporating non-local features. We show that retrieval-based context aggrega- tion that leverages global collection wide in- formation outperforms existing ad-hoc aggre- gation methods. We demonstrate that our re- trieval methods can leverage unlabeled data as a simple means of unsupervised model adap- tation. Our results show that retrieval features consistently improve named entity recogni- tion effectiveness on both CoNLL and topic- specific book collections.