Combining local and non-local information with dual decomposition for named entity recognition from text

Named entity recognition (NER) is the task of segmenting and classifying occurrences of names in text. In NER, local contextual cues provide important evidence, but non-local information from the whole document could also prove useful: for example, it is useful to know that “Mary Kay Inc.” has been mentioned in a document to classify subsequent mentions of “Mary Kay” as an organization and not as a person. Previous works for NER typically model the problem as a sequence labeling problem, coupling the predictions of neighboring words with a Markov model such as conditional random fields. We propose applying the dual decomposition approach to combine a local sentential model and a non-local label consistency model for NER. The dual decomposition approach is a fusion approach which combines two models by constraining them to agree on their predictions on the test data. Empirically, we show that this approach outperforms the local sentential models on four out of five data sets.