Yandongl writeup of Borthwick 1998

This is a review of Borthwick_1998_exploiting_diverse_knowledge_sources_via_maximum_entropy_in_named_entity_recognition by user:Yandongl.

This paper presents the novel idea of using maximum entropy for NER task. The notations look much like later-appearing CRF. Authors spend lots of space introducing features, and the features have binary values. Simply rule-based feature selection method is adopted.

Finally they use Viterbi,like HMM, to find a sensible assignment with maximum probability.

Results showed that MENE itself doesn't necessarily perform well. However, when combined with some weak systems (Proteus/Manitoba) the new system outperforms their strong baseline. Also, the result is comparable to HMM's/