KeisukeKamataki writeup of Borthwick et al.
From Cohen Courses
Jump to navigationJump to searchThis is a review of Borthwick_1998_exploiting_diverse_knowledge_sources_via_maximum_entropy_in_named_entity_recognition by user:KeisukeKamataki.
- They used Maximum Entropy modeling for model construction and Viterbi decoding for named entity(proper noun) detection. They used a set of 5 types of features and achieved very good performance like 95 of F-Measure value of same domain training/testing data. Using only lexical feature also achieved the F-measure 88.13.
- I like: They give good explanations about features they used and clearly stated the some tweaks to reduce redundant features which sound reasonable and could be applicable in other general use. They also explored different amount of training data and limited feature (with only lexical feature) and showed how much amount of training data is required to get satisfactory performance.
