Klein et al, CONLL 2003

From Cohen Courses
Jump to navigationJump to search


Dan Klein, Joseph Smarr, Huy Nguyen and Christopher D. Manning. 2003. Named Entity Recognition with Character-Level Model. In Proceedings of CoNLL-2003.

Online version

ACL Anthology


In this paper, the authors propose using character representations instead of word representations in the Named Entity Recognition task. The first model proposed is the character-level HMM with minimal context information and the second model is the maximum-entropy conditional markov model with rich context features.

In character-level HMM, each character is represented with one state which depends only on the previous state. And each character observation depends on the current state and on the previous n-1 observations. In order to prevent characters of a word getting different state labels, they represent each state with a pair(t,k) where t is entity type and k is length of time of being in that state. They limit the use of k to n-gram history and represent the final state with F. They experiment with two model in which they used the preceding context in one , no context in the other one. Using context information slightly gave better performance then the model without the context. But still both models got much better results then using a word-level HMM.

The authors also built a character-base conditional markov model ,where they used previous classification decisions as features. They also used joint-tag sequence, longer distance sequence and tag-sequence features, letter type pattern features and additional more context features. Overall they showed that rich features sets produced from the context is helping.

The paper does not give any information on the used data set.

A previous paper that uses character-level approach was the Cucerzan and Yarowsky, SIGDAT 1999. In that paper the authors used the prefix and suffix tries but in this paper all the characters are used.