A Systematic Cross-Comparison of Sequence Classifiers

Abstract

In the CoNLL 2003 NER shared task, more than two thirds of the submitted systems used a feature-rich representation of the task. Most of them used the maximum entropy principle to combine the features together. Others used large margin linear classifiers, such as SVM and RRM. In this paper, we compare several common classifiers under exactly the same conditions, demonstrating that the ranking of systems in the shared task is due to feature selection and other causes and not due to inherent qualities of the algorithms, which should be ranked otherwise. We demonstrate that whole-sequence models generally outperform local models, and that large margin classifiers generally outperform maximum entropy-based classifiers.

Integrating unsupervised and supervised word segmentation: The role of goodness measures

“…CRFs are a statistical sequence modeling framework that is reported to outperform other popular learning models, including MaxEnt (maximum entropy), in a number of natural language processing applications [35]. CRFs modeling is first applied to Chinese word segmentation in [31], treating it as a binary decision task to determine whether a character is the beginning of a word in a sentence.…”

A Simple and Efficient Model Pruning Method for Conditional Random Fields

“…CRFs often outperform maximum entropy Markov model (MEMM) [2], another popular structure learning method. The main reason is that, among directed graphical models, CRFs do not suffer from the label bias problem as much as MEMM and other conditional Markov models do [1].…”

Scaling Conditional Random Fields by One-Against-the-Other Decomposition

“…It often outperforms a maximum entropy (MaxEnt) model [2] , another popular machine learning method in NLP. The main reason is that, among directed graphical models, CRF does not suffer from the label bias problem as much as the MaxEnt and other conditional Markov models do [1] .…”

“…(2) L-BFGS [4] is a typical algorithm for CRFs training. To label an unseen sequence, we compute the most likely labeling Y * as in (3) by Viterbi algorithm [5] .…”