Table 16 The ten most important features related to difficult (D) and easy (E) classes measured by information gain
From: A detailed error analysis of 13 kernel methods for protein-protein interaction extraction
Difficult (D) | Easy (E) | |||||
|---|---|---|---|---|---|---|
Rank | Feature name | ± | IG | Feature name | ± | IG |
1 | sentence length (char) | − | 0.0089 | label entropy in ST | + | 0.110 |
2 | label entropy in ST (SP) | − | 0.0086 | sentence length (char) | + | 0.090 |
3 | dep frequency in DG | − | 0.0079 | label entropy in DG | + | 0.089 |
4 | # of proteins in sentence | − | 0.0078 | nn frequency in DG | − | 0.081 |
5 | sentence length (word) | − | 0.0069 | appos frequency in DG | − | 0.079 |
6 | conj_and frequency in DG | − | 0.0069 | conj_and frequency in DG | − | 0.076 |
7 | prep_with frequency in DG | − | 0.0066 | dep frequency in DG | − | 0.073 |
8 | prep_with occurrence in DG | − | 0.0066 | det frequency in DG | − | 0.069 |
9 | nsubjpass frequency in DG | − | 0.0059 | amod frequency in DG | − | 0.063 |
10 | prep_in frequency in DG | − | 0.0057 | dobj frequency in DG | − | 0.062 |