1. Perceptron Learning for Chinese Word Segmentation
Yaoyong Li, Chuanjiang Miao, Kalina Bontcheva, Hamish Cunningham
Department of Computer Science
University of Sheffield

2. Outline Perceptron learning for Chinese word segmentation (CWS)
Different feature sets
Open task
Result analysis
2(15)

3. Character Based CWS
Check every character to see which of the following four categories it belongs to.
beginning, end or middle character of a multi-character word;
or a single character word
Convert CWS into four binary classification problems
Large training dataset, needs fast algorithm
3(15)

4. Perceptron Algorithm Simple, fast and effective
On-line or batch learning
Checks learning instance one by one
Binary classification
In application, a character is assigned the class which classifier has the maximal output
4(15)

5. Uneven Margins Perceptron
Perceptron with margin has better generalisation capability than the original Perceptron
Uneven margins make Perceptron handling imbalanced data better
Uneven margins Perceptron is as simple and efficient as Perceptron
5(15)

6. Results for four classifications
F1(%) of four classifiers using 4-fold CV on training set of the four corpora
beginning
middle
end
single
combination as
95.64
90.07
95.47
95.27
95.5
cityu
96.64
90.06
96.43
95.14
95.1
msr
96.36
89.79
96.00
94.99
94.9
pku
96.09
89.99
96.18
94.12
6(15)

7. Comparison of PAUM with SVM
Averaged F1(%) and computation time on three subsets and whole data of cityu corpus by 4-fold CV
100
1000
5000
53019
PAUM
73.55
(4s)
78.00
(14s)
88.08
(92s)
95.13
(1.03h)
SVM
75.50
(3.8m)
79.15
(1.1h)
88.78
(13.7h)
------
7(15)

8. Features
The features of the character c0 are from the 5 neighbouring characters, {c-2 c-1 c0 c1 c2 }
1-order features {c-2 c-1 c0 c1 c2}
2-order features {c-2c-1 c-1c-0 c0c1 c1c2 c-1c1}
Α海运 业 雄踞全球之首Ω
8(15)

9. Different feature sets
Different kernels correspond to different feature sets.
Linear kernel amounts to using 1-order features
quadratic kernel: all 1- and 2-order features
semi-quadratic kernel: all 1-order features and some 2-order features, as shown in last slide
linear
quadratic
semi-quadratic
cityu
81.30
94.78
95.13
msr
79.80
94.92
pku
82.33
94.80
95.05
9(15)

10. Open task
Replacement of some special text with symbol in order to achieve better generalisation
Replace every English text with a symbol E
Replace every Arabic number with a symbol N
Smaller training data and less computation time
10(15)

11. Experimental results for open task
Comparison between close and open task using 4-fold CV on training sets of four corpora, F1 (%) and computation time
Only text
Text with E
Text with E & N as
95.53 (8.88h)
95.65 (7.66h)
95.78 (7.07h)
cityu
95.13 (1.03h)
95.25 (0.86h)
95.25 (0.82h)
msr
94.92 (2.62h)
94.98 (1.69h)
95.00 (1.62h)
pku
95.05 (0.70h)
95.08 (0.63h)
95.15 (0.60h)
11(15)

12. Official Results F1 (%) from official results Close task Open task as
94.4
94.8
cityu
93.6
msr
95.6
95.4
pku
92.7
93.8
12(15)

13. Analysis
Comparison with best ones and those from 4-fold CV on training set for Close Task: F1 (%)
Ours official
Best
4-fold
training
Unknown characters rate (%) as
94.4
95.2
95.5
0.484
cityu
93.6
94.3
95.1
0.924
msr
95.6
96.4
94.9
0.034
pku
92.7
95.0
0.215
13(15)

14. Analysis (2)
Comparison with best ones and those from 4-fold CV on training set for Open Task: F1 (%)
test
4-fold
training
Unknown characters rate (%) as
94.8
95.78
0.042
cityu
93.6
95.25
0.86
msr
95.4
95.00
0.031
pku
93.8
95.15
0.119
14(15)

15. Conclusions A simple and fast learning algorithm for WSD
The results are encouraging
Future works:
Better way to deal with unknown characters
More features
15(15)