434 JOURNAL OF LAW AND POLICY
The performance of the models based on word features has
similar characteristics. It steadily grows or remains practically
stable until about 1,500 features and then drops significantly.
The drop is much more abrupt in the case of Society texts
Figure 3: Performance of the cross-topic attribution models
(training on Politics, test on Society).
Figure 4: Performance of the cross-topic attribution models
(training on Politics, test on World).
30
40
50
60
70
80
90
100
0 5000 10000 15000
Accuracy (%)
Features
Words
Char 3-
grams
30
40
50
60
70
80
90
100
0 5000 10000 15000
Accuracy (%)
Features
Words
Char 3-
grams