talks.info

Ripper 分类算法

A simple text classification problem: classify postings to /usr/msgs
as talk announcements or "other".  There are 818 messages, going back
to Aug 93.  I used messages numbered 500 and up as test cases.
Average words/message is around 160.  I used a simple lex program to
tokenize the data. Class labels were obtained by manual inspection.

Doug McIlroy suggested this dataset, and also suggested that egrep -i
'talk|abstract' would be hard to beat. (Which is correct).

		recall	prec.	 #errors  %err	fp	fn

doug's egrep			  22	  6.77 	8	14

rocchio -m acc	75.9	 96.9	  22	  6.77	2	20
rocchio -m F	95.2	 86.8	  16	  4.92	12	 4

ripper -L0.06 	83.1	 87.3	  24	  7.38	10	14	
ripper -L0.125 	89.2	 89.2	  18	  5.54	9	 9
ripper -L0.25 	95.2	 91.9	  11	  3.38	7	 4
ripper -L0.5 	94.0	 91.8	  12	  3.69	7	 5
ripper -L1 	89.2	 92.5	  15	  4.62	6	 9
ripper -L1.5 	79.5	 95.7	  20	  6.15	3	17
ripper -L2 	77.1	100.0	  19	  5.85	0	19