en
0.25
0.5
0.75
1.25
1.5
1.75
2
Fast Support Vector Machines for Structural Kernels
Published on 2011-10-033116 Views
In this paper, we propose three important enhancements of the approximate cutting plane algorithm (CPA) to train Support Vector Machines with structural kernels: (i) we exploit a compact yet exact r
Related categories
Presentation
Fast Support Vector Machines for Structural Kernels00:00
Structured Data00:19
Ex: Predicate-Argument Identification (1)01:20
Ex: Predicate-Argument Identification (2)01:46
Explicit feature vector representation01:52
State of the art in many tasks02:18
Training SVMs with Structural Kernels02:42
Key ideas in the paper03:01
CPA in a nutshell (1)03:49
CPA in a nutshell (2)04:22
CPA in a nutshell (3)04:59
CPA in a nutshell (4)05:09
CPA in a nutshell (5)05:14
Computing most violated constraint (MVC) (1)05:19
Computing most violated constraint (MVC) (2)06:00
Computing most violated constraint (MVC) (3)06:16
Computational bottleneck (1)06:41
Computational bottleneck (2)06:44
Approximate CPA + Structural Kernels07:14
Compact model representation using DAGs07:49
Three syntactic trees and the resulting DAG (1)08:36
Three syntactic trees and the resulting DAG (2)08:56
Three syntactic trees and the resulting DAG (3)09:04
Three syntactic trees and the resulting DAG (4)09:08
Computational bottleneck (3)09:20
SDAG (1)09:51
SDAG (2)09:57
SDAG+ (1)10:16
SDAG+ (2)10:32
Example: computing Kdag (1)10:45
Example: computing Kdag (2)10:55
Example: computing Kdag (3)11:15
Experimental setup11:28
Semantic Role Labeling (SRL) dataset12:03
Speedups during training on SRL dataset (100k)12:34
Classification speedups on SRL dataset (100k)13:46
Parallelization14:11
Speedups due to parallelization on 50k YA dataset14:44
Handling class-‐imbalance problem15:06
Results on QA classification: TREC and Yahoo! Answers16:08
Conclusions16:33
Future work17:26
Thank you!18:19