FREQTM
- TinySVM with tree kernels on marked labeled ordred trees for node relation labeling (e.g., SRL task).
- Plus methods for the fast training with the above tree kernels described in
"Speeding up Training with Tree Kernels for Node Relation Labeling", Jun'ichi Kazama and Kentaro Torisawa. In Proceedings of HLT-EMNLP 2005
- "FREQTM" is the name of the core algorithm for the speed-up, which is a modification of the FREQT algortihm [Asai et al. 02].
Original TinySVM
To find the original TinySVM developed by Taku Kudo, go here.
Since I modified the original TinySVM for my purpose, bugs not existed in the original may have been introduced. Do not bother the original author before you check (e.g., by using diff) whether the problem really stems from the original.
FREQT
I learned the FREQT algorithm and started my FREQTM code by consulting the implementation of FREQT by Taku Kudo.
Download
2005/10/15
TinySVM-0.09-wm-2005-10-15.tar.gz
Note:
- As we mentioned in page 4 of the HLT-EMNLP05 paper, this version does not check whether |sup(F_i)| > M (M is the maximum support size of the malicious subtrees found so far).
- This version contains experimental implementation of pruning using mark information. This is enabled with -k option for tree2fv.
- FREQTM for normal labeled ordered trees are also implemented but the corresponding speed-up code is not yet implemented.
- Bug I found that pattern output function (-T -O) does not work correctly. Will be fixed in the next release.
Documents
Under construction. (see README in the downloaded package at now).
ToDo
- documentation
- templatize tree kernels
- FREQTM support for other tree kernels such as "elastic" Kashima's kernel
- more efficient enumeration
- fast classification
Contact
If you find bugs or have comments please let me know.
Jun'ichi Kazama