micter
micter copied to clipboard
micter is a micro word segmenter which splits a sentence into words.
Micter
Micter is a machine learning based word segmenter. Word splitting is a bit difficult task in agglutinative language like Japanese. Micter could be used for such languages (if you can arrange learning data.)
Micter's learning algorithm is Support Vector Machine with L1 regularization. Optimization algorithm is FOBOS(Forward-Backward Splitting). To know details of FOBOS, you may want to read to "Efficient Learning using Forward-Backward Splitting" by Duchi and Singer. (pdf)
How to build
./waf configure
./waf build
You need python and g++ 4.1 or higher to build micter for now.
How to use
$ ./build/default/micter-train -m modelfile.txt learndata1.txt learndata2.txt ...
$ ./build/default/micter -m modelfile.txt
type some sentence here.
there is a model file learned with japanese blog data.
$ wget http://kodou.net/~tkng/micter/micter.model
$ ./build/default/micter -m micter.model
type some japanese sentence here.
micter has a benchmark mode.
./build/default/micter --bench -m modelfile.txt learndata1.txt
will output accuracy, precision and recall of closed test.
learning data format
1 word in 1 line. empty line is treated as a sentence break.
Performance Test
test data: crowled Japanese blog entries (roughly 250MB for train, 250MB for test. character encoding is UTF-8.)
train data and test data is splitted with mecab.
open test:
accuracy: 0.941353
precision: 0.976557
recall: 0.913756
closed test (test with train data):
accuracy: 0.942232
precision: 0.977311
recall: 0.914956
TODO
- strict parsing of model file (current implemantation is not robust).
- implement feature vector iterator.