How do I load and use a CRF trained with Mallet?

Question

I've trained a CRF using GenericAcrfTui, it writes an ACRF to a file. I'm not quite sure how to load and use the trained CRF but

import cc.mallet.grmm.learning.ACRF;
import cc.mallet.util.FileUtils;
ACRF c = (ACRF) FileUtils.readObject(Paths.get("acrf.ser.gz").toFile());

seems to work. However, the labeling seems incorrect and seems to rely on the labels that I pass as input. How do I label using a loaded ACRF?

Here's how I do my labeling:

GenericAcrfData2TokenSequence instanceMaker = new GenericAcrfData2TokenSequence();
instanceMaker.setDataAlphabet(c.getInputAlphabet());
instanceMaker.setIncludeTokenText(true);
instanceMaker.setFeaturesIncludeToken(true);
instanceMaker.setLabelsAtEnd(false);
Pipe pipe = new SerialPipes(new Pipe[] {
        instanceMaker,
        new TokenSequence2FeatureVectorSequence(c.getInputAlphabet(),
                true, false),
});
InstanceList testing = new InstanceList(pipe);
Iterator<Instance> testSource = new LineGroupIterator(
    // initialize the labels to O
        new StringReader("O O ---- what W=the@1 W=hell@2\n"
                    + "O O ---- the W=what@-1 W=hell@1\n"
                    + "O O ---- hell W=what@-2 W=the@-1"),
        Pattern.compile("^\\s*$"), true);
testing.addThruPipe(testSource);
System.out.println(c.getBestLabels(testing.get(0)));

I got that by looking at GenericAcrfTui. Some things I tried:

• When I tried giving different initial labels (other than "O"), then the resulting labeling changed but this doesn't help because I can't guess what labels to give initially, otherwise I wouldn't need a tagger.
• I tried not giving any initial labels at all but that just caused exceptions, it seems that Mallet really wants those labels.

I noticed that there's also the SimpleTagger that can be used to train a CRF but I think that I will still have the same problem using that to label new input.

Any help with labeling using a CRF from SimpleTagger or GenericAcrfTui would help.

BTW I usually use CRF++ but for this task, I want to build my own graph because I'm using dependency parse features.

Answer 1

I figured it out!

The problem was that the pipe didn't know the target alphabet. The solution is to use the CRF's Pipe, like so:

Pipe pipe = crf.getInputPipe();

instead of doing that crazyness to make my own Pipe.

Now if anyone knows a nicer way to make a new Instance using a query, that would be good too, I just copied what the trainer does.