Using Vowpal Wabbit for NER

9

1

The Vowpal Wabbit (VW) apparently supports sequence tagging functionality via SEARN. The problem is that I cannot find anywhere detailed parameter list with explanations and with some examples. The best I could find is Zinkov's blog entry with a very short example. The main wiki page barely mentions SEARN.

In the checked out source code I found demo folder with some NER sample data. Unfortunately, the script running all the tests does not show how to run on this data. At least it was informative enough to see what is the expected format: almost the same as standard VW data format, except that entries are separated by blank lines (this is important).

My current understanding is to run the following command:

cat train.txt | vw -c --passes 10 --searn 25 --searn_task sequence \
--searn_passes_per_policy 2 -b 30 -f twpos.vw

where

--searn 25 - the total number of NER labels (?)

--searn_task sequence - sequence tagging task (?)

--searn_passes_per_policy 2 - not clear what it does

Other parameters are standard to VW and need no additional explanation. Perhaps there are more parameters specific to SEARN? What is their importance and impact? How to tune them? Any rules of thumb?

Any pointers to examples will be appreciated.

Vladislavs Dovgalecs

Posted 2015-06-06T07:00:13.363

Reputation: 471

Answers

1

John Langford's documentation on GitHub could help. You can find something on the Learning to Search Sub System page.

Brian Spiering

Posted 2015-06-06T07:00:13.363

Reputation: 10 864