On 10/4/11 10:12 AM, Riccardo Tasso wrote:
Hi, are there any kind of guideline to choose good parameters for training models of POS tagging and Chunking?
For maxent a small cutoff between 3 to 5 works well and around 100 iterations, as far as I know using more iterations in maxent slightly improves your results, but it may take time.
The best thing to do is, to create a test set, and then measure the different settings on this test set, we have for both components a built-in evaluator. You might as well want to experiment
with new features. Jörn
