-
Notifications
You must be signed in to change notification settings - Fork 20
Hyperparameter study
André Pires edited this page May 22, 2017
·
1 revision
In order to improve and check the influence of the hyperparamenter for each tool, I ran the tools again with new configurations. Because the repeated 10-fold cross validation took a long time to run, I decided to perform only repeated validation, with a split of 70% training and 30% testing with the HAREM golden collection.
Tool | Parameter | Default | Values |
---|---|---|---|
Stanford CoreNLP | max iterations | infinite | - |
OpenNLP | iterations | 100 | [70 - 130, 125, 135, 150 - 180, 200] |
cutoff | 5 | [0, 3 - 7, 10] | |
spaCy | iterations | 10 | [10 - 110] |
NLTK | [ME] max iterations | 10 | [10 - 120] |
[ME] min_ll | 0 | - | |
[ME] min_lldelta | 0.1 | [0 - 0.2 (0.05)] | |
[DT] entropy cutoff | 0.05 | [0.03 - 0.13 (x0.01)] | |
[DT] depth cutoff | 100 | 2, 5, [10 - 120] | |
[DT] support cutoff | 10 | 3, [7 - 20] |