 Use active learning as a sampling framework: - greedily annotate each sequence with current model, sort by increasing correctness probability -> this favors longer sequences which are harder to predict - same as above but normalize by sequence length n - fraction of models that disagree on the label of a word. Normalized by n - you can use label propagation to further decide on next samples to pick |
Artificial Intelligence + NLP + deep learning > AI > Machine Learning > Neural Networks > Deep Learning >