Hi, I used your model on two of my datasets on different tasks. Compared to the benchmark your model performed similarly on the tasks. There is one exception, for on task, where I have many classes >140. The accuracy drops to 5% (benchmark 55%-65%). Does the model have a problem, when predicting many classes? I'm trying windows size = 5 and pmi <= 10000 right now and already tried different learning rates (3e-4, 1e-3, 3e-3). Are they any other parameters you think I should try?
Some facts to my datasets: 270000 documents, average length of documents 690 words).
Thanks!
Hi, I used your model on two of my datasets on different tasks. Compared to the benchmark your model performed similarly on the tasks. There is one exception, for on task, where I have many classes >140. The accuracy drops to 5% (benchmark 55%-65%). Does the model have a problem, when predicting many classes? I'm trying windows size = 5 and pmi <= 10000 right now and already tried different learning rates (3e-4, 1e-3, 3e-3). Are they any other parameters you think I should try? Some facts to my datasets: 270000 documents, average length of documents 690 words). Thanks!