Erik Chan

I tried the BiLSTM_Model (val accuracy stayed at 0.99 throughout training), but results don't seem to be good. Accuracy for [CLS], [SEP], [BOS] are 98-99% accurate. But B-TERM, I-TERM, and...

Providing an example of my training data: ``` dict = { "O": 1, "[CLS]": 2, "[SEP]": 3, "[BOS]": 4, "B-TERM1": 6, "I-TERM1": 7 } >>> train_x[10] ['[CLS]', 'during', 'the', 'track',...

> What is the evaluation result when you evaluate on the training set and the validation set? Maybe this is because of not enough dataset? Good idea. Evaluation result of...

Seems like the model isn't learning anything. @BrikerMan do you think this is a problem with the library? Or something I am not doing correctly?

As I imagined, the model treats each of the 26 different tags as separate entities, which makes it harder for the model to learn. Accuracy for [CLS], [SEP], [BOS] are...

Looking at the results, my guess is the simpler model with one keyword is better. But the main problem where the model isn't learning anything is the main issue. Theoretically,...

While debugging I've found that even when the training input is the same as the desired output the model is still inaccurate. For example: ``` # TWO INPUTS [CLS] 这是一只狗和这是一只红猫...

> Here is the new numeric embedding, just add this to your code and replace the old numeric embedding. I just tried this, results are identical. Looks like the numeric...

While I haven't tried this myself, I think sorting the record by id before sorting by another attribute might solve it. I haven't tried it myself, but I imagine it...

Are you referring to showing the Open Graph data? A preview of the link including the text, image/video, etc? It would be great to show preview/embed of youtube videos