Bidirectional model for imdb: am I interpreting it right? (read inside for details)

Part 2 & Alumni (2018)
1

So if I’m not mistaken, this is my understanding:

use the same vocabulary and everything, but flip the data around, so that you have anther new set of data, and model&encoder. OKAY.

Then I use the new model and train my classifier on that, and get some predictions.

… and I combine the prediction of original classifier and the new classifier to get better result?

Am I getting this right/wrong?

1 Like
2

That’s right. You just average the predictions of the two models.

3 Likes
3

Ensemble forward and backward prediction

2 Likes
4

What exactly do you mean by the predictions… (I am getting little confused…)

5

The probability vectors of size vocab_length for the next word (predicted word).

2 Likes
6

You just average the predictions of the two models.

I think “predictions” means the output of the classifiers, not

The probability vectors of size vocab_length for the next word (predicted word).

(Is it possible to ensemble the result of the two language models?)

7

What do you mean by it, please elaborate backward prediction