Verified error fixes and readability improvement

[ahwang16]

This version is verified to work on Google Colab using a GPU and Python3. This fixes typos in the original version that prevented the program from running, eliminates the dependency on contrib, and improves readability by changing a couple variable names and spacing.

[glicerico]

Thanks for this PR @ahwang16 .
Did you train on switchboard? Can you please share what accuracy did your trained model achieve?

[ahwang16]

I did train on switchboard, but I don't have much information because my summer internship ended in August and I have not returned to Google yet. Without any sort of parameter search/optimization, my trained model achieved a test accuracy of 0.743 (using 300-dimensional GloVe embeddings). I later used BERT embeddings, which pushed the accuracy up to 0.809.

[glicerico]

Thanks, sounds quite promising, specially using BERT embeddings.
Can you offer some insight into how to use real data?
Currently I see the test data declared as:

data = [[[1,2,3,4],[1,2,3],[2,3,5]],[[1,0], [4]],[[1,2,8,4],[1,1,3],[2,3,9,1,3,1,9]], [[1,2,3,4,5,7,8,9],[9,1,2,4],[8,9,0,1,2]],[[1,2,4,3,2,3],[9,8,7,5,5,5,5,5,5,5,5]],[[1,2,3,4,5,6,9],[9,1,0,0,2,4,6,5,4]],[[1,2,3,4,5,6,7,8,9],[9,1,2,4],[8,9,0,1,2]],[[1]] , [[1,2,11,2,3,2,1,1,3,4,4], [6,5,3,2,1,1,4,5,6,7], [9,8,1], [1,6,4,3,5,7,8], [0,9,2,4,6,2,4,6], [5,2,2,5,6,7,3,7,2,2,1], [0,0,0,1,2,7,5,3,7,5,3,6], [1,3,6,6,3,3,3,5,6,7,2,4,2,1], [1,2,4,5,2,3,1,5,1,1,2], [9,0,1,0,0,1,3,3,5,3,2], [0,9,2,3,0,2,1,5,5,6], [9,0,0,1,4,2,4,10,13,11,12], [0,0,1,2,3,0,1,1,0,1,2], [0,0,1,3,1,12,13,3,12,3], [0,9,1,2,3,4,1,3,2]]]
labels = [[1,2,1],[0, 3],[1,2,1],[1,0,2], [2,1], [1,1], [2,1,2], [4], [0,1,2,0,2,4,2,1,0,1,0,2,1,2,0]]

I guess the labels correspond to speech acts. What about the arrays of data? Are they word ids for each utterance?

[ahwang16]

For real data, I used the Switchboard Dialogue Act Corpus (SWDA) and an internal dataset (not sure if it is available to the public).

The multilayered structure of data can be pretty confusing. In the sample test data, each integer is a token, a list of tokens is an utterance, and a list of utterances is a dialogue. The labels correspond to dialogue acts at the utterance level.

So you can understand the data format like:

data =
[
    [
         [word1, word2, ...], # utt 1 of dialogue 1
         [word1, ... ], # utt 2 of dialogue 1
         ...
    ] # dialogue 1
    ...
] # entire dataset

labels = 
[
    [DA of utt 1, DA of utt 2, ...], # dialogue 1
    ...
]

For the two-tier architecture in this code, the data should be parsed at the utterance level for the first layer and the dialogue level for the second layer.

[glicerico]

Thanks for the description!